One embedding at a time

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.6.7",
+	"version": "0.7.3",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/ai.ts

@@ -8,6 +8,8 @@ export type AbortablePromise<T> = Promise<T> & {
 };
 
 export type AiOptions = {
+	/** Token to pull models from hugging face */
+	hfToken?: string;
 	/** Path to models */
 	path?: string;
 	/** ASR model: whisper-tiny, whisper-base */

src/asr.ts

@@ -1,44 +1,43 @@
 import { pipeline } from '@xenova/transformers';
 import { parentPort } from 'worker_threads';
-import * as fs from 'node:fs';
-import wavefile from 'wavefile';
 import { spawn } from 'node:child_process';
+import { execSync } from 'node:child_process';
+import { mkdtempSync, rmSync, readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
+import wavefile from 'wavefile';
 
 let whisperPipeline: any;
 
 export async function canDiarization(): Promise<boolean> {
 	return new Promise((resolve) => {
-		const proc = spawn('python3', ['-c', 'import pyannote.audio']);
+		const proc = spawn('python', ['-c', 'import pyannote.audio']);
 		proc.on('close', (code: number) => resolve(code === 0));
 		proc.on('error', () => resolve(false));
 	});
 }
 
-async function runDiarization(audioPath: string, torchHome: string): Promise<any[]> {
+async function runDiarization(audioPath: string, dir: string, token: string): Promise<any[]> {
 	const script = `
 import sys
 import json
 import os
 from pyannote.audio import Pipeline
 
-os.environ['TORCH_HOME'] = "${torchHome}"
-pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
-diarization = pipeline(sys.argv[1])
+os.environ['TORCH_HOME'] = r"${dir}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${token}")
+output = pipeline(sys.argv[1])
 
 segments = []
-for turn, _, speaker in diarization.itertracks(yield_label=True):
-    segments.append({
-        "start": turn.start,
-        "end": turn.end,
-        "speaker": speaker
-    })
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
 
 print(json.dumps(segments))
 `;
 
 	return new Promise((resolve, reject) => {
 		let output = '';
-		const proc = spawn('python3', ['-c', script, audioPath]);
+		const proc = spawn('python', ['-c', script, audioPath]);
 		proc.stdout.on('data', (data: Buffer) => output += data.toString());
 		proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
 		proc.on('close', (code: number) => {
@@ -71,54 +70,65 @@ function combineSpeakerTranscript(chunks: any[], speakers: any[]): string {
 		const speaker = speakers.find((s: any) => time >= s.start && time <= s.end);
 		const speakerNum = speaker ? speakerMap.get(speaker.speaker) : 1;
 		if (speakerNum !== currentSpeaker) {
-			if(currentText) lines.push(`[speaker ${currentSpeaker}]: ${currentText.trim()}`);
+			if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
 			currentSpeaker = speakerNum;
 			currentText = chunk.text;
 		} else {
 			currentText += chunk.text;
 		}
 	});
-	if(currentText) lines.push(`[speaker ${currentSpeaker}]: ${currentText.trim()}`);
+	if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
 	return lines.join('\n');
 }
 
-parentPort?.on('message', async ({ file, speaker, model, modelDir }) => {
+function prepareAudioBuffer(file: string): [string, Float32Array] {
+	let wav: any, tmp;
 	try {
-		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
-
-		// Prepare audio file (convert to mono channel wave)
-		const wav = new wavefile.WaveFile(fs.readFileSync(file));
+		wav = new wavefile.WaveFile(readFileSync(file));
+	} catch(err) {
+		tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		wav = new wavefile.WaveFile(readFileSync(tmp));
+	} finally {
 		wav.toBitDepth('32f');
 		wav.toSampleRate(16000);
 		const samples = wav.getSamples();
-		let buffer;
-		if(Array.isArray(samples)) { // stereo to mono - average the channels
+		if(Array.isArray(samples)) {
 			const left = samples[0];
 			const right = samples[1];
-			buffer = new Float32Array(left.length);
+			const buffer = new Float32Array(left.length);
 			for (let i = 0; i < left.length; i++) buffer[i] = (left[i] + right[i]) / 2;
-		} else {
-			buffer = samples;
+			return [tmp || file, buffer];
 		}
-
-		// Transcribe
-		const transcriptResult = await whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false});
-		if(!speaker) {
-			parentPort?.postMessage({ text: transcriptResult.text?.trim() || null });
-			return;
+		return [tmp || file, samples];
 	}
+}
 
