Compare commits

...

14 Commits
1.4.1 ... 1.6.7

Author SHA1 Message Date
d42c240362 Memorization optimziations
All checks were successful
Publish Library / Build NPM Project (push) Successful in 1m2s
Publish Library / Tag Version (push) Successful in 10s
2026-08-31 12:38:40 -04:00
c1a16096ae Keep message progress on abort
All checks were successful
Publish Library / Build NPM Project (push) Successful in 43s
Publish Library / Tag Version (push) Successful in 14s
2026-08-29 21:18:28 -04:00
ff0ee0b60e Patched memory merging
All checks were successful
Publish Library / Build NPM Project (push) Successful in 45s
Publish Library / Tag Version (push) Successful in 10s
2026-08-28 16:48:46 -04:00
0a6f1e4d62 Refined memory management prompts
All checks were successful
Publish Library / Build NPM Project (push) Successful in 1m18s
Publish Library / Tag Version (push) Successful in 20s
2026-08-25 10:03:36 -04:00
08a351e028 Better memory management
All checks were successful
Publish Library / Build NPM Project (push) Successful in 59s
Publish Library / Tag Version (push) Successful in 11s
2026-08-24 14:42:10 -04:00
85c01d3ef1 Added official file support
All checks were successful
Publish Library / Build NPM Project (push) Successful in 30s
Publish Library / Tag Version (push) Successful in 10s
2026-08-17 15:50:48 -04:00
5826573d5c Added official file support
All checks were successful
Publish Library / Build NPM Project (push) Successful in 50s
Publish Library / Tag Version (push) Successful in 13s
2026-08-17 15:16:32 -04:00
797a40a566 Added official file support
All checks were successful
Publish Library / Build NPM Project (push) Successful in 58s
Publish Library / Tag Version (push) Successful in 13s
2026-08-16 15:40:50 -04:00
7308927a3c max token rename
All checks were successful
Publish Library / Build NPM Project (push) Successful in 35s
Publish Library / Tag Version (push) Successful in 14s
2026-08-05 16:16:30 -04:00
04f038ba65 Memory prompt refinement
All checks were successful
Publish Library / Build NPM Project (push) Successful in 38s
Publish Library / Tag Version (push) Successful in 19s
2026-08-05 13:14:21 -04:00
d42f58d710 Memory refinement
All checks were successful
Publish Library / Build NPM Project (push) Successful in 54s
Publish Library / Tag Version (push) Successful in 11s
2026-08-05 12:22:13 -04:00
878a8794ee Rebuild graph edges on changes
All checks were successful
Publish Library / Build NPM Project (push) Successful in 46s
Publish Library / Tag Version (push) Successful in 19s
2026-08-04 17:05:58 -04:00
3f1289d993 Small agent tweaks
All checks were successful
Publish Library / Build NPM Project (push) Successful in 49s
Publish Library / Tag Version (push) Successful in 9s
2026-08-04 14:33:28 -04:00
077f75cdd9 Fixed delegate agent history... again
All checks were successful
Publish Library / Build NPM Project (push) Successful in 48s
Publish Library / Tag Version (push) Successful in 13s
2026-08-04 13:58:47 -04:00
12 changed files with 1433 additions and 1287 deletions

View File

@@ -119,7 +119,7 @@ const ai = new Ai({
system: 'You are a helpful assistant.',
compress: {max: 90_000, min: 50_000}, // Compress chat history to min tokens when max is reached
temperature: 0.8,
max_tokens: 100_000,
maxTokens: 100_000,
memoryModel: 'gpt-4o', // Cheap model for managing memories in background, defaults to current model
models: {
'claude-3-5-sonnet': {proto: 'anthropic', token: process.env.ANTHROPIC_TOKEN},
@@ -186,7 +186,7 @@ console.log(chunks);
// Manually compile history into memories at end of conversation
// Happens automatically when coverstaions are compressed
await ai.language.updateMemory(history, memory);
await ai.language.memorize(history, memory);
// Summarize text
const summary = await ai.language.summarize(longText, 200);

