Fixed LLM multi message responses

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.9",
+	"version": "0.1.22",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/ai.ts

@@ -1,25 +1,21 @@
-import {$} from '@ztimson/node-utils';
 import {createWorker} from 'tesseract.js';
 import {LLM, LLMOptions} from './llm';
 import fs from 'node:fs/promises';
 import Path from 'node:path';
 import * as tf from '@tensorflow/tfjs';
+import {spawn} from 'node:child_process';
 
 export type AiOptions = LLMOptions & {
 	whisper?: {
 		/** Whisper binary location */
 		binary: string;
-		/** Model */
+		/** Model: `ggml-base.en.bin` */
-		model: WhisperModel;
+		model: string;
 		/** Path to models */
 		path: string;
-		/** Path to storage location for temporary files */
-		temp?: string;
 	}
 }
 
-export type WhisperModel = 'tiny' | 'base' | 'small' | 'medium' | 'large';
-
 export class Ai {
 	private downloads: {[key: string]: Promise<string>} = {};
 	private whisperModel!: string;
@@ -30,7 +26,7 @@ export class Ai {
 	constructor(public readonly options: AiOptions) {
 		this.llm = new LLM(this, options);
 		if(this.options.whisper?.binary) {
-			this.whisperModel = Path.join(<string>this.options.whisper?.path, this.options.whisper?.model + this.options.whisper?.model.endsWith('.bin') ? '' : '.bin');
+			this.whisperModel = this.options.whisper?.model.endsWith('.bin') ? this.options.whisper?.model : this.options.whisper?.model + '.bin';
 			this.downloadAsrModel();
 		}
 	}
@@ -41,17 +37,23 @@ export class Ai {
 	 * @param model Whisper model
 	 * @returns {Promise<any>} Extracted text
 	 */
-	async asr(path: string, model?: WhisperModel): Promise<string | null> {
+	asr(path: string, model: string = this.whisperModel): {abort: () => void, response: Promise<string | null>} {
 		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		const m = await this.downloadAsrModel(model);
+		let abort: any = () => {};
-		const name = Math.random().toString(36).substring(2, 10) + '-' + path.split('/').pop() + '.txt';
+		const response = new Promise<string | null>((resolve, reject) => {
-		const output = Path.join(this.options.whisper.temp || '/tmp', name);
+			this.downloadAsrModel(model).then(m => {
-		console.log('model:', this.options.whisper?.model);
+				let output = '';
-		console.log(this.whisperModel);
+				const proc = spawn(<string>this.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
-		console.log(`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${m} -f ${path} -otxt -of ${output}`);
+				abort = () => proc.kill('SIGTERM');
-		await $`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${m} -f ${path} -otxt -of ${output}`;
+				proc.on('error', (err: Error) => reject(err));
-		return fs.readFile(output, 'utf-8').then(text => text?.trim() || null)
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
-			.finally(() => fs.rm(output, {force: true}).catch(() => {}));
+				proc.on('close', (code: number) => {
+					if(code === 0) resolve(output.trim() || null);
+					else reject(new Error(`Exit code ${code}`));
+				});
+			});
+		});
+		return {response, abort};
 	}
 
 	/**
@@ -60,22 +62,20 @@ export class Ai {
 	 * @param {string} model Whisper model that will be downloaded
 	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
 	 */
-	async downloadAsrModel(model?: string): Promise<string> {
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
 		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		let m;
+		if(!model.endsWith('.bin')) model += '.bin';
-		if(model) m = model?.endsWith('.bin') ? model : model + '.bin';
+		const p = Path.join(this.options.whisper.path, model);
-		else m = <string>this.whisperModel.split('/').at(-1);
-		const p = Path.join(this.options.whisper.path, m);
 		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
-		if(!!this.downloads[m]) return this.downloads[m];
+		if(!!this.downloads[model]) return this.downloads[model];
-		this.downloads[m] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${m}`)
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
 			.then(resp => resp.arrayBuffer())
 			.then(arr => Buffer.from(arr)).then(async buffer => {
-				await fs.writeFile(Path.join((<any>this.options.whisper).path, m), buffer);
+				await fs.writeFile(p, buffer);
-				delete this.downloads[m];
+				delete this.downloads[model];
 				return p;
 			});
-		return this.downloads[m];
+		return this.downloads[model];
 	}
 
 	/**

src/antrhopic.ts

@@ -1,5 +1,5 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
 import {Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {AbortablePromise, LLMProvider} from './provider.ts';
@@ -19,7 +19,7 @@ export class Anthropic extends LLMProvider {
 				if(history[orgI].role == 'assistant') {
 					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
 						i++;
-						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input});
+						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
 					});
 				} else if(history[orgI].role == 'user') {
 					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
@@ -29,6 +29,7 @@ export class Anthropic extends LLMProvider {
 				}
 				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
 			}
+			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
 		}
 		return history.filter(h => !!h.content);
 	}
@@ -44,13 +45,14 @@ export class Anthropic extends LLMProvider {
 				i++;
 			}
 		}
-		return history;
+		return history.map(({timestamp, ...h}) => h);
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			const original = deepCopy(history);
 			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -71,13 +73,15 @@ export class Anthropic extends LLMProvider {
 				stream: !!options.stream,
 			};
 
-			// Run tool changes
+			let resp: any, isFirstMessage = true;
-			let resp: any;
+			const assistantMessages: string[] = [];
 			do {
 				resp = await this.client.messages.create(requestParams);
 
