Re-organized functions and added semantic embeddings

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.8",
+	"version": "0.2.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -27,6 +27,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.67.0",
 		"@tensorflow/tfjs": "^4.22.0",
+		"@xenova/transformers": "^2.17.2",
 		"@ztimson/node-utils": "^1.0.4",
 		"@ztimson/utils": "^0.27.9",
 		"ollama": "^0.6.0",

src/ai.ts

@@ -1,125 +1,32 @@
-import {$} from '@ztimson/node-utils';
-import {createWorker} from 'tesseract.js';
 import {LLM, LLMOptions} from './llm';
-import fs from 'node:fs/promises';
+import { Audio } from './audio.ts';
-import Path from 'node:path';
+import {Vision} from './vision.ts';
-import * as tf from '@tensorflow/tfjs';
 
 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;
 
-	/** Large Language Models */
+	/** Audio processing AI */
-	llm!: LLM;
+	audio!: Audio;
+	/** Language processing AI */
+	language!: LLM;
+	/** Vision processing AI */
+	vision!: Vision;
 
 	constructor(public readonly options: AiOptions) {
-		this.llm = new LLM(this, options);
+		this.audio = new Audio(this);
-		if(this.options.whisper?.binary) {
+		this.language = new LLM(this);
-			this.whisperModel = Path.join(<string>this.options.whisper?.path, this.options.whisper?.model + this.options.whisper?.model.endsWith('.bin') ? '' : '.bin');
+		this.vision = new Vision(this);
-			this.downloadAsrModel();
-		}
-	}
-
-	/**
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	async asr(path: string, model?: WhisperModel): Promise<string | null> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		await this.downloadAsrModel(model);
-		const name = Math.random().toString(36).substring(2, 10) + '-' + path.split('/').pop() + '.txt';
-		const output = Path.join(this.options.whisper.temp || '/tmp', name);
-		console.log(`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${model ? Path.join(this.options.whisper.path, model) : this.whisperModel} -f ${path} -otxt -of ${output}`);
-		await $`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${model ? Path.join(this.options.whisper.path, model) : this.whisperModel} -f ${path} -otxt -of ${output}`;
-		return fs.readFile(output, 'utf-8').then(text => text?.trim() || null)
-			.finally(() => fs.rm(output, {force: true}).catch(() => {}));
-	}
-
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
-	async downloadAsrModel(model?: string): Promise<string> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		let m;
-		if(model) m = model?.endsWith('.bin') ? model : model + '.bin';
-		else m = <string>this.whisperModel.split('/').at(-1);
-		const p = Path.join(this.options.whisper.path, m);
-		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
-		if(!!this.downloads[m]) return this.downloads[m];
-		this.downloads[m] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${m}`)
-			.then(resp => resp.arrayBuffer())
-			.then(arr => Buffer.from(arr)).then(async buffer => {
-				await fs.writeFile(Path.join((<any>this.options.whisper).path, m), buffer);
-				delete this.downloads[m];
-				return p;
-			});
-		return this.downloads[m];
-	}
-
-	/**
-	 * Convert image to text using Optical Character Recognition
-	 * @param {string} path Path to image
-	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
-	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
-		let worker: any;
-		return {
-			abort: () => { worker?.terminate(); },
-			response: new Promise(async res => {
-				worker = await createWorker('eng');
-				const {data} = await worker.recognize(path);
-				await worker.terminate();
-				res(data.text.trim() || null);
-			})
-		}
-	}
-
-	/**
-	 * Compare the difference between two strings using tensor math
-	 * @param target Text that will checked
-	 * @param {string} searchTerms Multiple search terms to check against target
-	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
-	 */
-	semanticSimilarity(target: string, ...searchTerms: string[]) {
-		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-
-		const vector = (text: string, dimensions: number = 10): number[] => {
-			return text.toLowerCase().split('').map((char, index) =>
-				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
-		}
-
-		const cosineSimilarity = (v1: number[], v2: number[]): number => {
-			if (v1.length !== v2.length) throw new Error('Vectors must be same length');
-			const tensor1 = tf.tensor1d(v1), tensor2 = tf.tensor1d(v2)
-			const dotProduct = tf.dot(tensor1, tensor2)
-			const magnitude1 = tf.norm(tensor1)
-			const magnitude2 = tf.norm(tensor2)
-			if(magnitude1.dataSync()[0] === 0 || magnitude2.dataSync()[0] === 0) return 0
-			return dotProduct.dataSync()[0] / (magnitude1.dataSync()[0] * magnitude2.dataSync()[0])
-		}
-
-		const v = vector(target);
-		const similarities = searchTerms.map(t => vector(t)).map(refVector => cosineSimilarity(v, refVector))
-		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
 	}
 }