526
package-lock.json generated
View File

@@ -1,21 +1,22 @@
{
"name": "@ztimson/ai-utils",
"version": "1.2.6",
"version": "1.6.6",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "@ztimson/ai-utils",
"version": "1.2.6",
"version": "1.6.6",
"license": "MIT",
"dependencies": {
"@anthropic-ai/sdk": "^0.102.0",
"@huggingface/transformers": "^4.2.0",
"@tensorflow/tfjs": "^4.22.0",
"@ztimson/node-utils": "^1.0.7",
"@ztimson/utils": "^0.29.4",
"@ztimson/utils": "^0.30.8",
"cheerio": "^1.2.0",
"openai": "^6.42.0",
"pdf-parse": "^2.4.5",
"tesseract.js": "^7.0.0"
},
"devDependencies": {
@@ -56,39 +57,12 @@
"node": ">=6.9.0"
}
},
"node_modules/@emnapi/core": {
"version": "2.0.0-alpha.3",
"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-2.0.0-alpha.3.tgz",
"integrity": "sha512-AZypUeJ/yByuxyS7BlSNRDOMLMlROYtjYdIAuBmJssVz1UJDSeYxLrdizhXCFYhedC5bqd/ASy8EuNXbVVXp9g==",
"dev": true,
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"@emnapi/wasi-threads": "2.0.1",
"tslib": "^2.4.0"
}
},
"node_modules/@emnapi/runtime": {
"version": "2.0.0-alpha.3",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-2.0.0-alpha.3.tgz",
"integrity": "sha512-hFPAhMUjJD9BSyCANEISPOogeXC9Zo9ZQl7L6vKnaVsMkCtzznaW/naYypeyl0Gv5rYfWYsZbpixTMpjDJzQeA==",
"dev": true,
"version": "1.11.3",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.3.tgz",
"integrity": "sha512-Xz4Tpyki7XyrpbUK1jR1AhdAdaXyhhY4lZ3neLodmhpuWfy2PAQN5B46sAiU4liOXGLkHypn/qU+jvfWSCYYLA==",
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@emnapi/wasi-threads": {
"version": "2.0.1",
"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-2.0.1.tgz",
"integrity": "sha512-9DsSk+o5NBX0CCJT8s0EROGSGxjR/tKu6aBTaVyq+SjAEQH4XcdcRxPBRzsBLizTTJ49MJjF+jgu3qnO9GLQcQ==",
"dev": true,
"license": "MIT",
"optional": true,
"peer": true,
"dependencies": {
"tslib": "^2.4.0"
}
@@ -577,16 +551,6 @@
"url": "https://opencollective.com/libvips"
}
},
"node_modules/@img/sharp-wasm32/node_modules/@emnapi/runtime": {
"version": "1.11.3",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.3.tgz",
"integrity": "sha512-Xz4Tpyki7XyrpbUK1jR1AhdAdaXyhhY4lZ3neLodmhpuWfy2PAQN5B46sAiU4liOXGLkHypn/qU+jvfWSCYYLA==",
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@img/sharp-win32-arm64": {
"version": "0.34.5",
"resolved": "https://registry.npmjs.org/@img/sharp-win32-arm64/-/sharp-win32-arm64-0.34.5.tgz",
@@ -694,32 +658,209 @@
"@jridgewell/sourcemap-codec": "^1.4.14"
}
},
"node_modules/@napi-rs/wasm-runtime": {
"version": "1.2.0",
"resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.2.0.tgz",
"integrity": "sha512-kDoONqMa+VnZ4vvvu/ZUurpJ4gkZU57e7g69qpNgWhYcZFPUHZM2CEMKm+cG6ufDVALbjMvfmMjFVqaK7uEMnA==",
"dev": true,
"node_modules/@napi-rs/canvas": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas/-/canvas-0.1.80.tgz",
"integrity": "sha512-DxuT1ClnIPts1kQx8FBmkk4BQDTfI5kIzywAaMjQSXfNnra5UFU9PwurXrl+Je3bJ6BGsp/zmshVVFbCmyI+ww==",
"license": "MIT",
"workspaces": [
"e2e/*"
],
"engines": {
"node": ">= 10"
},
"optionalDependencies": {
"@napi-rs/canvas-android-arm64": "0.1.80",
"@napi-rs/canvas-darwin-arm64": "0.1.80",
"@napi-rs/canvas-darwin-x64": "0.1.80",
"@napi-rs/canvas-linux-arm-gnueabihf": "0.1.80",
"@napi-rs/canvas-linux-arm64-gnu": "0.1.80",
"@napi-rs/canvas-linux-arm64-musl": "0.1.80",
"@napi-rs/canvas-linux-riscv64-gnu": "0.1.80",
"@napi-rs/canvas-linux-x64-gnu": "0.1.80",
"@napi-rs/canvas-linux-x64-musl": "0.1.80",
"@napi-rs/canvas-win32-x64-msvc": "0.1.80"
}
},
"node_modules/@napi-rs/canvas-android-arm64": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-android-arm64/-/canvas-android-arm64-0.1.80.tgz",
"integrity": "sha512-sk7xhN/MoXeuExlggf91pNziBxLPVUqF2CAVnB57KLG/pz7+U5TKG8eXdc3pm0d7Od0WreB6ZKLj37sX9muGOQ==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"dependencies": {
"@tybys/wasm-util": "^0.10.3"
},
"os": [
"android"
],
"engines": {
"node": "^20.19.0 || ^22.13.0 || >=23.5.0"
"node": ">= 10"
}
},
"funding": {
"type": "github",
"url": "https://github.com/sponsors/Brooooooklyn"
"node_modules/@napi-rs/canvas-darwin-arm64": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-arm64/-/canvas-darwin-arm64-0.1.80.tgz",
"integrity": "sha512-O64APRTXRUiAz0P8gErkfEr3lipLJgM6pjATwavZ22ebhjYl/SUbpgM0xcWPQBNMP1n29afAC/Us5PX1vg+JNQ==",
"cpu": [
"arm64"
],
"license": "MIT",
"optional": true,
"os": [
"darwin"
],
"engines": {
"node": ">= 10"
}
},
"peerDependencies": {
"@emnapi/core": "^2.0.0-alpha.3",
"@emnapi/runtime": "^2.0.0-alpha.3"
"node_modules/@napi-rs/canvas-darwin-x64": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-darwin-x64/-/canvas-darwin-x64-0.1.80.tgz",
"integrity": "sha512-FqqSU7qFce0Cp3pwnTjVkKjjOtxMqRe6lmINxpIZYaZNnVI0H5FtsaraZJ36SiTHNjZlUB69/HhxNDT1Aaa9vA==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"darwin"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-arm-gnueabihf": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm-gnueabihf/-/canvas-linux-arm-gnueabihf-0.1.80.tgz",
"integrity": "sha512-eyWz0ddBDQc7/JbAtY4OtZ5SpK8tR4JsCYEZjCE3dI8pqoWUC8oMwYSBGCYfsx2w47cQgQCgMVRVTFiiO38hHQ==",
"cpu": [
"arm"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-arm64-gnu": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-gnu/-/canvas-linux-arm64-gnu-0.1.80.tgz",
"integrity": "sha512-qwA63t8A86bnxhuA/GwOkK3jvb+XTQaTiVML0vAWoHyoZYTjNs7BzoOONDgTnNtr8/yHrq64XXzUoLqDzU+Uuw==",
"cpu": [
"arm64"
],
"libc": [
"glibc"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-arm64-musl": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-arm64-musl/-/canvas-linux-arm64-musl-0.1.80.tgz",
"integrity": "sha512-1XbCOz/ymhj24lFaIXtWnwv/6eFHXDrjP0jYkc6iHQ9q8oXKzUX1Lc6bu+wuGiLhGh2GS/2JlfORC5ZcXimRcg==",
"cpu": [
"arm64"
],
"libc": [
"musl"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-riscv64-gnu": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-riscv64-gnu/-/canvas-linux-riscv64-gnu-0.1.80.tgz",
"integrity": "sha512-XTzR125w5ZMs0lJcxRlS1K3P5RaZ9RmUsPtd1uGt+EfDyYMu4c6SEROYsxyatbbu/2+lPe7MPHOO/0a0x7L/gw==",
"cpu": [
"riscv64"
],
"libc": [
"glibc"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-x64-gnu": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-gnu/-/canvas-linux-x64-gnu-0.1.80.tgz",
"integrity": "sha512-BeXAmhKg1kX3UCrJsYbdQd3hIMDH/K6HnP/pG2LuITaXhXBiNdh//TVVVVCBbJzVQaV5gK/4ZOCMrQW9mvuTqA==",
"cpu": [
"x64"
],
"libc": [
"glibc"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-linux-x64-musl": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-x64-musl/-/canvas-linux-x64-musl-0.1.80.tgz",
"integrity": "sha512-x0XvZWdHbkgdgucJsRxprX/4o4sEed7qo9rCQA9ugiS9qE2QvP0RIiEugtZhfLH3cyI+jIRFJHV4Fuz+1BHHMg==",
"cpu": [
"x64"
],
"libc": [
"musl"
],
"license": "MIT",
"optional": true,
"os": [
"linux"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@napi-rs/canvas-win32-x64-msvc": {
"version": "0.1.80",
"resolved": "https://registry.npmjs.org/@napi-rs/canvas-win32-x64-msvc/-/canvas-win32-x64-msvc-0.1.80.tgz",
"integrity": "sha512-Z8jPsM6df5V8B1HrCHB05+bDiCxjE9QA//3YrkKIdVDEwn5RKaqOxCJDRJkl48cJbylcrJbW4HxZbTte8juuPg==",
"cpu": [
"x64"
],
"license": "MIT",
"optional": true,
"os": [
"win32"
],
"engines": {
"node": ">= 10"
}
},
"node_modules/@oxc-project/types": {
"version": "0.139.0",
"resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.139.0.tgz",
"integrity": "sha512-r9gHphtCs+1M7J0pw6Sn/hh/Wpa/iQrOOkrNAlVLF/gHq+/CJmHIWKKUUhdWjcD6CIa8idarspCsASiXCXvFUw==",
"version": "0.144.0",
"resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.144.0.tgz",
"integrity": "sha512-nuhZIOLuI6TFQ32I/WnUx+SCPY7SdSKwgnFHydAuoS1+Z4BRcaP+RRJmGzl9lw+0OFF7UmaESf7KQRXaNLHypg==",
"dev": true,
"license": "MIT",
"funding": {
@@ -784,9 +925,9 @@
"license": "BSD-3-Clause"
},
"node_modules/@rolldown/binding-android-arm64": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.1.5.tgz",
"integrity": "sha512-lZg8fqIv2v7FF237bwMgzGZEJvGL79/s5knJ/i6FmsGF4XXlzccZ4jb+TrFIxtSSxFtIpdsgrPZeMk1I9AFcyQ==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.2.4.tgz",
"integrity": "sha512-jHC2cnyKz5xU2fhECtFl8OZ83cYNt13GZQD+0uMJ/X3o+ijmd56okHhTUwxVSHPx1IRVIJEZ1/1pPzeLCU6XKA==",
"cpu": [
"arm64"
],
@@ -801,9 +942,9 @@
}
},
"node_modules/@rolldown/binding-darwin-arm64": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.1.5.tgz",
"integrity": "sha512-51Bnx9pNiMRKSUNtBfySkNJ9vMU9Hh3I1ozDd6gyPPYzaXCfnptUcEZxXGYFn+ul2dtcMUiqGR1Yai2K10uoTw==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.2.4.tgz",
"integrity": "sha512-Dc5mPD8F5F/FS8i01syd7FTF6yB2fVthH/TRkjwJkzUK6EpoxHtqvZQP5Zwq80/5z19TWYHIg1KOHboCgVx/aQ==",
"cpu": [
"arm64"
],
@@ -818,9 +959,9 @@
}
},
"node_modules/@rolldown/binding-darwin-x64": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.1.5.tgz",
"integrity": "sha512-Tm+gbfC0aHu1tBA/JvKQh32S0K6YgCHkiAF4/W6xX0K0RmNuc94VeK419dJoE65R5aRxmo+noZQSWrAMF6yb6g==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.2.4.tgz",
"integrity": "sha512-fpDm4oBo6SqLvWUYCmFhdde3U9KH2fRNNMeAnAPAIwxRL345xutL0EtEUcuoxsoazdJGv/MuDBQHlCDrtbvqOg==",
"cpu": [
"x64"
],
@@ -835,9 +976,9 @@
}
},
"node_modules/@rolldown/binding-freebsd-x64": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.1.5.tgz",
"integrity": "sha512-JMzDKCCXq93YccG5gz3hvOs1oXRKAf0XYpfOS88e+wZrC8Iugj6j68867vrYZkvpDDpKn/KoKORThmchMpF6TA==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.2.4.tgz",
"integrity": "sha512-rSJoreDE/HoIzoaib6MTp5jQtCTdMHKIvItAKT/ImS6Y6Ww76oUaeMyp4Vc/fAgd/ehji068IxetHXAnqUwN9A==",
"cpu": [
"x64"
],
@@ -852,9 +993,9 @@
}
},
"node_modules/@rolldown/binding-linux-arm-gnueabihf": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.1.5.tgz",
"integrity": "sha512-uML21j2K5TfPGutKxub+M+nLjZIrWjXQ5Grx4lCe/nimTj9B4L63zHpjXLl4y0L3mcm2htEQIb06oCG/szerNw==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.2.4.tgz",
"integrity": "sha512-/jm8OGHgn7oGaJu3i/qZI9spUGcJ+y/lk43ttQ/iO1tOd9NissG6o97bighBCiL+BKRngmcDuR6ikfwYdJmVuQ==",
"cpu": [
"arm"
],
@@ -869,9 +1010,9 @@
}
},
"node_modules/@rolldown/binding-linux-arm64-gnu": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.1.5.tgz",
"integrity": "sha512-navSiuTMogvnQoZoM/v+l3ZWo50/NTwSHSzheABx/RCnmUPaKwq9qSo4Br2OYRs21+Fz8uFqITZM3H4opOB0/Q==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.2.4.tgz",
"integrity": "sha512-tIP06BeD9EqvECBrPZ+sqdPlYrT+aYaAiu1wYziVx5elRK/ftm33JxVDy2bXGbr6J0CrtirCkR87/X5a2euEng==",
"cpu": [
"arm64"
],
@@ -889,9 +1030,9 @@
}
},
"node_modules/@rolldown/binding-linux-arm64-musl": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.1.5.tgz",
"integrity": "sha512-lAryqH7IteztmCXQXk0etKj4wBQ7Gx5S6LjKhsgp9zb8I5bsuvU/2llH1hDQcjsFeqIsovMVN339/8pUDDBXxA==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.2.4.tgz",
"integrity": "sha512-Ql1Q0EQqVThvn9VAVlwNzsUvbSFtCMGjLpRRi4pk5i7NZZ4n5ISiLMjHYtus4VQ2PvkSw24zyaCVsiS+sXPj1w==",
"cpu": [
"arm64"
],
@@ -909,9 +1050,9 @@
}
},
"node_modules/@rolldown/binding-linux-ppc64-gnu": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.1.5.tgz",
"integrity": "sha512-fsK/sNBnxzBlL4O1JNrZakVQxPspqpED5dLtNsZS9oOKmtSpdNIzxH2kkol5HYTWJN47sE20ztMJPxfZ89qGOg==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.2.4.tgz",
"integrity": "sha512-GjbjXD4XXfN19D0LZNbmiCBUoDiRACsYHr0yaIbbn8aFsXjHZifcYqu/W5Er5X2X990WjHXFrxarn5chzItorQ==",
"cpu": [
"ppc64"
],
@@ -929,9 +1070,9 @@
}
},
"node_modules/@rolldown/binding-linux-s390x-gnu": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.1.5.tgz",
"integrity": "sha512-gLYb4BIadlfTOYT5gO503n8zQjXflgzpD0FcyKh0Mzx3rqCZKnHoJWV9xe1KXUJ5lx2JfcSHr/mhzS0PC/McAA==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.2.4.tgz",
"integrity": "sha512-p5WR0NOwaRmJ/B1b6IjEFLLivwEsf3PrdBIhRbhTCQisbo2SvHHpG4ELB/+FgQNnB88LTOF86upmJmbvZdQ2lw==",
"cpu": [
"s390x"
],
@@ -949,9 +1090,9 @@
}
},
"node_modules/@rolldown/binding-linux-x64-gnu": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.1.5.tgz",
"integrity": "sha512-FjcpEKUyJygHgs1o50VYNvkt5+7Le/VEdYt0AkRpkL33MnyQfwr8l5mXwMmfmTbyMPr5vJLC+8/Gd9gXnwU1QQ==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.2.4.tgz",
"integrity": "sha512-4/GyVjmhR+Tc6HLJvwc1sOhPqAZtySiSMesOZyX6JQ5XBxoTDEMKQzvo07NIK6nTon/SivlZqvhzvuVBNQhObQ==",
"cpu": [
"x64"
],
@@ -969,9 +1110,9 @@
}
},
"node_modules/@rolldown/binding-linux-x64-musl": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.1.5.tgz",
"integrity": "sha512-Me+PfPI2TMeOQk0gYWfLQZtTktrmzbr8cDboqX83XKc7UrgAi55gF+2dUkWdxd19n55Essp2yeca+O9N5rBxHg==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.2.4.tgz",
"integrity": "sha512-l9eeLsCNvPpmSXUej0etw/J1eqV0Jj1D5G/xG6YTijmE6dkv6E2QezgWbTfQk63v952DPqrjOCoiqxq7Bw0YUQ==",
"cpu": [
"x64"
],
@@ -989,9 +1130,9 @@
}
},
"node_modules/@rolldown/binding-openharmony-arm64": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.1.5.tgz",
"integrity": "sha512-yc5WrLzXks6zCQfn9Oxr8pORKyl/pF+QjHmW/Qx3qu0oyrrNC+y2JLTU1E2rcWYAmzlnqngWXHQjy51VzW70Vw==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.2.4.tgz",
"integrity": "sha512-e0F355MSTMm3+UOqtV3L24gFUp2N5m1f8L/7d56deik6va+AXdrt9F8LbzGpeWGWRbZEDq4m8NVnJDeBtf9DZg==",
"cpu": [
"arm64"
],
@@ -1005,63 +1146,10 @@
"node": "^20.19.0 || >=22.12.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.1.5.tgz",
"integrity": "sha512-VbQGPX2b4r48TAMIM2cjgluIM1HYutm4pcTEJsle7iEP7sB1dFqtPLBVbdLAZCxy1txCcPxf4QFf4v8uvltPqA==",
"cpu": [
"wasm32"
],
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"@emnapi/core": "1.11.1",
"@emnapi/runtime": "1.11.1",
"@napi-rs/wasm-runtime": "^1.1.6"
},
"engines": {
"node": "^20.19.0 || >=22.12.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/core": {
"version": "1.11.1",
"resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.11.1.tgz",
"integrity": "sha512-RSvbQmHzdKzNsLYa/wHrbc3KN4sYLKAdPZxqiM2HATqv/SBk2/ENSHpvXGaLOMcsAyz0poEGqkmmKYG3OWiJEQ==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"@emnapi/wasi-threads": "1.2.2",
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/runtime": {
"version": "1.11.1",
"resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.11.1.tgz",
"integrity": "sha512-vgj7R3y3Wgx24IQaGPA/R6YFXLHVMOZ0uVEyIQPaWs+rd1AzfEMXlAC22FYwO1XkKR6NPsq7mUandH8oIRdZFw==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
"version": "1.2.2",
"resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.2.tgz",
"integrity": "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@rolldown/binding-win32-arm64-msvc": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.1.5.tgz",
"integrity": "sha512-gHv82k63z4qpV5+Q1y/12KrK0ltWBukVDI8nZcbT7Tt/ZlOIVwppazneq0F93oDxTo3IgAMEDIoQh3E2n6mVsw==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.2.4.tgz",
"integrity": "sha512-AWLi0uBRYh6QlE7OKhiz+phZC0qwtij2QZmhmOdsLdFn64m7oMpooE9ICE3lhm9xMb4SpDo2WbHcxX1iFLFtqw==",
"cpu": [
"arm64"
],
@@ -1076,9 +1164,9 @@
}
},
"node_modules/@rolldown/binding-win32-x64-msvc": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.1.5.tgz",
"integrity": "sha512-tTZuDBPw85tEN5PQi1pnEBzDy0Z49HtScLAbD5t6hyeU92A95pRWaSMw1GZZi/RwgSgUIl0xrSlXIT/9QzvYSA==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.2.4.tgz",
"integrity": "sha512-UwSDJOg3dqCAejWdxclJjCsh3Qq4vLYMDxmyHqo1btz3stK2VqgwNd3mm5tuIwzSlGIQ/1H9Hr+Zn09mrezNqQ==",
"cpu": [
"x64"
],
@@ -1317,17 +1405,6 @@
"@tensorflow/tfjs-core": "4.22.0"
}
},
"node_modules/@tybys/wasm-util": {
"version": "0.10.3",
"resolved": "https://registry.npmjs.org/@tybys/wasm-util/-/wasm-util-0.10.3.tgz",
"integrity": "sha512-F3fo1MYrRJYL3zER0OUOmkutjr1Vp23m7OsSgp7nq4SP6OqX6C/56XFIPAl5bt3zaBRjmW7SGz3u/6LwFpYcOg==",
"dev": true,
"license": "MIT",
"optional": true,
"dependencies": {
"tslib": "^2.4.0"
}
},
"node_modules/@types/estree": {
"version": "1.0.9",
"resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.9.tgz",
@@ -1448,9 +1525,9 @@
"license": "MIT"
},
"node_modules/@ztimson/utils": {
"version": "0.29.7",
"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.29.7.tgz",
"integrity": "sha512-cjQ9+RjC5X7gKNA/hJHDf7OtyYCa+5E0PDc76lIaATwNAxXCSx2IO9r2wHiHtZGV5bldjnFmw7aOV8Jmq7SgKQ==",
"version": "0.30.8",
"resolved": "https://registry.npmjs.org/@ztimson/utils/-/utils-0.30.8.tgz",
"integrity": "sha512-+vBjcinqckqMHkP95xWiQeQz2E7Q1oS0b+Odjp+F9rvQ4z0US4JdodjqhEaqh+RAO/yP77x4Eu0a04yBB4HNdw==",
"license": "MIT",
"dependencies": {
"var-persist": "^1.0.1"
@@ -1544,9 +1621,9 @@
"license": "MIT"
},
"node_modules/brace-expansion": {
"version": "2.1.3",
"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.3.tgz",
"integrity": "sha512-DRdx5neNsG/QXbniLFWi2YmC/68oeOOmKz6zOjVk6ZS1ZLXgLIKqVEc6hWsmkjBbgii0SwaBTcJ5XKj5gzY/4A==",
"version": "2.1.4",
"resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.4.tgz",
"integrity": "sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg==",
"dev": true,
"license": "MIT",
"dependencies": {
@@ -3011,9 +3088,9 @@
"license": "MIT"
},
"node_modules/nanoid": {
"version": "3.3.16",
"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.16.tgz",
"integrity": "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q==",
"version": "3.3.18",
"resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.18.tgz",
"integrity": "sha512-DTg4MJbGMWkfi6VZFdNt2/caMbQy4Ou+Op/hJQvGEWcnVfoA1QA+xzRKAzw9jD6+GVOOeYr/mIcuDSdug6F6+w==",
"dev": true,
"funding": [
{
@@ -3233,6 +3310,38 @@
"dev": true,
"license": "MIT"
},
"node_modules/pdf-parse": {
"version": "2.4.5",
"resolved": "https://registry.npmjs.org/pdf-parse/-/pdf-parse-2.4.5.tgz",
"integrity": "sha512-mHU89HGh7v+4u2ubfnevJ03lmPgQ5WU4CxAVmTSh/sxVTEDYd1er/dKS/A6vg77NX47KTEoihq8jZBLr8Cxuwg==",
"license": "Apache-2.0",
"dependencies": {
"@napi-rs/canvas": "0.1.80",
"pdfjs-dist": "5.4.296"
},
"bin": {
"pdf-parse": "bin/cli.mjs"
},
"engines": {
"node": ">=20.16.0 <21 || >=22.3.0"
},
"funding": {
"type": "github",
"url": "https://github.com/sponsors/mehmet-kozan"
}
},
"node_modules/pdfjs-dist": {
"version": "5.4.296",
"resolved": "https://registry.npmjs.org/pdfjs-dist/-/pdfjs-dist-5.4.296.tgz",
"integrity": "sha512-DlOzet0HO7OEnmUmB6wWGJrrdvbyJKftI1bhMitK7O2N8W2gc757yyYBbINy9IDafXAV9wmKr9t7xsTaNKRG5Q==",
"license": "Apache-2.0",
"engines": {
"node": ">=20.16.0 || >=22.3.0"
},
"optionalDependencies": {
"@napi-rs/canvas": "^0.1.80"
}
},
"node_modules/picocolors": {
"version": "1.1.1",
"resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz",
@@ -3272,9 +3381,9 @@
"license": "MIT"
},
"node_modules/postcss": {
"version": "8.5.25",
"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.25.tgz",
"integrity": "sha512-DTPx3RWSSnWyzLxQnlH0rJP+EW5ekl16ZU4/psbIhA0e53kJfdgaN5vKM+xP7yJtXVu+nfdVFmlgFDEKAe4Pyw==",
"version": "8.5.26",
"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.26.tgz",
"integrity": "sha512-u82N74LFzG8ca+dD8puPnplTXoGH4fTPpVGuIbt36G3qvNlkvfD0lEAZSxaly3KX8TS/L1A1gsCEmvKmBcVbkQ==",
"dev": true,
"funding": [
{
@@ -3292,7 +3401,7 @@
],
"license": "MIT",
"dependencies": {
"nanoid": "^3.3.16",
"nanoid": "^3.3.17",
"picocolors": "^1.1.1",
"source-map-js": "^1.2.1"
},
@@ -3434,13 +3543,13 @@
"license": "BSD-3-Clause"
},
"node_modules/rolldown": {
"version": "1.1.5",
"resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.1.5.tgz",
"integrity": "sha512-t9z29cJjXf/vxQ8dyhCSpt6H6aSwHTk8cT5I3iy6SMXuFpk5mB6PL6XfC8PCwrPTx93udwKUm9HRteAlTGBLiA==",
"version": "1.2.4",
"resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.2.4.tgz",
"integrity": "sha512-rSr7irW0K7QRWzjdJXqZowkcRdDtjRduh43rBltnVKd0VFq839l1lJoDvGJb6gl7+4rTTCrPWu+YfujUL8Ug7w==",
"dev": true,
"license": "MIT",
"dependencies": {
"@oxc-project/types": "=0.139.0",
"@oxc-project/types": "=0.144.0",
"@rolldown/pluginutils": "^1.0.0"
},
"bin": {
@@ -3450,21 +3559,20 @@
"node": "^20.19.0 || >=22.12.0"
},
"optionalDependencies": {
"@rolldown/binding-android-arm64": "1.1.5",
"@rolldown/binding-darwin-arm64": "1.1.5",
"@rolldown/binding-darwin-x64": "1.1.5",
"@rolldown/binding-freebsd-x64": "1.1.5",
"@rolldown/binding-linux-arm-gnueabihf": "1.1.5",
"@rolldown/binding-linux-arm64-gnu": "1.1.5",
"@rolldown/binding-linux-arm64-musl": "1.1.5",
"@rolldown/binding-linux-ppc64-gnu": "1.1.5",
"@rolldown/binding-linux-s390x-gnu": "1.1.5",
"@rolldown/binding-linux-x64-gnu": "1.1.5",
"@rolldown/binding-linux-x64-musl": "1.1.5",
"@rolldown/binding-openharmony-arm64": "1.1.5",
"@rolldown/binding-wasm32-wasi": "1.1.5",
"@rolldown/binding-win32-arm64-msvc": "1.1.5",
"@rolldown/binding-win32-x64-msvc": "1.1.5"
"@rolldown/binding-android-arm64": "1.2.4",
"@rolldown/binding-darwin-arm64": "1.2.4",
"@rolldown/binding-darwin-x64": "1.2.4",
"@rolldown/binding-freebsd-x64": "1.2.4",
"@rolldown/binding-linux-arm-gnueabihf": "1.2.4",
"@rolldown/binding-linux-arm64-gnu": "1.2.4",
"@rolldown/binding-linux-arm64-musl": "1.2.4",
"@rolldown/binding-linux-ppc64-gnu": "1.2.4",
"@rolldown/binding-linux-s390x-gnu": "1.2.4",
"@rolldown/binding-linux-x64-gnu": "1.2.4",
"@rolldown/binding-linux-x64-musl": "1.2.4",
"@rolldown/binding-openharmony-arm64": "1.2.4",
"@rolldown/binding-win32-arm64-msvc": "1.2.4",
"@rolldown/binding-win32-x64-msvc": "1.2.4"
}
},
"node_modules/safe-buffer": {
@@ -4021,16 +4129,16 @@
}
},
"node_modules/vite": {
"version": "8.1.5",
"resolved": "https://registry.npmjs.org/vite/-/vite-8.1.5.tgz",
"integrity": "sha512-7ULLwsCdYx/nRyrpiEwvqb5TFHrMVZyBt+rg/OAXT7rgj/z+DtTDyKFeLAdDkubDVDKD8jOsndmy7m55XcfUsw==",
"version": "8.2.1",
"resolved": "https://registry.npmjs.org/vite/-/vite-8.2.1.tgz",
"integrity": "sha512-EU/eS7BH3XROHh2YnBefjM6DBKA6ZeMZEYQbj7NLWg5wHYlhB8B/Mayd5XsgWq+NFYccDOTemRpdETWR6Ka/lw==",
"dev": true,
"license": "MIT",
"dependencies": {
"lightningcss": "^1.32.0",
"lightningcss": "^1.33.0",
"picomatch": "^4.0.5",
"postcss": "^8.5.17",
"rolldown": "~1.1.5",
"postcss": "^8.5.25",
"rolldown": "~1.2.1",
"tinyglobby": "^0.2.17"
},
"bin": {
@@ -4047,7 +4155,7 @@
},
"peerDependencies": {
"@types/node": "^20.19.0 || >=22.12.0",
"@vitejs/devtools": "^0.3.0",
"@vitejs/devtools": "^0.4.0",
"esbuild": "^0.27.0 || ^0.28.0",
"jiti": ">=1.21.0",
"less": "^4.0.0",
@@ -4132,9 +4240,9 @@
"license": "MIT"
},
"node_modules/wasm-feature-detect": {
"version": "1.8.0",
"resolved": "https://registry.npmjs.org/wasm-feature-detect/-/wasm-feature-detect-1.8.0.tgz",
"integrity": "sha512-zksaLKM2fVlnB5jQQDqKXXwYHLQUVH9es+5TOOHwGOVJOCeRBCiPjwSg+3tN2AdTCzjgli4jijCH290kXb/zWQ==",
"version": "1.9.0",
"resolved": "https://registry.npmjs.org/wasm-feature-detect/-/wasm-feature-detect-1.9.0.tgz",
"integrity": "sha512-zonE+xlIIYtxPy++L24ow0hAD8CICb4+FgPyROd3buyXIqsJvUEDkBgfCCoXOd1Hu3DUr0GOfnPIdcGV+YpNaA==",
"license": "Apache-2.0"
},
"node_modules/webidl-conversions": {

