Optimized cosineSimilarity

Re-organized functions and added semantic embeddings
2025-12-19 15:22:06 -05:00 · 2025-12-19 11:16:05 -05:00
9 changed files with 987 additions and 123 deletions
--- a/package-lock.json
+++ b/package-lock.json
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.22",
+	"version": "0.2.1",
 	"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",
--- a/src/ai.ts
+++ b/src/ai.ts
@@ -1,9 +1,6 @@
-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';
+import { Audio } from './audio.ts';
+import {Vision} from './vision.ts';

 export type AiOptions = LLMOptions & {
 	whisper?: {
@@ -20,108 +17,16 @@ export class Ai {
 	private downloads: {[key: string]: Promise<string>} = {};
 	private whisperModel!: string;

-	/** Large Language Models */
-	llm!: LLM;
+	/** Audio processing AI */
+	audio!: Audio;
+	/** Language processing AI */
+	language!: LLM;
+	/** Vision processing AI */
+	vision!: Vision;

 	constructor(public readonly options: AiOptions) {
-		this.llm = new LLM(this, options);
-		if(this.options.whisper?.binary) {
-			this.whisperModel = this.options.whisper?.model.endsWith('.bin') ? this.options.whisper?.model : this.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.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.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 ${code}`));
-				});
-			});
-		});
-		return {response, abort};
-	}
-
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
-	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		if(!model.endsWith('.bin')) model += '.bin';
-		const p = Path.join(this.options.whisper.path, model);
-		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
-		if(!!this.downloads[model]) return this.downloads[model];
-		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${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];
-	}
-
-	/**
-	 * 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}
+		this.audio = new Audio(this);
+		this.language = new LLM(this);
+		this.vision = new Vision(this);
 	}
 }
--- a/src/antrhopic.ts
+++ b/src/antrhopic.ts
@@ -53,7 +53,7 @@ export class Anthropic extends LLMProvider {
 		const response = new Promise<any>(async (res, rej) => {
 			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);
+			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,
--- a/src/audio.ts
+++ b/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];
+	}
+}
--- a/src/llm.ts
+++ b/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';
@@ -5,6 +6,7 @@ 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 */
@@ -81,12 +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) {
-		if(options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, options.anthropic.token, options.anthropic.model);
-		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
-		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.model);
+	constructor(public readonly ai: Ai) {
+		this.embedModel = pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+		if(ai.options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, ai.options.anthropic.token, ai.options.anthropic.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);
 	}

 	/**
@@ -99,11 +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;
-			else model = [this.options.model, (<any>this.options)[this.options.model]?.model];
+			if(typeof this.ai.options.model == 'object') model = this.ai.options.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]});
@@ -117,7 +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()) {
@@ -132,6 +136,69 @@ export class LLM {
 		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
 	}

+	cosineSimilarity(v1: number[], v2: number[]): number {
+		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
+		let dotProduct = 0, normA = 0, normB = 0;
+		for (let i = 0; i < v1.length; i++) {
+			dotProduct += v1[i] * v2[i];
+			normA += v1[i] * v1[i];
+			normB += v2[i] * v2[i];
+		}
+		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
+		return denominator === 0 ? 0 : dotProduct / denominator;
+	}
+
+	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),
+		})));
+	}
+
 	/**
 	 * Estimate variable as tokens
 	 * @param history Object to size
@@ -142,6 +209,23 @@ 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 v = vector(target);
+		const similarities = searchTerms.map(t => vector(t)).map(refVector => this.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
--- a/src/ollama.ts
+++ b/src/ollama.ts
@@ -46,7 +46,7 @@ export class Ollama extends LLMProvider {
 				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 = {
--- a/src/open-ai.ts
+++ b/src/open-ai.ts
@@ -65,7 +65,7 @@ export class OpenAi extends LLMProvider {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
 			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,
--- a/src/vision.ts
+++ b/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);
+			})
+		}
+	}
+}
Author	SHA1	Message	Date
ztimson	bb6933f0d5	Optimized cosineSimilarity All checks were successful Publish Library / Build NPM Project (push) Successful in 42s Details Publish Library / Tag Version (push) Successful in 7s Details	2025-12-19 15:22:06 -05:00
ztimson	435c6127b1	Re-organized functions and added semantic embeddings All checks were successful Publish Library / Build NPM Project (push) Successful in 46s Details Publish Library / Tag Version (push) Successful in 8s Details	2025-12-19 11:16:05 -05:00