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);
 	}
@@ -44,14 +45,15 @@ 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()}]);
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
+			const original = deepCopy(history);
+			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
 			const requestParams: any = {
 				model: options.model || this.model,
 				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
@@ -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/audio.ts

@@ -0,0 +1,63 @@
+import {spawn} from 'node:child_process';
+import fs from 'node:fs/promises';
+import Path from 'node:path';
+import {Ai} from './ai.ts';
+
+export class Audio {
+	private downloads: {[key: string]: Promise<string>} = {};
+	private whisperModel!: string;
+
+	constructor(private ai: Ai) {
+		if(ai.options.whisper?.binary) {
+			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
+			this.downloadAsrModel();
+		}
+	}
+
+	/**
+	 * Convert audio to text using Auditory Speech Recognition
+	 * @param {string} path Path to audio
+	 * @param model Whisper model
+	 * @returns {Promise<any>} Extracted text
+	 */
+	asr(path: string, model: string = this.whisperModel): {abort: () => void, response: Promise<string | null>} {
+		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		let abort: any = () => {};
+		const response = new Promise<string | null>((resolve, reject) => {
+			this.downloadAsrModel(model).then(m => {
+				let output = '';
+				const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
+				abort = () => proc.kill('SIGTERM');
+				proc.on('error', (err: Error) => reject(err));
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(code === 0) resolve(output.trim() || null);
+					else reject(new Error(`Exit code ${code}`));
+				});
+			});
+		});
+		return {response, abort};
+	}
+
+	/**
+	 * Downloads the specified Whisper model if it is not already present locally.
+	 *
+	 * @param {string} model Whisper model that will be downloaded
+	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
+	 */
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
+		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		if(!model.endsWith('.bin')) model += '.bin';
+		const p = Path.join(this.ai.options.whisper.path, model);
+		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
+		if(!!this.downloads[model]) return this.downloads[model];
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
+			.then(resp => resp.arrayBuffer())
+			.then(arr => Buffer.from(arr)).then(async buffer => {
+				await fs.writeFile(p, buffer);
+				delete this.downloads[model];
+				return p;
+			});
+		return this.downloads[model];
+	}
+}

src/llm.ts

@@ -1,3 +1,4 @@
+import {pipeline} from '@xenova/transformers';
 import {JSONAttemptParse} from '@ztimson/utils';
 import {Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
@@ -5,12 +6,15 @@ import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
 import {AbortablePromise, LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
+import * as tf from '@tensorflow/tfjs';
 
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Timestamp */
+	timestamp?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -24,6 +28,8 @@ export type LLMMessage = {
 	content: undefined | string;
 	/** Tool error */
 	error: undefined | string;
+	/** Timestamp */
+	timestamp?: number;
 }
 
 export type LLMOptions = {
@@ -77,12 +83,14 @@ export type LLMRequest = {
 }
 
 export class LLM {
+	private embedModel: any;
 	private providers: {[key: string]: LLMProvider} = {};
 
-	constructor(public readonly ai: Ai, public readonly options: LLMOptions) {
+	constructor(public readonly ai: Ai) {
-		if(options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, options.anthropic.token, options.anthropic.model);
+		this.embedModel = pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
-		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
+		if(ai.options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, ai.options.anthropic.token, ai.options.anthropic.model);
-		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.model);
+		if(ai.options.ollama?.host) this.providers.ollama = new Ollama(this.ai, ai.options.ollama.host, ai.options.ollama.model);
+		if(ai.options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, ai.options.openAi.token, ai.options.openAi.model);
 	}
 