View File

@@ -1,6 +1,6 @@
{
"name": "@ztimson/ai-utils",
"version": "1.4.1",
"version": "1.6.7",
"description": "AI Utility library",
"author": "Zak Timson",
"license": "MIT",
@@ -26,12 +26,13 @@
},
"dependencies": {
"@anthropic-ai/sdk": "^0.102.0",
"@tensorflow/tfjs": "^4.22.0",
"@huggingface/transformers": "^4.2.0",
"@tensorflow/tfjs": "^4.22.0",
"@ztimson/node-utils": "^1.0.7",
"@ztimson/utils": "^0.29.4",
"@ztimson/utils": "^0.30.8",
"cheerio": "^1.2.0",
"openai": "^6.42.0",
"pdf-parse": "^2.4.5",
"tesseract.js": "^7.0.0"
},
"devDependencies": {

View File

@@ -4,7 +4,7 @@ import { Audio } from './audio.ts';
import {Vision} from './vision.ts';
export type AbortablePromise<T> = Promise<T> & {
abort: () => any
abort: (keep?: boolean) => any
};
export type AiOptions = {

View File

@@ -24,53 +24,40 @@ export class Anthropic extends LLMProvider {
return client;
}
private toStandard(history: any[]): LLMMessage[] {
const timestamp = Date.now();
const messages: LLMMessage[] = [];
for(let h of history) {
if(typeof h.content == 'string') {
messages.push(<any>{timestamp, ...h});
} else {
const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
if(textContent) messages.push({role: h.role, content: textContent, timestamp: timestamp, duration: h.duration, tps: h.tps});
h.content.forEach((c: any) => {
if(c.type == 'tool_use') {
messages.push({role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: h.timestamp, content: undefined, duration: h.duration, tps: h.tps});
} else if(c.type == 'tool_result') {
const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
if(m) m[c.is_error ? 'error' : 'content'] = c.content;
}
});
}
}
return messages;
private toWireContent(content: any): any {
if(!Array.isArray(content)) return content;
return content.map(c => c.type === 'image'
? {type: 'image', source: {type: 'base64', media_type: c.mime, data: c.data}}
: {type: 'text', text: c.text});
}
private fromStandard(history: LLMMessage[]): any[] {
for(let i = 0; i < history.length; i++) {
if(history[i].role == 'tool') {
const h: any = history[i];
history.splice(i, 1,
/** Convert standard history -> Anthropic wire format */
private toWire(history: LLMMessage[]): any[] {
const wire: any[] = [];
for(const h of history) {
if(h.role === 'tool') {
wire.push(
{role: 'assistant', content: [{type: 'tool_use', id: h.id, name: h.name, input: h.args}]},
{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content}]}
)
i++;
{role: 'user', content: [{type: 'tool_result', tool_use_id: h.id, is_error: !!h.error, content: h.error || h.content || ''}]}
);
} else {
wire.push({role: h.role, content: this.toWireContent(h.content)});
}
}
return history;
return wire;
}
ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
const controller = new AbortController();
return Object.assign(new Promise<any>(async (res) => {
let history = this.fromStandard([
...(options.history || []).filter(h => h.role !== 'system'),
{role: 'user', content: message, timestamp: Date.now()}
]);
return Object.assign(new Promise<any>(async (res, rej) => {
if(!options.history) options.history = [];
const history = options.history;
if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
const tools = options.tools || this.ai.options.llm?.tools || [];
const requestParams: any = {
model: options.model || this.model,
max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
max_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || 4096,
system: options.system || this.ai.options.llm?.system || '',
temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
tools: tools.map(t => ({
@@ -80,54 +67,43 @@ export class Anthropic extends LLMProvider {
type: 'object',
properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
},
fn: undefined
}
})),
messages: history,
stream: !!options.stream,
};
// Add structured output support
if(options.schema) {
requestParams.output_config = {
format: {
type: 'json_schema',
schema: convertSchema(options.schema)
}
};
requestParams.output_config = {format: {type: 'json_schema', schema: convertSchema(options.schema)}};
}
let resp: any, terminal = false, duration = 0, tps = 0;
try {
let terminal = false;
do {
requestParams.messages = history.map(({timestamp, ...m}) => m);
requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'));
const callStart = Date.now();
resp = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
const resp: any = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
throw err;
});
let usage: any;
let usage: any, content: any[] = [];
if(options.stream) {
resp.content = [];
for await (const chunk of resp) {
if(controller.signal.aborted) break;
if(chunk.type === 'content_block_start') {
if(chunk.content_block.type === 'text') {
resp.content.push({type: 'text', text: ''});
} else if(chunk.content_block.type === 'tool_use') {
resp.content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: <any>''});
}
if(chunk.content_block.type === 'text') content.push({type: 'text', text: ''});
else if(chunk.content_block.type === 'tool_use') content.push({type: 'tool_use', id: chunk.content_block.id, name: chunk.content_block.name, input: ''});
} else if(chunk.type === 'content_block_delta') {
if(chunk.delta.type === 'text_delta') {
const text = chunk.delta.text;
resp.content.at(-1).text += text;
options.stream({text});
content.at(-1).text += chunk.delta.text;
options.stream({text: chunk.delta.text});
} else if(chunk.delta.type === 'input_json_delta') {
resp.content.at(-1).input += chunk.delta.partial_json;
content.at(-1).input += chunk.delta.partial_json;
}
} else if(chunk.type === 'content_block_stop') {
const last = resp.content.at(-1);
if(last?.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
const last = content.at(-1);
if(last?.type === 'tool_use') last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
} else if(chunk.type === 'message_delta') {
if(chunk.usage) usage = chunk.usage;
} else if(chunk.type === 'message_stop') {
@@ -136,49 +112,52 @@ export class Anthropic extends LLMProvider {
}
} else {
usage = resp.usage;
content = resp.content;
}
duration = Date.now() - callStart;
tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
const duration = Date.now() - callStart;
const tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
const toolCalls = content.filter((c: any) => c.type === 'tool_use');
if(toolCalls.length && !controller.signal.aborted) {
history.push({role: 'assistant', content: resp.content, timestamp: Date.now(), duration, tps});
const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
const tool = tools.find(findByProp('name', toolCall.name));
if(options.stream) options.stream({tool: toolCall.name});
if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
const entries = toolCalls.map((tc: any) => {
const entry: any = {role: 'tool', id: tc.id, name: tc.name, args: tc.input, content: undefined, timestamp: Date.now()};
history.push(entry);
return {tc, entry};
});
await Promise.all(entries.map(async ({tc, entry}: any) => {
const tool = tools.find(findByProp('name', tc.name));
if(options.stream) options.stream({tool: tc.name});
if(!tool) { entry.error = 'Tool not found'; return; }
try {
const toolStream = options.stream && ((chunk: any) => {
if(chunk.done) { terminal = true; return; }
options.stream!(chunk);
});
const result = await tool.fn(toolCall.input, toolStream, this.ai, toolCall.id);
return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
} catch(err: any) {
return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
entry.error = err?.message || err?.toString() || 'Unknown';
}
}));
history.push({role: 'user', content: results, timestamp: Date.now()});
requestParams.messages = history;
} else {
terminal = true;
const text = content.filter((c: any) => c.type === 'text').map((c: any) => c.text).join('\n\n').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
}
} while (!terminal && !controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
} while(!terminal && !controller.signal.aborted);
if(!terminal) {
const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
}
history = this.toStandard(history);
if(options.history) options.history.splice(0, options.history.length, ...history);
if(options.stream) options.stream({done: true});
const turnStart = history.map(h => h.role).lastIndexOf('user');
const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
if(h.role === 'assistant') return str + (h.content || '');
return str;
}, '').trim();
const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
} catch(err) {
rej(err);
}
}), {abort: () => controller.abort()});
}
}

View File

@@ -7,10 +7,74 @@ export type MemoryNode = {
backlinks: string[];
}
export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
export function extractLinks(content: string): string[] {
if (!content) return [];
const matches = content.matchAll(/\[\[([^\]|]+)(?:\|[^\]]*)?\]\]/g);
return [...new Set([...matches].map(m => m[1].trim()))];
}
/**
* Incrementally patch the graph for a set of changed memories, instead of
* re-scanning every document. Only the changed memories' own content is
* re-parsed for links; affected targets have their backlinks patched.
* Does NOT handle node deletion — full rebuildGraph() is still required
* when a memory is removed, since that needs a backlink sweep across
* everyone who might reference it.
*/
export function patchGraph(mems: Memory[], nodes: MemoryNode[], changed: Memory[]): MemoryNode[] {
const nameSet = new Set(mems.map(m => m.name));
const byName = new Map(nodes.map(n => [n.name, n]));
const ensureNode = (name: string): MemoryNode => {
let n = byName.get(name);
if (!n) {
n = {name, missing: !nameSet.has(name), links: [], backlinks: []};
byName.set(name, n);
}
return n;
};
for (const m of changed) {
const node = ensureNode(m.name);
node.missing = false; // real memory, promotes any pre-existing ghost entry
const oldLinks = m.links ?? [];
const newLinks = extractLinks(m.content).filter(l => l !== m.name);
for (const target of oldLinks.filter(l => !newLinks.includes(l))) {
const t = byName.get(target);
if (!t) continue;
t.backlinks = t.backlinks.filter(n => n !== m.name);
if (t.missing && !t.backlinks.length) byName.delete(target); // fully dereferenced ghost
}
for (const target of newLinks.filter(l => !oldLinks.includes(l))) {
const t = ensureNode(target);
if (!t.backlinks.includes(m.name)) t.backlinks.push(m.name);
}
m.links = newLinks;
node.links = newLinks;
}
for (const m of mems) {
const n = byName.get(m.name);
if (n) m.backlinks = n.backlinks;
}
return [...byName.values()];
}
export function rebuildGraph(memories: Memory[] | MemoryCache): MemoryNode[] {
const mems = memories instanceof MemoryCache ? memories.memories : memories;
const nameSet = new Set(mems.map(m => m.name));
const ghosts = new Set<string>();
for (const m of mems) m.links = extractLinks(m.content).filter(l => l !== m.name);
for (const m of mems) m.backlinks = [];
for (const m of mems) {
for (const link of m.links) {
const target = mems.find(t => t.name === link);
if (target) target.backlinks.push(m.name);
}
}
const nodes: MemoryNode[] = mems.map(m => ({
name: m.name,
@@ -19,6 +83,7 @@ export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[]
backlinks: m.backlinks,
}));
const ghosts = new Set<string>();
for (const node of nodes) {
for (const link of node.links) {
if (!nameSet.has(link)) ghosts.add(link);
@@ -31,30 +96,28 @@ export function buildMemoryGraph(memories: Memory[] | MemoryCache): MemoryNode[]
name,
missing: true,
links: [],
backlinks: nodes
.filter(n => n.links.includes(name))
.map(n => n.name),
}))
backlinks: nodes.filter(n => n.links.includes(name)).map(n => n.name),
})),
];
}
export function renderMemoryGraph(nodes) {
export function renderMemoryGraph(nodes: MemoryNode[]): string {
if (!nodes.length) return 'No memories yet.';
const groups = new Map();
const groups = new Map<string, (MemoryNode & {label: string})[]>();
for (const node of nodes) {
const [prefix, ...rest] = node.name.split('/');
const group = rest.length ? prefix : 'Root';
const label = rest.length ? rest.join('/') : node.name;
if (!groups.has(group)) groups.set(group, []);
groups.get(group).push({...node, label});
groups.get(group)!.push({...node, label});
}
const ghostCount = nodes.filter(n => n.missing).length;
const lines = [`Memory Graph (${nodes.length} nodes, ${ghostCount} ghost${ghostCount === 1 ? '' : 's'})`, ''];
for (const group of [...groups.keys()].sort()) {
const items = groups.get(group).sort((a, b) => a.label.localeCompare(b.label));
const items = groups.get(group)!.sort((a, b) => a.label.localeCompare(b.label));
lines.push(`${group}/`);
items.forEach((n, i) => {
const last = i === items.length - 1;