 				// Streaming mode
 				if(options.stream) {
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
 					resp.content = [];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
@@ -108,6 +112,7 @@ export class Anthropic extends LLMProvider {
 				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push({role: 'assistant', content: resp.content});
+					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = options.tools?.find(findByProp('name', toolCall.name));
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
@@ -122,12 +127,11 @@ export class Anthropic extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {
+			res(this.toStandard([...history, {role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')}]));
-				role: 'assistant',
-				content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')
-			}]));
 		});
+
 		return Object.assign(response, {abort: () => controller.abort()});
 	}
 }

src/llm.ts

@@ -11,6 +11,8 @@ export type LLMMessage = {
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Timestamp */
+	timestamp?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -24,6 +26,8 @@ export type LLMMessage = {
 	content: undefined | string;
 	/** Tool error */
 	error: undefined | string;
+	/** Timestamp */
+	timestamp?: number;
 }
 
 export type LLMOptions = {
@@ -125,7 +129,7 @@ export class LLM {
 		const recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
 		const summary = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
-		return [{role: 'assistant', content: `Conversation Summary: ${summary}`}, ...recent];
+		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
 	}
 
 	/**

src/ollama.ts

@@ -22,15 +22,17 @@ export class Ollama extends LLMProvider {
 				}
 			} else if(history[i].role == 'tool') {
 				const error = history[i].content.startsWith('{"error":');
-				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content};
+				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content, timestamp: history[i].timestamp};
 			}
+			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
 		}
 		return history;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
 		return history.map((h: any) => {
-			if(h.role != 'tool') return h;
+			const {timestamp, ...rest} = h;
+			if(h.role != 'tool') return rest;
 			return {role: 'tool', tool_name: h.name, content: h.error || h.content}
 		});
 	}
@@ -39,7 +41,7 @@ export class Ollama extends LLMProvider {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
 			let system = options.system || this.ai.options.system;
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			if(history[0].roll == 'system') {
 				if(!system) system = history.shift();
 				else history.shift();
@@ -70,11 +72,12 @@ export class Ollama extends LLMProvider {
 				}))
 			}
 
-			// Run tool chains
+			let resp: any, isFirstMessage = true;
-			let resp: any;
 			do {
 				resp = await this.client.chat(requestParams);
 				if(options.stream) {
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
 					resp.message = {role: 'assistant', content: '', tool_calls: []};
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
@@ -87,7 +90,6 @@ export class Ollama extends LLMProvider {
 					}
 				}
 
-				// Run tools
 				if(resp.message?.tool_calls?.length && !controller.signal.aborted) {
 					history.push(resp.message);
 					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
@@ -105,9 +107,11 @@ export class Ollama extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.message?.tool_calls?.length);
+
 			if(options.stream) options.stream({done: true});
 			res(this.toStandard([...history, {role: 'assistant', content: resp.message?.content}]));
 		});
+
 		return Object.assign(response, {abort: () => controller.abort()});
 	}
 }

src/open-ai.ts

@@ -20,7 +20,8 @@ export class OpenAi extends LLMProvider {
 					role: 'tool',
 					id: tc.id,
 					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {})
+					args: JSONAttemptParse(tc.function.arguments, {}),
+					timestamp: h.timestamp
 				}));
 				history.splice(i, 1, ...tools);
 				i += tools.length - 1;
@@ -33,7 +34,7 @@ export class OpenAi extends LLMProvider {
 				history.splice(i, 1);
 				i--;
 			}
-
+			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
 		}
 		return history;
 	}
@@ -46,14 +47,15 @@ export class OpenAi extends LLMProvider {
 					content: null,
 					tool_calls: [{ id: h.id, type: 'function', function: { name: h.name, arguments: JSON.stringify(h.args) } }],
 					refusal: null,
-					annotations: [],
+					annotations: []
 				}, {
 					role: 'tool',
 					tool_call_id: h.id,
 					content: h.error || h.content
 				});
 			} else {
-				result.push(h);
+				const {timestamp, ...rest} = h;
+				result.push(rest);
 			}
 			return result;
 		}, [] as any[]);
@@ -62,7 +64,7 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
 
 			const requestParams: any = {
@@ -85,23 +87,25 @@ export class OpenAi extends LLMProvider {
 				}))
 			};
 
-			// Tool call and streaming logic similar to other providers
+			let resp: any, isFirstMessage = true;
-			let resp: any;
 			do {
 				resp = await this.client.chat.completions.create(requestParams);
-
-				// Implement streaming and tool call handling
 				if(options.stream) {
-					resp.choices = [];
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
+					resp.choices = [{message: {content: '', tool_calls: []}}];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
 						if(chunk.choices[0].delta.content) {
+							resp.choices[0].message.content += chunk.choices[0].delta.content;
 							options.stream({text: chunk.choices[0].delta.content});
 						}
+						if(chunk.choices[0].delta.tool_calls) {
+							resp.choices[0].message.tool_calls = chunk.choices[0].delta.tool_calls;
+						}
 					}
 				}
 
-				// Run tools
 				const toolCalls = resp.choices[0].message.tool_calls || [];
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push(resp.choices[0].message);
@@ -124,7 +128,6 @@ export class OpenAi extends LLMProvider {
 			if(options.stream) options.stream({done: true});
 			res(this.toStandard([...history, {role: 'assistant', content: resp.choices[0].message.content || ''}]));
 		});
-
 		return Object.assign(response, {abort: () => controller.abort()});
 	}
 }