 	/**
@@ -95,11 +103,11 @@ export class LLM {
 		let model: any = [null, null];
 		if(options.model) {
 			if(typeof options.model == 'object') model = options.model;
-			else model = [options.model, (<any>this.options)[options.model]?.model];
+			else model = [options.model, (<any>this.ai.options)[options.model]?.model];
 		}
 		if(!options.model || model[1] == null) {
-			if(typeof this.options.model == 'object') model = this.options.model;
+			if(typeof this.ai.options.model == 'object') model = this.ai.options.model;
-			else model = [this.options.model, (<any>this.options)[this.options.model]?.model];
+			else model = [this.ai.options.model, (<any>this.ai.options)[this.ai.options.model]?.model];
 		}
 		if(!model[0] || !model[1]) throw new Error(`Unknown LLM provider or model: ${model[0]} / ${model[1]}`);
 		return this.providers[model[0]].ask(message, {...options, model: model[1]});
@@ -113,7 +121,7 @@ export class LLM {
 	 * @param {LLMRequest} options LLM options
 	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
 	 */
-	async compress(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
 		if(this.estimateTokens(history) < max) return history;
 		let keep = 0, tokens = 0;
 		for(let m of history.toReversed()) {
@@ -125,7 +133,58 @@ export class LLM {
 		const recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
 		const summary = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
-		return [{role: 'assistant', content: `Conversation Summary: ${summary}`}, ...recent];
+		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
+	}
+
+	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+		const objString = (obj: any, path = ''): string[] => {
+			if(obj === null || obj === undefined) return [];
+			return Object.entries(obj).flatMap(([key, value]) => {
+				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
+				if(typeof value === 'object' && value !== null && !Array.isArray(value)) return objString(value, p);
+				const valueStr = Array.isArray(value) ? value.join(', ') : String(value);
+				return `${p}: ${valueStr}`;
+			});
+		};
+
+		const embed = async (text: string): Promise<number[]> => {
+			const model = await this.embedModel;
+			const output = await model(text, {pooling: 'mean', normalize: true});
+			return Array.from(output.data);
+		};
+
+		// Tokenize
+		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
+		const tokens = lines.flatMap(line => [...line.split(/\s+/).filter(w => w.trim()), '\n']);
+
+		// Chunk
+		const chunks: string[] = [];
+		let start = 0;
+		while (start < tokens.length) {
+			let end = start;
+			let text = '';
+			// Build chunk
+			while (end < tokens.length) {
+				const nextToken = tokens[end];
+				const testText = text + (text ? ' ' : '') + nextToken;
+				const testTokens = this.estimateTokens(testText.replace(/\s*\n\s*/g, '\n'));
+				if (testTokens > maxTokens && text) break;
+				text = testText;
+				end++;
+			}
+			// Save chunk
+			const cleanText = text.replace(/\s*\n\s*/g, '\n').trim();
+			if(cleanText) chunks.push(cleanText);
+			start = end - overlapTokens;
+			if (start <= end - tokens.length + end) start = end; // Safety: prevent infinite loop
+		}
+
+		return Promise.all(chunks.map(async (text, index) => ({
+			index,
+			embedding: await embed(text),
+			text,
+			tokens: this.estimateTokens(text),
+		})));
 	}
 
 	/**
@@ -138,6 +197,35 @@ export class LLM {
 		return Math.ceil((text.length / 4) * 1.2);
 	}
 
+	/**
+	 * Compare the difference between two strings using tensor math
+	 * @param target Text that will checked
+	 * @param {string} searchTerms Multiple search terms to check against target
+	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
+	 */
+	fuzzyMatch(target: string, ...searchTerms: string[]) {
+		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
+
+		const vector = (text: string, dimensions: number = 10): number[] => {
+			return text.toLowerCase().split('').map((char, index) =>
+				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
+		}
+
+		const cosineSimilarity = (v1: number[], v2: number[]): number => {
+			if (v1.length !== v2.length) throw new Error('Vectors must be same length');
+			const tensor1 = tf.tensor1d(v1), tensor2 = tf.tensor1d(v2)
+			const dotProduct = tf.dot(tensor1, tensor2)
+			const magnitude1 = tf.norm(tensor1)
+			const magnitude2 = tf.norm(tensor2)
+			if(magnitude1.dataSync()[0] === 0 || magnitude2.dataSync()[0] === 0) return 0
+			return dotProduct.dataSync()[0] / (magnitude1.dataSync()[0] * magnitude2.dataSync()[0])
+		}
+
+		const v = vector(target);
+		const similarities = searchTerms.map(t => vector(t)).map(refVector => cosineSimilarity(v, refVector))
+		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
+	}
+
 	/**
 	 * Ask a question with JSON response
 	 * @param {string} message Question

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}
 		});
 	}
@@ -39,12 +41,12 @@ 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();
 			}
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min);
+			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min);
 			if(options.system) history.unshift({role: 'system', content: system})
 
 			const requestParams: any = {
@@ -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[]);
@@ -62,8 +64,8 @@ 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);
+			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
 
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -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()});
 	}
 }

src/vision.ts

@@ -0,0 +1,25 @@
+import {createWorker} from 'tesseract.js';
+import {Ai} from './ai.ts';
+
+export class Vision {
+
+	constructor(private ai: Ai) { }
+
+	/**
+	 * Convert image to text using Optical Character Recognition
+	 * @param {string} path Path to image
+	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
+	 */
+	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
+		let worker: any;
+		return {
+			abort: () => { worker?.terminate(); },
+			response: new Promise(async res => {
+				worker = await createWorker('eng');
+				const {data} = await worker.recognize(path);
+				await worker.terminate();
+				res(data.text.trim() || null);
+			})
+		}
+	}
+}