View File

@@ -15,6 +15,7 @@ interface KDNode<T> {
axis: number;
left: KDNode<T> | null;
right: KDNode<T> | null;
deleted?: boolean;
}
// ─── Distance helpers ─────────────────────────────────────────────────────────
@@ -95,6 +96,7 @@ class BoundedMaxHeap<T> {
*
* Supports:
* - Insertion of labeled points
* - Lazy (tombstone) removal, physically purged on rebalance()
* - k-nearest-neighbor (KNN) search
* - Radius search (all points within a given distance)
* - Euclidean and cosine distance metrics
@@ -103,9 +105,11 @@ class BoundedMaxHeap<T> {
export class KDTree<T = unknown> {
private root: KDNode<T> | null = null;
private _size = 0;
private readonly dims: number;
private _tombstones = 0;
private readonly distanceFn: (a: number[], b: number[]) => number;
readonly dims: number;
/**
* @param dims Dimensionality of all vectors (must be consistent).
* @param metric Distance metric to use. Default: "euclidean".
@@ -128,9 +132,15 @@ export class KDTree<T = unknown> {
}
}
/** Total number of points stored in the tree. */
/** Total number of live points stored in the tree (excludes tombstoned). */
get size(): number { return this._size; }
/** Fraction of physical nodes that are tombstoned (pending removal on next rebalance). */
get tombstoneRatio(): number {
const total = this._size + this._tombstones;
return total ? this._tombstones / total : 0;
}
// ── Insertion ──────────────────────────────────────────────────────────────
/**
@@ -143,10 +153,36 @@ export class KDTree<T = unknown> {
this._size++;
}
// ── Removal ────────────────────────────────────────────────────────────────
/**
* Lazily remove all live points whose payload matches `predicate`.
* O(n) traversal, but avoids a full tree rebuild. Call `rebalance()`
* periodically (e.g. once tombstoneRatio crosses ~0.25) to reclaim space
* and restore optimal query depth.
* @returns number of points removed
*/
remove(predicate: (payload: T) => boolean): number {
let removed = 0;
const visit = (node: KDNode<T> | null): void => {
if (!node) return;
if (!node.deleted && predicate(node.point.payload)) {
node.deleted = true;
removed++;
}
visit(node.left);
visit(node.right);
};
visit(this.root);
this._size -= removed;
this._tombstones += removed;
return removed;
}
// ── KNN search ─────────────────────────────────────────────────────────────
/**
* Find the k nearest neighbors to `query`.
* Find the k nearest live neighbors to `query`.
* Returns results sorted by distance ascending.
*/
knn(query: number[], k: number): KNNResult<T>[] {
@@ -170,7 +206,7 @@ export class KDTree<T = unknown> {
// ── Radius search ──────────────────────────────────────────────────────────
/**
* Return all points whose distance to `query` is ≤ `radius`,
* Return all live points whose distance to `query` is ≤ `radius`,
* sorted by distance ascending.
*/
radiusSearch(query: number[], radius: number): KNNResult<T>[] {
@@ -185,7 +221,7 @@ export class KDTree<T = unknown> {
// ── Conversion ─────────────────────────────────────────────────────────────
/** Collect all points in the tree (order not guaranteed). */
/** Collect all live points in the tree (order not guaranteed). */
toArray(): KDPoint<T>[] {
const out: KDPoint<T>[] = [];
this.collect(this.root, out);
@@ -193,12 +229,14 @@ export class KDTree<T = unknown> {
}
/**
* Rebuild the tree from its current points as a balanced tree.
* Useful after many individual insertions to restore O(log n) query time.
* Rebuild the tree from its current live points as a balanced tree.
* Physically purges tombstones and restores O(log n) query time.
*/
rebalance(): void {
const points = this.toArray();
this.root = points.length ? this.buildBalanced(points, 0) : null;
this._size = points.length;
this._tombstones = 0;
}
// ── Private: build ─────────────────────────────────────────────────────────
@@ -250,8 +288,10 @@ export class KDTree<T = unknown> {
): void {
if (node === null) return;
if (!node.deleted) {
const dist = this.distanceFn(query, node.point.vector);
heap.push({ point: node.point, distance: dist });
}
const axis = node.axis;
const diff = query[axis] - node.point.vector[axis];
@@ -284,10 +324,12 @@ export class KDTree<T = unknown> {
): void {
if (node === null) return;
if (!node.deleted) {
const dist = this.distanceFn(query, node.point.vector);
if (dist <= radius) {
results.push({ point: node.point, distance: dist });
}
}
const axis = node.axis;
const diff = query[axis] - node.point.vector[axis];
@@ -309,7 +351,7 @@ export class KDTree<T = unknown> {
private collect(node: KDNode<T> | null, out: KDPoint<T>[]): void {
if (node === null) return;
out.push(node.point);
if (!node.deleted) out.push(node.point);
this.collect(node.left, out);
this.collect(node.right, out);
}

View File

@@ -1,15 +1,20 @@
import {snakeCase} from '@ztimson/utils';
import {clean, makeUnique, snakeCase} from '@ztimson/utils';
import {AbortablePromise, Ai} from './ai.ts';
import {Anthropic} from './antrhopic.ts';
import {OpenAi} from './open-ai.ts';
import {LLMProvider} from './provider.ts';
import {AiTool, AiToolArg} from './tools.ts';
import {fileURLToPath} from 'url';
import {dirname, join} from 'path';
import {spawn} from 'node:child_process';
import {Memory, MemoryCache, MemoryManager, MemoryOptions} from './memory.ts';
import {Memory, MemoryCache, MemoryManager, MemoryOptions, stripHeader} from './memory.ts';
import {mkdtempSync} from 'node:fs';
import fs from 'node:fs/promises';
import {tmpdir} from 'node:os';
import {dirname, join, basename, extname} from 'path';
import { PDFParse } from 'pdf-parse';
const MAX_AGENT_DEPTH = 5;
const PDF_OCR_PAGE_THRESHOLD = 12; // above this many pages, OCR scanned pages instead of feeding images to the model
export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
@@ -27,13 +32,32 @@ export type Agent = {
agents?: string[] | null;
}
export type LLMFile = {
/** Path to file on disk */
path?: string;
/** File content: raw text, base64-encoded binary, or a Buffer */
content?: string | Buffer;
/** Original filename, used to infer type from extension */
name?: string;
/** Mime type override, inferred from extension if omitted */
mime?: string;
/** @internal set once extraction has run, skips re-processing next turn */
extracted?: boolean;
};
export type LLMMessage = {
/** Message originator */
role: 'assistant' | 'system' | 'user';
/** Message content */
content: string | any;
/** Files attached to request */
files?: LLMFile[];
/** Timestamp */
timestamp?: number;
/** Response duration in ms */
duration?: number;
/** Tokens per second */
tps?: number;
} | {
/** Tool call */
role: 'tool';
@@ -63,7 +87,7 @@ export type LLMRequest = {
/** Message history */
history?: LLMMessage[];
/** Max tokens for request */
max_tokens?: number;
maxTokens?: number;
/** 0 = Rigid Logic, 1 = Balanced, 2 = Hyper Creative **/
temperature?: number;
/** Available tools */
@@ -84,6 +108,8 @@ export type LLMRequest = {
mcp?: McpServer[];
/** Subagents exposed as delegatable/wrapped tools */
agents?: Agent[];
/** Attach files to request */
files?: LLMFile[];
/** @internal recursion guard for nested agent delegation */
_agentDepth?: number;
}
@@ -107,6 +133,11 @@ export type Skill = {
}
class LLM {
private static AUDIO_EXT = ['wav','mp3','m4a','flac','ogg','aac','wma'];
private static IMAGE_EXT = ['png','jpg','jpeg','bmp','gif','tiff','webp'];
private static TEXT_EXT = ['txt','md','csv','json','xml','html','js','ts','py','yaml','yml','log'];
private static PDF_EXT = ['pdf'];
private memoryManager!: MemoryManager;
defaultModel!: string;
@@ -122,27 +153,141 @@ class LLM {
this.memoryManager = new MemoryManager(this);
}
private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
private async loadBuffer(file: LLMFile, asText: boolean): Promise<Buffer> {
if(file.path) return fs.readFile(file.path);
if(Buffer.isBuffer(file.content)) return file.content;
if(typeof file.content === 'string') return Buffer.from(file.content, asText ? 'utf-8' : 'base64');
throw new Error('No path or content provided');
}
private async writeTemp(name: string, buffer: Buffer): Promise<string> {
const path = join(mkdtempSync(join(tmpdir(), 'ai-file-')), name);
await fs.writeFile(path, buffer);
return path;
}
/**
* Extract text from a PDF. Pages with no text layer (scanned/image-only) are handled as either:
* - Rendered to images and returned alongside the text so the (vision-capable) model can read them directly
* - OCR'd via Tesseract when the doc is too large to reasonably pass as images
*/
private async resolvePdf(buffer: Buffer): Promise<{text: string, images: {mime: string, data: string}[]}> {
const parser = new PDFParse({data: buffer});
try {
const {text, pages} = await parser.getText();
const scanned = (pages || []).filter(p => !p.text?.trim());
if(!scanned.length) return {text: text.trim() || '[Empty PDF]', images: []};
const total = pages.length;
const pageNums = scanned.map(p => p.num);
const {pages: shots} = await parser.getScreenshot({partial: pageNums});
if(total <= PDF_OCR_PAGE_THRESHOLD) {
return {
text: text.trim(),
images: shots.map(s => ({mime: 'image/png', data: Buffer.from(s.data).toString('base64')}))
};
}
const ocrText = await Promise.all(shots.map(async (s, i) => {
const path = await this.writeTemp(`page-${pageNums[i]}.png`, Buffer.from(s.data));
try {
return await this.ai.vision.ocr(path) || '';
} finally {
fs.rm(dirname(path), {recursive: true, force: true}).catch(() => {});
}
}));
return {text: [text.trim(), ...ocrText].filter(Boolean).join('\n\n'), images: []};
} finally {
await parser.destroy();
}
}
private async resolveFile(file: LLMFile): Promise<{text?: string, images?: {mime: string, data: string}[]}> {
const name = file.name || (file.path ? basename(file.path) : 'file');
// Already resolved on a previous turn, reuse cached text
if(file.extracted) return {text: `<file name="${name}">\n${file.content}\n</file>`};
const ext = extname(name).slice(1).toLowerCase();
const mime = file.mime || '';
const isAudio = mime.startsWith('audio/') || LLM.AUDIO_EXT.includes(ext);
const isImage = mime.startsWith('image/') || LLM.IMAGE_EXT.includes(ext);
const isPdf = mime === 'application/pdf' || LLM.PDF_EXT.includes(ext);
const isText = mime.startsWith('text/') || LLM.TEXT_EXT.includes(ext);
let tmpDir: string | null = null;
try {
if(isImage) {
const data = (await this.loadBuffer(file, false)).toString('base64');
return {images: [{mime: mime || `image/${ext === 'jpg' ? 'jpeg' : ext}`, data}]};
}
if(isPdf) {
const {text, images} = await this.resolvePdf(await this.loadBuffer(file, false));
// Only cache/skip re-processing when we didn't need to hand off images (OCR'd or fully text-based)
if(!images.length) {
file.content = text;
file.extracted = true;
delete file.path;
}
return {text: `<file name="${name}">\n${text || '[Scanned PDF - see attached page images]'}\n</file>`, images};
}
let text: string;
if(isAudio) {
let path = file.path;
if(!path) {
const buffer = await this.loadBuffer(file, false);
path = await this.writeTemp(name, buffer);
tmpDir = dirname(path);
}
text = await this.ai.audio.asr(path) || '';
} else if(isText) {
text = (await this.loadBuffer(file, true)).toString('utf-8');
} else {
text = typeof file.content === 'string' ? file.content : `[Binary file, unable to extract: ${name}]`;
}
file.content = text;
file.extracted = true;
delete file.path;
return {text: `<file name="${name}">\n${text}\n</file>`};
} catch(err: any) {
return {text: `<file name="${name}">Failed to process: ${err.message}</file>`};
} finally {
if(tmpDir) fs.rm(tmpDir, {recursive: true, force: true}).catch(() => {});
}
}
private async resolveFiles(files: LLMFile[]): Promise<{text: string, images: {mime: string, data: string}[]}> {
const resolved = await Promise.all(files.map(f => this.resolveFile(f)));
return {
text: resolved.filter(r => r.text).map(r => r.text).join('\n\n'),
images: resolved.flatMap(r => r.images || [])
};
}
private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: ((keep?: boolean) => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
return agents.map(a => {
const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
return {
name: toolName,
description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
args: <any>(a.delegate ? {} : {
context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
args: clean<any>({
context: !a.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
}),
fn: async (args: any, stream: any, ai: any, id?: string) => {
if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
// Opt-in only, self always excluded regardless of whitelist
const nested = (a.agents || [])
.map(name => allAgents.find(x => x.name === name))
.filter((x): x is Agent => !!x && x.name !== a.name);
const request = this.ask(a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`, {
system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
const request = this.ask(q, {
system: `You are a specialized subagent being called from an orchestrator
${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation' : 'You are wrapped in a tool call that will be analysis by an LLM'}
Dispense with greetings and focus on your instructions using available tools and returning only the final result unless specifically instructed to converse
${a.system}`,
model: a.model || undefined,
@@ -201,7 +346,7 @@ ${a.system}`,
const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
return {
prompt: `You have access to the following MCP tools:\n${list}`,
prompt: `## MCP\nYou have access to the following MCP tools:\n${list}`,
tools: allTools
};
}
@@ -210,7 +355,7 @@ ${a.system}`,
if(!skills?.length) return {prompt: '', tools: []};
const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
return {
prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
prompt: `## Skills\nYou have access to the following skill documents, whenever there is overlap between a question and a skill file, use \`skill_read\` to get instructions and background knowledge:\n${list}`,
tools: [{
name: 'skill_read',
description: 'Read the full content of a skill/knowledge document',
@@ -252,11 +397,13 @@ ${a.system}`,
if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
let request: AbortablePromise<string> | null = null;
let aborted = false;
const nestedAborts: (() => void)[] = [];
const abort = () => {
let keepOnAbort = true;
const nestedAborts: ((keep?: boolean) => void)[] = [];
const abort = (keep = true) => {
aborted = true;
request?.abort?.();
nestedAborts.forEach(a => a());
keepOnAbort = keep;
request?.abort?.(keep);
nestedAborts.forEach(a => a(keep));
};
let promise: any;
@@ -266,6 +413,24 @@ ${a.system}`,
let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
const prompts: string[] = [];
let history = options.history || [];
const historyStart = history.length;
const files = options.files || [];
if(message || files.length) history.push({role: 'user', content: message || '', timestamp: Date.now()});
// Accumulate streamed text so it can be committed to history if aborted mid-generation
let partialText = '';
const onStream = options.stream;
const stream = (chunk: {text?: string, tool?: string, done?: true}) => {
if(chunk.text) partialText += chunk.text;
return onStream?.(chunk);
};
/** Commit (keep) or discard this turn's progress on abort, then throw */
const abortNow = (): never => {
if(keepOnAbort) { if(partialText) history.push({role: 'assistant', content: partialText, timestamp: Date.now()}); }
else history.splice(historyStart, history.length - historyStart);
throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
};
// MCP
const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -294,8 +459,8 @@ ${a.system}`,
const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
if(mems.length) {
if(mem.inject) {
const pool = 15; // candidates considered, cheap since only refs are listed
const budget = mem.maxTokens ?? 2000; // actual content injected
const pool = 15;
const budget = mem.maxTokens ?? 2000;
const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
let used = 0;
@@ -309,33 +474,64 @@ ${a.system}`,
} else listed.push(r);
}
prompts.unshift(`You have access to the following memory files:
${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
${preloaded.length ? `
Relevant memories have been preloaded:
${preloaded.map(r => `
**${r.name}**
${r.description}
${r.content}
`).join('\n---\n')}
` : ''}${listed.length ? `
Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.name).join(', ')}
` : ''}`.trim());
prompts.unshift(`## Memory
You have a background memory process which has prefetched relevant information${mem.update ? ' and will create new memories from this conversation' : ''} for you
Assume it is perfect and never mention this process to anyone ever
Always use your memories to craft a personalized response, they contain links / [[wiki links]] which you use navigate between them
${mem.tool ? `You can access memory files via the \`memory_search\` and \`memory_recall\` tools
When you need information about the user, \`memory_recall\` \`People/User\` before asking (fetch if not included bellow)
When you need information not provided, attempt 1-3 \`memory_search\` calls with distinct queries before asking` : ''}
${preloaded.length ? `### Prefetched Memories (Most relevant first):
${preloaded.map(r => `Memory: ${r.name}
Description: ${r.description}
Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
\`\`\`
${stripHeader(r.content)}
\`\`\``).join('\n\n')}` : ''}
${mem.tool && listed.length ? '\n' + listed.map(r => `Memory: ${r.name}
Description: ${r.description}
Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
<!-- Truncated -->`).join('\n\n') : ''}`.trim())
}
if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
}
}
if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
if(aborted) abortNow();
const lastMsg = history[history.length - 1];
if(files.length && lastMsg?.role === 'user') lastMsg.files = files;
const restores: {msg: LLMMessage, content: any}[] = [];
for(const msg of history) {
if(msg.role !== 'user' || !msg.files?.length) continue;
const {text, images} = await this.resolveFiles(msg.files);
if(!text && !images.length) continue;
restores.push({msg, content: msg.content});
const merged = text ? [msg.content, text].filter(Boolean).join('\n\n') : msg.content;
msg.content = images.length
? [...images.map(i => ({type: 'image', mime: i.mime, data: i.data})), {type: 'text', text: merged}]
: merged;
}
const toolTimings = new Map<string, {duration: number, tps: number}>();
tools = this.wrapToolTiming(tools, toolTimings);
if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
if(aborted) abortNow();
prompts.unshift(options.system || this.ai.options.llm?.system || '');
request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
let resp = await request;
request = this.models[m].ask('', {...options, tools, stream, system: prompts.filter(Boolean).join('\n\n')});
let resp: string;
try {
resp = await request;
} catch(err: any) {
if(aborted) return abortNow();
throw err;
}
// Strip the file injection shim
restores.forEach(({msg, content}) => msg.content = content);
// Capture meta (duration / tps)
for(const h of history) {
@@ -364,14 +560,6 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
return Object.assign(promise, {abort});
}
/**
* Digest full conversation history into memory documents.
* Call on session end to persist the conversation.
*/
async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
}
/**
* Compress chat history to reduce context size
* @param {LLMMessage[]} history Chatlog that will be compressed
@@ -551,6 +739,14 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
};
}
/**
* Digest full conversation history into memory documents.
* Call on session end to persist the conversation.
*/
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
}
/**
* Create a summary of some text
* @param {string} text Text to summarize

View File

@@ -1,9 +1,13 @@
import {MemoryNode, patchGraph, rebuildGraph} from './helpers.ts';
import {LLMRequest, LLMMessage} from './llm.ts';
import {AiTool} from './tools.ts';
import {KDPoint, KDTree} from './kd-tree.ts';
const FACTS_HEADING = '## Facts';
import {KDTree} from './kd-tree.ts';
import {escapeRegex} from '@ztimson/utils';
const MERGE_THRESHOLD = 0.12;
const PENDING_HEADING = '## Pending';
const TREE_TOMBSTONE_LIMIT = 0.25;
const ALIAS_MATCH_THRESHOLD = 0.55;
const GENERIC_TEMPLATE = `# {{Title}}
## Summary
@@ -12,81 +16,16 @@ const GENERIC_TEMPLATE = `# {{Title}}
## Related`;
export class MemoryCache {
private tree: KDTree<MemoryRef>;
public memories: Memory[];
get length() { return this.memories.length; }
constructor(memories: Memory[]) {
this.memories = memories;
this.tree = this.buildTree();
}
private buildTree(): KDTree<MemoryRef> {
const embedded = this.memories.filter(m => m.embedding?.length);
if (!embedded.length) return new KDTree<MemoryRef>(0);
const dims = embedded[0].embedding.length;
const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
vector: m.embedding,
payload: {name: m.name, description: m.description},
}));
return new KDTree<MemoryRef>(dims, 'cosine', points);
}
search(query: number[], limit: number): MemoryRef[] {
const results = this.tree.knn(query, limit);
return results.map(r => r.point.payload);
}
add(memory: Memory): void {
this.memories.push(memory);
this.rebuild();
}
update(memory: Memory): void {
const idx = this.memories.findIndex(m => m.name === memory.name);
if (idx !== -1) {
this.memories[idx] = memory;
} else {
this.memories.push(memory);
}
this.rebuild();
}
remove(name: string): void {
const idx = this.memories.findIndex(m => m.name === name);
if (idx !== -1) {
this.memories.splice(idx, 1);
this.rebuild();
}
}
rebuild(): void {
this.tree = this.buildTree();
}
}
export type MemoryOptions = {
/** Memory object */
memory: Memory[] | MemoryCache;
/** Inject N memories into the system prompt */
inject?: boolean;
/** expose recall tool to LLM */
tool?: boolean;
/** Update memory on compression */
update?: boolean;
/** Max context size of memories to inject to each call (removed immediately after use) */
maxTokens?: number;
}
export type Memory = {
name: string;
description: string;
content: string;
/** Description embedding — indexed in the KD tree, used for merge/ANN candidate lookup */
embedding: number[];
/** Title-only embedding, weighted heaviest during recall ranking */
titleEmbedding?: number[];
/** Chunked body embeddings, best-chunk match used during recall ranking */
bodyEmbeddings?: number[][];
links: string[];
backlinks: string[];
}
@@ -94,6 +33,8 @@ export type Memory = {
type MemoryRef = {
name: string;
description: string;
/** Cosine distance from the query, present when returned from a search */
distance?: number;
}
type FactBucket = {
@@ -101,21 +42,9 @@ type FactBucket = {
facts: string[];
}
function extractLinks(content: string): string[] {
if (!content) return [];
const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
return [...new Set([...matches].map(m => m[1].trim()))];
}
export function rebuildGraph(memories: Memory[]): void {
for (const m of memories) m.links = extractLinks(m.content).filter(l => l !== m.name);
for (const m of memories) m.backlinks = [];
for (const m of memories) {
for (const link of m.links) {
const target = memories.find(t => t.name === link);
if (target) target.backlinks.push(m.name);
}
}
type FactAgentResult = {
buckets: FactBucket[];
journal: string;
}
function dedupeFacts(facts: string[]): string[] {
@@ -138,39 +67,171 @@ function cosineDistance(a: number[], b: number[]): number {
return denom === 0 ? 1 : 1 - dot / denom;
}
function getWeekMonday(date: Date = new Date()): string {
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
const day = d.getUTCDay();
const diff = day === 0 ? -6 : 1 - day;
d.setUTCDate(d.getUTCDate() + diff);
return d.toISOString().slice(0, 10);
function cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
return memories
.filter(m => m.embedding?.length)
.map(m => ({name: m.name, description: m.description, distance: cosineDistance(query, m.embedding)}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
}
/** Re-embed a node's title / description / body fields. Description embedding stays the KD-tree index key. */
async function embedMemoryFields(node: Memory, llm: any): Promise<void> {
const body = stripHeader(node.content);
const [titleE] = await llm.embedding(node.name.split('/').pop() || node.name);
const [descE] = await llm.embedding(node.description || '');
const bodyChunks = body ? await llm.embedding(body) : [];
if (titleE) node.titleEmbedding = titleE.embedding;
if (descE) node.embedding = descE.embedding;
node.bodyEmbeddings = bodyChunks.map((c: any) => c.embedding).filter(Boolean);
}
export function stripHeader(content: string): string {
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
}
export class MemoryCache {
private tree!: KDTree<MemoryRef>;
/** Tracks which memories are currently indexed in the tree, keyed by name -> embedding reference */
private indexed = new Map<string, number[]>();
public memories: Memory[];
public nodes: MemoryNode[] = [];
get length() { return this.memories.length; }
constructor(memories: Memory[]) {
this.memories = memories;
this.tree = new KDTree<MemoryRef>(0);
this.rebuild();
}
/** Incrementally sync the KD tree against `this.memories` instead of rebuilding from scratch */
private syncTree(): void {
const current = new Set(this.memories.map(m => m.name));
for (const [name, emb] of [...this.indexed]) {
const mem = this.memories.find(m => m.name === name);
if (!mem || !current.has(name) || mem.embedding !== emb) {
this.tree.remove(p => p.name === name);
this.indexed.delete(name);
}
}
for (const mem of this.memories) {
if (!mem.embedding?.length || this.indexed.has(mem.name)) continue;
if (this.tree.dims === 0) this.tree = new KDTree<MemoryRef>(mem.embedding.length, 'cosine');
if (mem.embedding.length !== this.tree.dims) continue; // guard against embedding model/dim drift
this.tree.insert({vector: mem.embedding, payload: {name: mem.name, description: mem.description}});
this.indexed.set(mem.name, mem.embedding);
}
if (this.tree.tombstoneRatio > TREE_TOMBSTONE_LIMIT) this.tree.rebalance();
}
search(query: number[], limit: number): MemoryRef[] {
if (!this.tree || this.tree.dims === 0) return [];
return this.tree.knn(query, limit).map(r => ({...r.point.payload, distance: r.distance}));
}
add(memory: Memory): void {
this.memories.push(memory);
this.rebuild([memory]);
}
update(memory: Memory): void {
const existing = this.memories.find(m => m.name === memory.name);
if (existing) Object.assign(existing, memory);
else this.memories.push(memory);
this.rebuild([existing ?? memory]);
}
remove(name: string): void {
const idx = this.memories.findIndex(m => m.name === name);
if (idx !== -1) {
this.memories.splice(idx, 1);
this.rebuild();
}
}
rebuild(changed?: Memory[]): void {
this.nodes = (changed?.length && this.nodes.length)
? patchGraph(this.memories, this.nodes, changed)
: rebuildGraph(this.memories);
this.syncTree();
}
}
class MemoryAccessor {
readonly list: Memory[];
private readonly cache: MemoryCache | null;
constructor(memories: Memory[] | MemoryCache) {
this.cache = memories instanceof MemoryCache ? memories : null;
this.list = this.cache ? this.cache.memories : <Memory[]>memories;
}
find(name: string): Memory | undefined {
return this.list.find(m => m.name === name);
}
commit(changed?: Memory[]): MemoryNode[] {
if (this.cache) {
this.cache.rebuild(changed);
return this.cache.nodes;
}
return rebuildGraph(this.list);
}
ghosts(): string[] {
const nodes = this.cache ? this.cache.nodes : rebuildGraph(this.list);
return nodes.filter(n => n.missing).map(n => n.name);
}
/** Cache path uses the KD tree's knn(); raw-array path (no cache available) falls back to a linear cosine scan */
search(vector: number[], limit: number): MemoryRef[] {
return this.cache ? this.cache.search(vector, limit) : cosineSearch(vector, this.list, limit);
}
forget(name: string): boolean {
const idx = this.list.findIndex(m => m.name === name);
if (idx === -1) return false;
this.list.splice(idx, 1);
this.commit();
return true;
}
async backfillEmbeddings(llm: any): Promise<number> {
const missing = this.list.filter(m => !m.embedding?.length);
if (!missing.length) return 0;
await Promise.all(missing.map(node => embedMemoryFields(node, llm)));
this.commit();
return missing.length;
}
}
export type MemoryOptions = {
/** Memory object */
memory: Memory[] | MemoryCache;
/** Inject N memories into the system prompt */
inject?: boolean;
/** expose recall tool to LLM */
tool?: boolean;
/** Update memory on compression */
update?: boolean;
/** Max context size of memories to inject to each call (removed immediately after use) */
maxTokens?: number;
}
export class MemoryManager {
private recentlyTouched = new Map<string, number>();
private mergeLock: Promise<any> = Promise.resolve();
private queues = new Map<string, {
dirty: boolean,
request: {abort?: () => void} | null,
task: Promise<void>,
}>();
private recentlyTouched = new Map<string, number>();
tools = {
read: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'memory_recall',
description: 'Read the full content of a memory document',
args: {
name: {type: 'string', description: 'Exact memory name', required: true},
},
fn: (args: any) => {
const mems = this.unwrap(memories);
const mem = mems.find(m => m.name === args.name);
if (!mem) return 'Document not found';
this.touch(mem.name);
return mem.content;
},
}),
forget: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'memory_forget',
description: 'Permanently delete a memory document and clean up all references to it',
@@ -182,6 +243,38 @@ export class MemoryManager {
return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
},
}),
read: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'memory_recall',
description: 'Read the full content of a memory document',
args: {
name: {type: 'string', description: 'Exact memory name', required: true},
},
fn: (args: any) => {
const mem = this.access(memories).find(args.name);
if (!mem) return 'Document not found';
this.touch(mem.name);
return mem.content;
},
}),
search: (memories: Memory[] | MemoryCache): AiTool => ({
name: 'memory_search',
description: 'Use embeddings to find the MOST relevant memories, even if NOT relevant',
args: {
query: {type: 'string', description: 'What to look for in the memories', required: true},
limit: {type: 'number', description: 'Number of memories to return', default: 1},
},
fn: async ({query, limit}) => {
const mem = await this.recollect(query, memories, limit)
return mem.map(m => `Memory: ${m.name}
Description: ${m.description}
Links: ${[...m.links, ...m.backlinks].join(', ')}
\`\`\`
${m.content}
\`\`\``).join('\n\n');
},
}),
};
constructor(private llm: any) {}
@@ -192,41 +285,18 @@ export class MemoryManager {
return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
}
private unwrap(memories: Memory[] | MemoryCache): Memory[] {
return memories instanceof MemoryCache ? memories.memories : memories;
private access(memories: Memory[] | MemoryCache): MemoryAccessor {
return new MemoryAccessor(memories);
}
private sync(memories: Memory[] | MemoryCache): void {
if (memories instanceof MemoryCache) memories.rebuild();
}
private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
if (!match) return {fm: new Map(), body: content};
const fm = new Map<string, string>();
for (const line of match[1].split('\n')) {
const i = line.indexOf(':');
if (i === -1) continue;
fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
}
return {fm, body: match[2]};
}
private writeFrontmatter(fm: Map<string, string>, body: string): string {
const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
}
private stripHeader(content: string): string {
return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
}
private touchHeader(node: Memory, body: string): string {
const {fm} = this.parseFrontmatter(node.content);
fm.set('name', node.name);
fm.set('description', node.description || '');
fm.set('modified', new Date().toISOString());
return this.writeFrontmatter(fm, body);
private stage(node: Memory, block: string): void {
this.ensureDoc(node);
const body = stripHeader(node.content);
const idx = body.indexOf(PENDING_HEADING);
const newBody = idx === -1
? `${body.trimEnd()}\n\n${PENDING_HEADING}\n${block}\n`
: `${body.slice(0, idx + PENDING_HEADING.length)}\n${block}${body.slice(idx + PENDING_HEADING.length)}`;
node.content = this.touchHeader(node, newBody);
}
private ensureDoc(node: Memory): void {
@@ -235,148 +305,147 @@ export class MemoryManager {
node.content = this.touchHeader(node, `# ${title}\n`);
}
private appendFacts(node: Memory, facts: string[]): void {
this.ensureDoc(node);
const body = this.stripHeader(node.content);
const bullets = facts.map(f => `- ${f}`).join('\n');
const idx = body.indexOf(FACTS_HEADING);
const newBody = idx === -1
? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
node.content = this.touchHeader(node, newBody);
private sanitizeDescription(text: string): string {
return (text ?? '').replace(/\s+/g, ' ').trim().slice(0, 240);
}
decay() {
for(const [name, ttl] of this.recentlyTouched) {
if(ttl <= 1) this.recentlyTouched.delete(name);
else this.recentlyTouched.set(name, ttl - 1);
}
private relink(memories: Memory[], from: string, to: string): void {
const pattern = new RegExp(`\\[\\[${escapeRegex(from)}\\]\\]`, 'g');
for (const m of memories) if (pattern.test(m.content)) m.content = m.content.replace(pattern, `[[${to}]]`);
}
touch(name: string, ttl = 2) {
this.recentlyTouched.set(name, ttl);
private normalizeLeaf(name: string): string {
return name.trim().toLowerCase().replace(/\s+/g, ' ');
}
getTouched(): string[] {
return [...this.recentlyTouched.keys()];
/**
* Resolve a fact-agent proposed subject to an existing node when it's an alias/rename of one.
* Exact match is checked first (cheap, and covers the common case since node names are
* already normalized at creation time). Only falls through to fuzzy alias matching against
* same-root candidates when there's no existing hit — i.e. only on likely-new-doc creation.
*/
private resolveSubject(subject: string, store: MemoryAccessor): string {
const trimmed = subject.trim();
const exact = store.find(trimmed);
if (exact) return exact.name;
const normalized = this.normalizeLeaf(trimmed);
const caseInsensitive = store.list.find(m => this.normalizeLeaf(m.name) === normalized);
if (caseInsensitive) return caseInsensitive.name;
const root = trimmed.split('/')[0];
const leaf = trimmed.split('/').slice(1).join('/') || trimmed;
const candidates = store.list.filter(m => m.name.split('/')[0] === root && m.name !== trimmed);
if (!candidates.length) return trimmed;
// fuzzyMatch requires >=2 terms; pad with an empty string when there's only one candidate
const leaves = candidates.map(m => m.name.split('/').slice(1).join('/') || m.name);
const probe = leaves.length > 1 ? leaves : [...leaves, ''];
const {max, similarities} = this.llm.fuzzyMatch(leaf, ...probe);
if (max >= ALIAS_MATCH_THRESHOLD) return candidates[similarities.indexOf(max)].name;
return trimmed;
}
forget(name: string, memories: Memory[] | MemoryCache): boolean {
const mem = this.unwrap(memories);
const idx = mem.findIndex(m => m.name === name);
if (idx === -1) return false;
private async factAgent(conversation: string, store: MemoryAccessor, options: LLMRequest): Promise<FactAgentResult> {
const ghosts = store.ghosts();
mem.splice(idx, 1);
rebuildGraph(mem);
this.sync(memories);
return true;
const response = await this.llm.ask(conversation, {
model: options.model,
temperature: 0.2,
system: `You are a fact extractor for Obsidian-style knowledge vaults. Analyze the conversation and produce:
1. Journal recap (single paragraph)
- "Captains Log" style record keeping
- What was discussed/worked on, decisions, user's events/state/mood, general context
- Leave empty only for trivial/empty exchanges/small talk
2. Fact buckets
- ONLY facts the USER explicitly stated about themselves, their work, projects, or decisions made during this conversation
- NEVER extract greetings, pleasantries, or anything the assistant itself said
- Extract the final/end state, not deltas
Path assignment rules:
- Reuse existing node names whenever possible, including when the subject is an alias/nickname of an existing node (e.g. "Rob" referring to an existing "People/Robert")
- Documents should be grouped and named by the root subject
- Person → People/Name
- Project → Projects/Name
- Concept → Concepts/Name
- A bug report, its investigation, should be nested and attached to the same root subject node
- Tickets/one-off tasks → file under the project/name/component they belong to
- Only create a new top-level node when the fact belongs to a genuinely new subject (person/project/concept)\`
Available nodes:
${this.listNodes(store.list).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
schema: {
journal: {type: 'string', description: 'Short day-to-day recap; empty if nothing happened.', required: false},
buckets: {type: 'array', description: 'Groups of facts to remember; empty array if nothing worth storing.', items: {
type: 'object', items: {
subject: {type: 'string', description: 'Exact node name or new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
facts: {type: 'array', description: 'Facts to store here', items: {type: 'string'}},
},
},
},
},
});
const buckets = new Map<string, string[]>();
for (const bucket of response.buckets ?? []) {
const subject = bucket.subject.trim();
const facts = buckets.get(subject) ?? [];
facts.push(...dedupeFacts(bucket.facts));
buckets.set(subject, facts);
}
async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
const mem = this.unwrap(memories);
if (!mem.length) return [];
const [e] = await this.llm.embedding(query);
if (!e) return [];
let vectorResults: MemoryRef[];
if (memories instanceof MemoryCache) vectorResults = memories.search(e.embedding, limit);
else vectorResults = this.cosineSearch(e.embedding, mem, limit);
const found = new Set<string>(vectorResults.map(r => r.name));
if (graphDepth > 0) {
const frontier = [...found];
for (let depth = 0; depth < graphDepth; depth++) {
const next: string[] = [];
for (const name of frontier) {
const node = mem.find(m => m.name === name);
if (!node) continue;
for (const link of node.links) {
if (!found.has(link) && mem.find(m => m.name === link)) {
found.add(link);
next.push(link);
}
}
}
frontier.splice(0, frontier.length, ...next);
if (!frontier.length) break;
}
return {
buckets: buckets.entries().toArray().map(([subject, facts]) => ({subject, facts})),
journal: (response.journal ?? '').trim(),
};
}
const vectorOrder = vectorResults.map(r => r.name);
const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
const ordered = [...vectorOrder, ...graphExpansions];
return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
}
private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
const scored = memories
.filter(m => m.embedding?.length)
.map(m => ({
ref: {name: m.name, description: m.description},
distance: cosineDistance(query, m.embedding),
}))
.sort((a, b) => a.distance - b.distance)
.slice(0, limit);
return scored.map(s => s.ref);
private getWeekMonday(date: Date = new Date()): string {
const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
const day = d.getUTCDay();
const diff = day === 0 ? -6 : 1 - day;
d.setUTCDate(d.getUTCDate() + diff);
return d.toISOString().slice(0, 10);
}
private listNodes(memories: Memory[]): MemoryRef[] {
return memories.map(m => ({name: m.name, description: m.description}));
}
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
const conversation = history
.filter(h => h.role === 'user' || h.role === 'assistant')
.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
if (!conversation) return [];
/** Finds the closest merge candidate via the KD tree's knn() instead of a manual O(n) cosine scan */
private async checkMerge(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest, threshold = MERGE_THRESHOLD): Promise<Memory | null> {
if (!node.embedding?.length || node.name.startsWith('Journal/')) return null;
const store = this.access(memories);
const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
// NOTE: adjust field names below (id/tool_call_id/name) to match your LLMMessage/tool-call schema.
const pending = {role: 'tool', name: 'memory_process', id: uid, content: 'Processing…'} as unknown as LLMMessage;
history.push(pending);
const candidate = store.search(node.embedding, 5)
.find(r => r.name !== node.name && !r.name.startsWith('Journal/') && r.distance !== undefined && r.distance <= threshold);
if (!candidate) return null;
const closest = store.find(candidate.name);
if (!closest) return null;
const mem = this.unwrap(memories);
const buckets = await this.factAgent(conversation, mem, options, getWeekMonday());
const touched: Memory[] = [];
const result = await this.mergeAgent(node, closest, options);
const merged: Memory = {name: result.name, description: this.sanitizeDescription(result.description), content: '', embedding: [], links: [], backlinks: []};
merged.content = this.touchHeader(merged, result.content);
await embedMemoryFields(merged, this.llm);
for (const {subject, facts} of buckets) {
let node = mem.find(m => m.name === subject);
if (!node) {
node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
mem.push(node);
}
this.appendFacts(node, facts);
const [e] = await this.llm.embedding(node.content);
if (e) node.embedding = e.embedding;
this.touch(node.name);
touched.push(node);
this.relink(store.list, node.name, merged.name);
this.relink(store.list, closest.name, merged.name);
this.queues.get(closest.name)?.request?.abort?.();
this.queues.delete(closest.name);
store.forget(node.name);
store.forget(closest.name);
store.list.push(merged);
store.commit();
return merged;
}
if (touched.length) {
rebuildGraph(mem);
this.sync(memories);
(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
for (const node of touched) this.reconcile(node, memories, options).catch(() => {});
} else {
(pending as any).content = 'Nothing worth remembering.';
}
return touched;
}
/** Manual/cron entry point. scope 'touched' only reconciles docs with a pending Facts inbox. */
async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
const mem = this.unwrap(memories);
const targets = scope === 'all' ? mem : mem.filter(m => m.content.includes(FACTS_HEADING));
await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
this.sync(memories);
}
/**
* Coalescing queue: if a doc is already reconciling, mark it dirty and abort the in-flight
* request. The loop below always re-reads node.content fresh, so nothing is ever dropped.
*/
private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
const key = node.name;
const existing = this.queues.get(key);
@@ -388,22 +457,28 @@ export class MemoryManager {
const entry = {dirty: false, request: null, task: Promise.resolve()};
this.queues.set(key, entry);
const mem = this.unwrap(memories);
const store = this.access(memories);
entry.task = (async () => {
let current = node, merged = false;
try {
do {
entry.dirty = false;
await this.reconcileDoc(node, mem, options, entry);
await this.docAgent(current, store.list, options, entry);
this.mergeLock = this.mergeLock.then(() => this.checkMerge(current, memories, options));
const result = await this.mergeLock;
if (result) { current = result; merged = true; }
} while (entry.dirty);
})().finally(() => {
} finally {
store.commit(merged ? undefined : [node]);
this.queues.delete(key);
rebuildGraph(mem);
this.sync(memories);
});
}
})();
return entry.task;
}
private async reconcileDoc(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
const currentBody = this.stripHeader(node.content);
private async docAgent(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
if(!memories.includes(node)) return;
const currentBody = stripHeader(node.content);
let update;
try {
for (let i = 0; i < 2 && !update?.content; i++) {
@@ -411,27 +486,27 @@ export class MemoryManager {
model: options.model,
temperature: 0.3,
schema: {
description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
description: {type: 'string', description: 'One factual sentence describing the document\'s ENTIRE SUBJECT MATTER — for use as a search/merge fingerprint', required: true},
content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
},
system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
system: `You are a knowledge base editor maintaining one Obsidian-style document.
If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
If it has a "## Pending" section, fold all new material into the appropriate part, resolve overlap, then remove the section entirely. If no section, just tidy per the rules below.
Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
Use this loose structure, adapting headings to what the content needs:
\`\`\`markdown
${GENERIC_TEMPLATE}
\`\`\`
Formatting rules:
- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, tables for 2D data
- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
- Create links for specific entities (person, place, project, program) and abstract concepts, but skip generics (car, red, dog)
- Keep the document concise, factual, and human-readable
- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
- Do not add frontmatter blocks, filler, preamble, or AI commentary
Rules:
- Contradictions: "## Pending" holds the newest information — bias toward it. Fold it in as the standing fact and drop the outdated statement, unless the old context adds meaningful nuance (e.g. "previously X, now Y"). This document should read as a source of truth, not an audit log
- Journals (Journal/...): keep entries as a chronological timeline; clean up grammar within entries but never delete history
- Use Obsidian markdown: # headings, **bold**, bullet/numbered lists, tables for 2D data
- Link specific entities and concepts with [[WikiLink]] (e.g., [[Projects/KiwixServer]]); skip generics
- Keep concise, factual, human-readable
- NO frontmatter, filler, preamble, or AI commentary
Other nodes in the vault (link to these instead of duplicating their content):
Available nodes to link to (don't duplicate their content):
${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
Current document:
@@ -450,53 +525,200 @@ ${currentBody}
}
if (!update?.content) return;
node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
node.description = node.name !== 'People/User' ? this.sanitizeDescription(update.description) : 'All information about the current user';
node.content = this.touchHeader(node, update.content);
const [e] = await this.llm.embedding(node.content);
if (e) node.embedding = e.embedding;
await embedMemoryFields(node, this.llm);
}
private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
const buckets = new Map<string, string[]>();
await this.llm.ask(conversation, {
private async mergeAgent(a: Memory, b: Memory, options: LLMRequest): Promise<{name: string, description: string, content: string}> {
const modifiedOf = (m: Memory) => this.parseFrontmatter(m.content).fm.get('modified') || 'unknown';
return this.llm.ask('', {
model: options.model,
temperature: 0.2,
system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
Rules:
- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
- ONLY extract decisions that were MADE during this conversation
- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
- DO NOT extract greetings, pleasantries, or generic exchanges
- DO NOT extract deltas or changes in facts; ONLY the end fact
- If nothing worth remembering was said, do not call any tools
When extracting facts, you MUST also decide the exact destination path:
- Use an existing node name if the facts clearly belong there
- All information primarily about the user should go under "People/User"
- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
- For journal entries, use "Journal"
Available nodes:
- Journal
${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
tools: [{
name: 'facts_extract',
description: 'Submit facts with their destination',
args: {
destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
facts: {type: 'string', description: 'Comma-separated facts', required: true},
temperature: 0.3,
schema: {
name: {type: 'string', description: 'New path for the merged doc, collection/subject format (e.g. Projects/Oxide) — only reuse an old title if it\'s genuinely the best fit', required: true},
description: {type: 'string', description: 'One factual sentence describing the merged document\'s subject matter', required: true},
content: {type: 'string', description: 'Fully reconciled body in markdown, without frontmatter', required: true},
},
fn: (args: any) => {
const subject = args.destination.trim().toLowerCase() === 'journal'
? `Journal/${weekKey}` : args.destination.trim();
const facts = buckets.get(subject) ?? [];
facts.push(...dedupeFacts(String(args.facts).split(',')));
buckets.set(subject, facts);
return 'Recorded';
},
}],
system: `You are a knowledge base editor merging two overlapping Obsidian documents into one.
Structure loosely:
\`\`\`markdown
${GENERIC_TEMPLATE}
\`\`\`
Combine both documents, resolve duplication. On contradictions, bias toward whichever document was modified more recently; drop the outdated statement unless the old context adds meaningful nuance.
Document A ("${a.name}", last modified ${modifiedOf(a)}):
\`\`\`markdown
${stripHeader(a.content)}
\`\`\`
Document B ("${b.name}", last modified ${modifiedOf(b)}):
\`\`\`markdown
${stripHeader(b.content)}
\`\`\``,
});
return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
}
private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
if (!match) return {fm: new Map(), body: content};
const fm = new Map<string, string>();
for (const line of match[1].split('\n')) {
const i = line.indexOf(':');
if (i === -1) continue;
const key = line.slice(0, i).trim();
const raw = line.slice(i + 1).trim();
let value = raw;
try { value = JSON.parse(raw); } catch { /* legacy unquoted value, keep raw */ }
fm.set(key, value);
}
return {fm, body: match[2]};
}
/**
* Writes the code-owned frontmatter block. `body` is passed through stripHeader() first so a
* model that ignores instructions and hallucinates its own `---` block can never corrupt or
* duplicate the real frontmatter — the LLM only ever gets to influence the body.
*/
private touchHeader(node: Memory, body: string): string {
const {fm} = this.parseFrontmatter(node.content);
fm.set('name', node.name);
fm.set('description', node.description || '');
fm.set('modified', new Date().toISOString());
return this.writeFrontmatter(fm, stripHeader(body));
}
private writeFrontmatter(fm: Map<string, string>, body: string): string {
const lines = [...fm.entries()].map(([k, v]) => `${k}: ${JSON.stringify(String(v).replace(/\s+/g, ' ').trim())}`);
return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
}
decay() {
for (const [name, ttl] of this.recentlyTouched) {
if (ttl <= 1) this.recentlyTouched.delete(name);
else this.recentlyTouched.set(name, ttl - 1);
}
}
touch(name: string, ttl = 2) {
this.recentlyTouched.set(name, ttl);
}
forget(name: string, memories: Memory[] | MemoryCache): boolean {
return this.access(memories).forget(name);
}
/** Ranks a candidate pool by weighted title/description/body similarity against the query embedding */
private rankByFields(query: number[], candidates: Memory[], limit: number): Memory[] {
const scored = candidates.map(m => {
const titleSim = m.titleEmbedding?.length ? 1 - cosineDistance(query, m.titleEmbedding) : 0;
const descSim = m.embedding?.length ? 1 - cosineDistance(query, m.embedding) : 0;
const bodySim = m.bodyEmbeddings?.length
? Math.max(...m.bodyEmbeddings.map(b => 1 - cosineDistance(query, b)))
: 0;
return {memory: m, score: titleSim * 0.5 + descSim * 0.35 + bodySim * 0.15};
});
return scored.sort((a, b) => b.score - a.score).slice(0, limit).map(s => s.memory);
}
async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
const store = this.access(memories);
if (!store.list.length) return [];
await store.backfillEmbeddings(this.llm);
const [e] = await this.llm.embedding(query);
if (!e) return [];
// Description embedding is the cheap ANN index key; pull a wider pool then re-rank by field weight
const pool = store.search(e.embedding, Math.max(limit * 3, limit));
const poolMemories = pool.map(r => store.find(r.name)).filter((m): m is Memory => !!m);
const ranked = this.rankByFields(e.embedding, poolMemories, limit);
const found = new Set<string>(ranked.map(m => m.name));
if (graphDepth > 0) {
let frontier = [...found];
for (let depth = 0; depth < graphDepth && frontier.length; depth++) {
const next: string[] = [];
for (const name of frontier) {
const node = store.find(name);
if (!node) continue;
for (const link of node.links) {
if (!found.has(link) && store.find(link)) {
found.add(link);
next.push(link);
}
}
}
frontier = next;
}
}
const rankedOrder = ranked.map(m => m.name);
const graphExpansions = [...found].filter(n => !rankedOrder.includes(n));
return [...rankedOrder, ...graphExpansions].map(n => store.find(n)!).filter(Boolean);
}
async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
const conversation = history
.filter(h => h.role === 'user' || h.role === 'assistant')
.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
if (!conversation) return [];
const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
const pending = {role: 'tool', name: 'memory_process', id: uid, content: conversation} as unknown as LLMMessage;
history.push(pending);
const store = this.access(memories);
const {buckets, journal} = await this.factAgent(conversation, store, options);
const touched: Memory[] = [];
if (journal) {
const journalName = `Journal/${this.getWeekMonday()}`;
let jnode = store.find(journalName);
if (!jnode) {
jnode = {name: journalName, description: '', content: '', embedding: [], links: [], backlinks: []};
store.list.push(jnode);
}
this.stage(jnode, `### ${new Date().toISOString().slice(0, 10)}\n${journal}`);
touched.push(jnode);
}
for (const {subject, facts} of buckets) {
const resolved = this.resolveSubject(subject, store);
let node = store.find(resolved);
if (!node) {
node = {name: resolved, description: '', content: '', embedding: [], links: [], backlinks: []};
store.list.push(node);
}
this.stage(node, facts.map(f => `- ${f}`).join('\n'));
touched.push(node);
}
await Promise.all(touched.map(async node => {
await embedMemoryFields(node, this.llm);
this.touch(node.name);
}));
if (touched.length) {
store.commit(touched);
(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
Promise.all(touched.map(node => this.reconcile(node, memories, options).catch(() => {})));
} else {
(pending as any).content = 'Nothing worth remembering.';
}
(touched as any).uid = uid;
return touched;
}
async reconcileAll(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
const store = this.access(memories);
const targets = scope === 'all' ? store.list : store.list.filter(m => m.content.includes(PENDING_HEADING));
await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
store.commit();
}
}

View File

@@ -25,75 +25,47 @@ export class OpenAi extends LLMProvider {
return client;
}
private toStandard(history: any[]): LLMMessage[] {
for(let i = 0; i < history.length; i++) {
const h = history[i];
if(h.role === 'assistant' && h.tool_calls) {
const items: any[] = [];
if(h.content) items.push({role: 'assistant', content: h.content, timestamp: h.timestamp, duration: h.duration, tps: h.tps});
items.push(...h.tool_calls.map((tc: any) => ({
role: 'tool',
id: tc.id,
name: tc.function.name,
args: JSONAttemptParse(tc.function.arguments, {}),
timestamp: h.timestamp,
duration: h.duration,
tps: h.tps
})));
history.splice(i, 1, ...items);
i += items.length - 1;
} else if(h.role === 'tool') {
const record = history.find(h2 => h.tool_call_id == h2.id);
if(record) {
if(h.content?.includes('"error":')) record.error = h.content;
else record.content = h.content || '';
}
history.splice(i, 1);
i--;
}
if(!history[i]?.timestamp) history[i].timestamp = Date.now();
}
return history;
private toWireContent(content: any): any {
if(!Array.isArray(content)) return content;
return content.map(c => c.type === 'image'
? {type: 'image_url', image_url: {url: `data:${c.mime};base64,${c.data}`}}
: {type: 'text', text: c.text});
}
private fromStandard(history: LLMMessage[]): any[] {
return history.reduce((result, h) => {
/** Convert standard history -> OpenAI wire format */
private toWire(history: LLMMessage[], system?: string): any[] {
const wire: any[] = [];
if(system) wire.push({role: 'system', content: system});
for(const h of history) {
if(h.role === 'tool') {
result.push({
wire.push({
role: 'assistant',
content: null,
tool_calls: [{id: h.id, type: 'function', function: {name: h.name, arguments: JSON.stringify(h.args)}}],
refusal: null,
annotations: [],
timestamp: h.timestamp,
}, {
role: 'tool',
tool_call_id: h.id,
content: h.error || h.content,
timestamp: h.timestamp,
content: h.error || h.content || '',
});
} else {
result.push(h);
wire.push({role: h.role, content: this.toWireContent(h.content)});
}
return result;
}, [] as any[]);
}
return wire;
}
ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
const controller = new AbortController();
return Object.assign(new Promise<any>(async (res, rej) => {
const base = (options.history || []).filter(h => h.role !== 'system');
let history = this.fromStandard([
...(options.system ? [{role: <any>'system', content: options.system, timestamp: Date.now()}] : []),
...base,
{role: 'user', content: message, timestamp: Date.now()}
]);
if(!options.history) options.history = [];
const history = options.history;
if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
const tools = options.tools || this.ai.options.llm?.tools || [];
const requestParams: any = {
model: options.model || this.model,
messages: history,
stream: !!options.stream,
max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
max_completion_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || undefined,
temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
tools: tools.map(t => ({
type: 'function',
@@ -111,56 +83,42 @@ export class OpenAi extends LLMProvider {
if(options.schema) {
const schema = convertSchema(options.schema);
requestParams.response_format = {
type: 'json_schema',
json_schema: {
name: 'response',
strict: true,
schema
requestParams.response_format = {type: 'json_schema', json_schema: {name: 'response', strict: true, schema}};
}
};
}
if(options.stream) requestParams.stream_options = {include_usage: true};
let resp: any, terminal = false, duration = 0, tps = 0;
try {
let terminal = false;
do {
requestParams.messages = history.map(({timestamp, ...m}) => m);
requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
const callStart = Date.now();
resp = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
const resp: any = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
err.message += `\n\nMessages:\n${JSON.stringify(requestParams.messages, null, 2)}`;
throw err;
});
let usage: any;
let usage: any, msg: any = {content: '', tool_calls: []};
if(options.stream) {
resp.choices = [{message: {role: 'assistant', content: '', tool_calls: [], timestamp: Date.now()}}];
for await (const chunk of resp) {
if(controller.signal.aborted) break;
if(chunk.usage) usage = chunk.usage;
if(chunk.choices[0]?.delta?.content) {
resp.choices[0].message.content += chunk.choices[0].delta.content;
msg.content += chunk.choices[0].delta.content;
options.stream({text: chunk.choices[0].delta.content});
}
if(chunk.choices[0]?.delta?.tool_calls) {
for(const deltaTC of chunk.choices[0].delta.tool_calls) {
const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
const existing = msg.tool_calls.find((tc: any) => tc.index === deltaTC.index);
if(existing) {
if(deltaTC.id) existing.id = deltaTC.id;
if(deltaTC.type) existing.type = deltaTC.type;
if(deltaTC.function) {
if(!existing.function) existing.function = {};
if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
}
if(deltaTC.function?.name) existing.function.name = deltaTC.function.name;
if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
} else {
resp.choices[0].message.tool_calls.push({
msg.tool_calls.push({
index: deltaTC.index,
id: deltaTC.id || '',
type: deltaTC.type || 'function',
function: {
name: deltaTC.function?.name || '',
arguments: deltaTC.function?.arguments || ''
}
function: {name: deltaTC.function?.name || '', arguments: deltaTC.function?.arguments || ''}
});
}
}
@@ -168,51 +126,51 @@ export class OpenAi extends LLMProvider {
}
} else {
usage = resp.usage;
msg = resp.choices[0].message;
}
duration = Date.now() - callStart;
tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
const duration = Date.now() - callStart;
const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
if(resp.error) throw new Error(resp.error);
const toolCalls = resp.choices[0].message.tool_calls || [];
const toolCalls = msg.tool_calls || [];
if(toolCalls.length && !controller.signal.aborted) {
history.push({...resp.choices[0].message, duration, tps});
const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
const tool = tools?.find(findByProp('name', toolCall.function.name));
if(options.stream) options.stream({tool: toolCall.function.name});
if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}', timestamp: Date.now()};
if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
const entries = toolCalls.map((tc: any) => {
const entry: any = {role: 'tool', id: tc.id, name: tc.function.name, args: JSONAttemptParse(tc.function.arguments, {}), content: undefined, timestamp: Date.now()};
history.push(entry);
return {tc, entry};
});
await Promise.all(entries.map(async ({tc, entry}: any) => {
const tool = tools.find(findByProp('name', tc.function.name));
if(options.stream) options.stream({tool: tc.function.name});
if(!tool) { entry.error = 'Tool not found'; return; }
try {
const args = JSONAttemptParse(toolCall.function.arguments, {});
const toolStream = options.stream && ((chunk: any) => {
if(chunk.done) { terminal = true; return; }
options.stream!(chunk);
});
const result = await tool.fn(args, toolStream, this.ai, toolCall.id);
return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result, timestamp: Date.now()};
const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
} catch(err: any) {
return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'}), timestamp: Date.now()};
entry.error = err?.message || err?.toString() || 'Unknown';
}
}));
history.push(...results);
requestParams.messages = history;
} else {
terminal = true;
const text = (msg.content || '').trim();
if(text) history.push({role: 'assistant', content: text, timestamp: Date.now(), duration, tps});
}
} while (!terminal && !controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
} while(!terminal && !controller.signal.aborted);
if(!terminal) {
const textContent = resp.choices[0].message.content || '';
history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
}
history = this.toStandard(history);
if(options.history) options.history.splice(0, options.history.length, ...history.filter(h => h.role !== 'system'));
if(options.stream) options.stream({done: true});
const turnStart = history.map(h => h.role).lastIndexOf('user');
const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
if(h.role === 'assistant') return str + (h.content || '');
return str;
}, '').trim();
const finalContent = history.slice(turnStart + 1).reduce((str, h) => h.role === 'assistant' ? str + (h.content || '') : str, '').trim();
res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
} catch(err) {
rej(err);
}
}), {abort: () => controller.abort()});
}
}

View File

@@ -1,167 +0,0 @@
import {describe, it, expect, vi, beforeEach} from 'vitest';
import LLM from '../src/llm';
const {FakeProvider, providerLog} = vi.hoisted(() => {
const providerLog: any[] = [];
class FakeProvider {
model: string;
constructor(...args: any[]) { this.model = args[args.length - 1]; }
ask(message: string, opts: any) {
let aborted = false;
const p = (async () => {
const script = (globalThis as any).__scripts?.[this.model];
const plan = script ? script(message, opts) : {text: ''};
providerLog.push({model: this.model, message, system: opts.system, tools: (opts.tools || []).map((t: any) => t.name)});
for (const c of plan.calls || []) {
if (aborted) break;
const tool = (opts.tools || []).find((t: any) => t.name === c.tool);
const id = c.id || `${c.tool}_${Math.random()}`;
const content = await tool.fn(c.args, opts.stream, null, id);
opts.history.push({role: 'tool', id, name: c.tool, args: c.args, content, timestamp: Date.now()});
}
const text = plan.text ?? '';
if (opts.stream && text) opts.stream({text, done: true});
opts.history.push({role: 'assistant', content: text, timestamp: Date.now(), duration: 10, tps: 5});
return text;
})();
return Object.assign(p, {abort: () => { aborted = true; }});
}
}
return {FakeProvider, providerLog};
});
vi.mock('../src/antrhopic.ts', () => ({Anthropic: FakeProvider}));
vi.mock('../src/open-ai.ts', () => ({OpenAi: FakeProvider}));
function makeAi(models: any) {
return {options: {llm: {models}}} as any;
}
beforeEach(() => {
providerLog.length = 0;
(globalThis as any).__scripts = {};
});
describe('LLM cross-provider interchangeability', () => {
it('runs identical tool calls the same way on an anthropic-backed model and an openai-backed model', async () => {
const ai = makeAi({
claude: {proto: 'anthropic', token: 'x'},
gpt: {proto: 'openai', token: 'y', host: 'http://local'},
});
const llm = new LLM(ai);
const calc = {
name: 'calc_add',
description: 'Add two numbers',
args: {a: {type: 'number', required: true}, b: {type: 'number', required: true}},
fn: (args: any) => String(args.a + args.b),
};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
(globalThis as any).__scripts.gpt = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
const historyA: any[] = [], historyB: any[] = [];
const respA = await llm.ask('add 2 and 3', {model: 'claude', tools: [calc], history: historyA});
const respB = await llm.ask('add 2 and 3', {model: 'gpt', tools: [calc], history: historyB});
expect(respA).toBe('Result: 5');
expect(respB).toBe('Result: 5');
expect(providerLog.find(l => l.model === 'claude')!.tools).toContain('calc_add');
expect(providerLog.find(l => l.model === 'gpt')!.tools).toContain('calc_add');
// tool timing gets recomputed from real execution regardless of proto
for (const h of [historyA.find(h => h.name === 'calc_add'), historyB.find(h => h.name === 'calc_add')]) {
expect(h.content).toBe('5');
expect(typeof h.duration).toBe('number');
expect(typeof h.tps).toBe('number');
}
});
it('lets the same shared history flow across model + proto swaps with different system prompts', async () => {
const ai = makeAi({
claude: {proto: 'anthropic', token: 'x'},
gpt: {proto: 'openai', token: 'y', host: 'http://local'},
});
const llm = new LLM(ai);
const history: any[] = [];
(globalThis as any).__scripts.claude = () => ({text: 'Hi from claude'});
(globalThis as any).__scripts.gpt = () => ({text: 'Hi from gpt'});
const r1 = await llm.ask('hello', {model: 'claude', system: 'You are terse.', history});
const r2 = await llm.ask('follow up', {model: 'gpt', system: 'You are verbose.', history});
expect(r1).toBe('Hi from claude');
expect(r2).toBe('Hi from gpt');
expect(history.filter(h => h.role === 'assistant').map(h => h.content)).toEqual(['Hi from claude', 'Hi from gpt']);
expect(providerLog[0].system).toContain('You are terse.');
expect(providerLog[1].system).toContain('You are verbose.');
});
it('exposes MCP tools the same way no matter which proto backs the model', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const mcp = [{name: 'weather', host: 'http://mcp.local'}];
global.fetch = vi.fn(async (url: string, opts?: any) => {
if (url.endsWith('/tools')) {
return {json: async () => ({tools: [{name: 'lookup', description: 'Look up weather', inputSchema: {properties: {city: {type: 'string'}}, required: ['city']}}]})} as any;
}
const body = JSON.parse(opts.body);
return {json: async () => ({content: [{text: `Sunny in ${body.arguments.city}`}]})} as any;
}) as any;
for (const model of ['claude', 'gpt']) {
(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'weather_lookup', args: {city: 'Rome'}}], text: 'done'});
const history: any[] = [];
await llm.ask('weather?', {model, mcp, history});
expect(history.find(h => h.name === 'weather_lookup')?.content).toBe('Sunny in Rome');
}
});
it('exposes and resolves skill documents identically across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const skills = [{name: 'Onboarding', description: 'How to onboard a user', content: 'Step 1...'}];
for (const model of ['claude', 'gpt']) {
(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'skill_read', args: {name: 'Onboarding'}}], text: 'done'});
const history: any[] = [];
await llm.ask('onboard me', {model, skills, history});
expect(history.find(h => h.name === 'skill_read')?.content).toContain('Step 1...');
}
});
it('delegate agent mutates the shared history directly and backfills the orchestrator response, across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const history: any[] = [{role: 'user', content: 'research quantum computing'}];
const researcher = {name: 'researcher', system: 'You research topics.', delegate: true, model: 'gpt'};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'agent_researcher', args: {}}], text: ''});
(globalThis as any).__scripts.gpt = () => ({text: 'Quantum computers use qubits.'});
const resp = await llm.ask('go', {model: 'claude', agents: [researcher], history});
expect(resp).toBe('Quantum computers use qubits.');
expect(history.some(h => h.role === 'assistant' && h.content === 'Quantum computers use qubits.')).toBe(true);
expect(history.find(h => h.name === 'agent_researcher')?.content).toBe('');
});
it('regular (non-delegate) subagent keeps its own isolated history separate from the parent, across protos', async () => {
const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
const llm = new LLM(ai);
const history: any[] = [];
const summarizer = {name: 'summarizer', system: 'You summarize text.', model: 'gpt'};
(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'subagent_summarizer', args: {context: 'a long article', instructions: 'summarize it'}}], text: 'Summary: short version'});
(globalThis as any).__scripts.gpt = () => ({text: 'short version'});
const resp = await llm.ask('summarize this', {model: 'claude', agents: [summarizer], history});
expect(resp).toBe('Summary: short version');
expect(history.find(h => h.name === 'subagent_summarizer')?.content).toBe('short version');
// isolated history - subagent's own assistant turn never leaks into the parent
expect(history.some(h => h.role === 'assistant' && h.content === 'short version')).toBe(false);
});
});

View File

@@ -1,256 +0,0 @@
import {describe, it, expect, vi, beforeEach} from 'vitest';
import {MemoryManager, MemoryCache, rebuildGraph, Memory} from '../src/memory';
function makeMemory(overrides: Partial<Memory> = {}): Memory {
return {
name: 'Test/Doc',
description: '',
content: '',
embedding: [],
links: [],
backlinks: [],
...overrides,
};
}
function makeLLM() {
return {
embedding: vi.fn(async (_text: string) => [{embedding: [1, 0, 0]}]),
ask: vi.fn(async () => undefined),
};
}
describe('rebuildGraph', () => {
it('extracts [[WikiLinks]] from content, excluding self-links', () => {
const a = makeMemory({name: 'A', content: '[[B]] and [[A]] and [[C]]'});
const b = makeMemory({name: 'B', content: 'no links here'});
const mem = [a, b];
rebuildGraph(mem);
expect(a.links).toEqual(['B', 'C']);
expect(b.links).toEqual([]);
});
it('computes backlinks only for links that resolve to a real node', () => {
const a = makeMemory({name: 'A', content: '[[B]] [[Missing]]'});
const b = makeMemory({name: 'B', content: ''});
const mem = [a, b];
rebuildGraph(mem);
expect(b.backlinks).toEqual(['A']);
expect(mem.find(m => m.name === 'Missing')).toBeUndefined();
});
it('resets stale backlinks on every rebuild (no leftover from a removed link)', () => {
const a = makeMemory({name: 'A', content: '[[B]]'});
const b = makeMemory({name: 'B', content: ''});
const mem = [a, b];
rebuildGraph(mem);
expect(b.backlinks).toEqual(['A']);
a.content = 'no more links';
rebuildGraph(mem);
expect(b.backlinks).toEqual([]);
});
});
describe('MemoryCache', () => {
it('finds nearest neighbor by embedding via KD-tree search', () => {
const close = makeMemory({name: 'Close', embedding: [1, 0, 0]});
const far = makeMemory({name: 'Far', embedding: [0, 0, 1]});
const cache = new MemoryCache([close, far]);
const results = cache.search([1, 0, 0], 1);
expect(results[0].name).toBe('Close');
});
it('rebuilds the tree on add/update/remove', () => {
const cache = new MemoryCache([makeMemory({name: 'A', embedding: [1, 0, 0]})]);
cache.add(makeMemory({name: 'B', embedding: [0, 1, 0]}));
expect(cache.search([0, 1, 0], 1)[0].name).toBe('B');
cache.remove('B');
expect(cache.search([0, 1, 0], 1)[0]?.name).not.toBe('B');
});
});
describe('MemoryManager.forget', () => {
it('removes the node and recomputes backlinks for the rest of the graph', () => {
const llm = makeLLM();
const mgr = new MemoryManager(llm);
const a = makeMemory({name: 'A', content: '[[B]]'});
const b = makeMemory({name: 'B', content: '[[C]]'});
const c = makeMemory({name: 'C', content: ''});
const mem = [a, b, c];
rebuildGraph(mem);
expect(c.backlinks).toEqual(['B']);
const ok = mgr.forget('B', mem);
expect(ok).toBe(true);
expect(mem.find(m => m.name === 'B')).toBeUndefined();
expect(a.links).toEqual(['B']);
expect(c.backlinks).toEqual([]);
});
it('returns false for an unknown name', () => {
const mgr = new MemoryManager(makeLLM());
expect(mgr.forget('Nope', [makeMemory({name: 'A'})])).toBe(false);
});
});
describe('MemoryManager.recollect', () => {
it('orders vector matches first, then expands one hop via links', async () => {
const llm = makeLLM();
llm.embedding.mockResolvedValue([{embedding: [1, 0, 0]}]);
const mgr = new MemoryManager(llm);
const near = makeMemory({name: 'Near', embedding: [1, 0, 0], content: '[[Linked]]'});
const linked = makeMemory({name: 'Linked', embedding: [0, 0, 1], content: ''});
const far = makeMemory({name: 'Far', embedding: [0, 1, 0], content: ''});
const mem = [near, linked, far];
rebuildGraph(mem);
const result = await mgr.recollect('query', mem, 1, 1);
expect(result.map(r => r.name)).toEqual(['Near', 'Linked']);
});
it('returns [] when there are no memories', async () => {
const mgr = new MemoryManager(makeLLM());
expect(await mgr.recollect('q', [])).toEqual([]);
});
});
describe('MemoryManager.memorize (fast path)', () => {
let llm: ReturnType<typeof makeLLM>;
let mgr: MemoryManager;
beforeEach(() => {
llm = makeLLM();
mgr = new MemoryManager(llm);
});
it('pushes a pending tool message, then resolves it to links once facts land', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) {
opts.tools[0].fn({destination: 'Projects/Oxide', facts: 'Uses a hybrid memory system'});
return undefined;
}
return {description: 'd', content: '# doc'};
});
const history: any[] = [{role: 'user', content: 'we use a hybrid memory system'}];
const touched = await mgr.memorize(history, [], {model: 'test'} as any);
const pending = history.find(h => h.name === 'memory_process');
expect(pending).toBeDefined();
expect(pending.content).toContain('[[Projects/Oxide]]');
expect(touched.map(t => t.name)).toEqual(['Projects/Oxide']);
});
it('creates a new node and appends facts under "## Facts" without calling the doc LLM', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) opts.tools[0].fn({destination: 'People/Sarah', facts: 'Works at Acme, Likes hiking'});
return undefined;
});
const mem: Memory[] = [];
await mgr.memorize([{role: 'user', content: 'Sarah works at Acme and likes hiking'}] as any, mem, {model: 'test'} as any);
const node = mem.find(m => m.name === 'People/Sarah')!;
expect(node).toBeDefined();
expect(node.content).toContain('## Facts');
expect(node.content).toContain('- Works at Acme');
expect(node.content).toContain('- Likes hiking');
// doc reconciler LLM (schema call) should NOT have been awaited synchronously in this fast path assertion
});
it('routes "journal" destination to Journal/{weekMonday}', async () => {
llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
if (opts.tools) opts.tools[0].fn({destination: 'journal', facts: 'Shipped v1'});
return undefined;
});
const mem: Memory[] = [];
const touched = await mgr.memorize([{role: 'user', content: 'shipped v1 today'}] as any, mem, {model: 'test'} as any);
expect(touched[0].name).toMatch(/^Journal\/\d{4}-\d{2}-\d{2}$/);
});
it('reports nothing to remember when no facts are extracted', async () => {
llm.ask.mockResolvedValue(undefined); // tools present but fn never called
const history: any[] = [{role: 'user', content: 'hey'}];
const touched = await mgr.memorize(history, [], {model: 'test'} as any);
expect(touched).toEqual([]);
expect(history.find(h => h.name === 'memory_process').content).toBe('Nothing worth remembering.');
});
it('returns [] and does nothing for an empty conversation', async () => {
const touched = await mgr.memorize([], [], {model: 'test'} as any);
expect(touched).toEqual([]);
expect(llm.ask).not.toHaveBeenCalled();
});
});
describe('MemoryManager reconcileVault', () => {
it('integrates the "## Facts" section via the doc LLM and removes it', async () => {
const llm = makeLLM();
llm.ask.mockResolvedValue({description: 'Tidy summary', content: '# Doc\n\nIntegrated fact.'});
const mgr = new MemoryManager(llm);
const node = makeMemory({
name: 'Projects/Oxide',
content: '---\nname: Projects/Oxide\n---\n\n# Doc\n\n## Facts\n- some raw fact\n',
});
const mem = [node];
await mgr.reconcileVault(mem, {model: 'test'} as any, 'all');
expect(node.content).not.toContain('## Facts');
expect(node.content).toContain('Integrated fact.');
expect(node.description).toBe('Tidy summary');
});
it('only targets docs with a pending Facts inbox when scope is "touched"', async () => {
const llm = makeLLM();
llm.ask.mockResolvedValue({description: 'd', content: '# clean'});
const mgr = new MemoryManager(llm);
const dirty = makeMemory({name: 'A', content: '## Facts\n- x'});
const clean = makeMemory({name: 'B', content: '# already tidy'});
await mgr.reconcileVault([dirty, clean], {model: 'test'} as any, 'touched');
expect(dirty.content).toContain('# clean'); // rewritten (frontmatter now wraps it)
expect(clean.content).toBe('# already tidy'); // untouched, never queued
});
});
describe('MemoryManager reconcile coalescing', () => {
it('coalesces a second call while one is in-flight: marks dirty, aborts, reuses the same task promise', () => {
const llm = makeLLM();
const abort = vi.fn();
let calls = 0;
llm.ask.mockImplementation(() => {
calls++;
const pending: any = new Promise(() => {}); // never resolves in this test
pending.abort = abort;
return pending;
});
const mgr: any = new MemoryManager(llm);
const node = makeMemory({name: 'Q', content: '# Q\n\n## Facts\n- f'});
const mem = [node];
const p1 = mgr.reconcile(node, mem, {model: 'test'});
const p2 = mgr.reconcile(node, mem, {model: 'test'});
expect(p2).toBe(p1); // same in-flight task, not a new queue entry
expect(abort).toHaveBeenCalledTimes(1); // second call aborted the in-flight request
expect(calls).toBe(1); // no second ask() fired synchronously — it'll rerun via the dirty loop
});
});