-		// Speaker Diarization
-		const hasDiarization = await canDiarization();
-		if(!hasDiarization) {
-			parentPort?.postMessage({ text: transcriptResult.text?.trim() || null, error: 'Speaker diarization unavailable' });
-			return;
-		}
+parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
+	try {
+		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
 
-		const speakers = await runDiarization(file, modelDir);
-		const combined = combineSpeakerTranscript(transcriptResult.chunks || [], speakers);
+		// Prepare audio file
+		const [f, buffer] = prepareAudioBuffer(file);
+
+		// Fetch transcript and speakers
+		const hasDiarization = speaker && await canDiarization();
+		const [transcript, speakers] = await Promise.all([
+			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
+			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(f, modelDir, token),
+		]);
+		if(file != f) rmSync(f, { recursive: true, force: true });
+
+		// Return any results / errors if no more processing required
+		const text = transcript.text?.trim() || null;
+		if(!speaker) return parentPort?.postMessage({ text });
+		if(!token) return parentPort?.postMessage({ text, error: 'HuggingFace token required' });
+		if(!hasDiarization) return parentPort?.postMessage({ text, error: 'Speaker diarization unavailable' });
+
+		// Combine transcript and speakers
+		const combined = combineSpeakerTranscript(transcript.chunks || [], speakers || []);
 		parentPort?.postMessage({ text: combined });
-	} catch (err) {
-		parentPort?.postMessage({ error: (err as Error).message });
+	} catch (err: any) {
+		parentPort?.postMessage({ error: err.stack || err.message });
 	}
 });

src/audio.ts

@@ -1,18 +1,19 @@
+import {fileURLToPath} from 'url';
 import {Worker} from 'worker_threads';
-import Path from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
 import {canDiarization} from './asr.ts';
+import {dirname, join} from 'path';
 
 export class Audio {
 	constructor(private ai: Ai) {}
 
-	asr(file: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
+	asr(file: string, options: { model?: string; speaker?: boolean | 'id' } = {}): AbortablePromise<string | null> {
 		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
 		let aborted = false;
 		const abort = () => { aborted = true; };
 
-		const p = new Promise<string | null>((resolve, reject) => {
-			const worker = new Worker(Path.join(import.meta.dirname, 'asr.js'));
+		let p = new Promise<string | null>((resolve, reject) => {
+			const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
 			const handleMessage = ({ text, warning, error }: any) => {
 				worker.terminate();
 				if(aborted) return;
@@ -31,8 +32,27 @@ export class Audio {
 			worker.on('exit', (code) => {
 				if(code !== 0 && !aborted) reject(new Error(`Worker exited with code ${code}`));
 			});
-			worker.postMessage({file, model, speaker, modelDir: this.ai.options.path});
+			worker.postMessage({file, model, speaker, modelDir: this.ai.options.path, token: this.ai.options.hfToken});
 		});
+
+		// Name speakers using AI
+		if(options.speaker == 'id') {
+			if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
+			p = p.then(async transcript => {
+				if(!transcript) return transcript;
+				let chunks = this.ai.language.chunk(transcript, 500, 0);
+				if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
+				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name"}', {
+					system: 'Use this following transcript to identify speakers. Only identify speakers you are sure about',
+					temperature: 0.1,
+				});
+				Object.entries(names).forEach(([speaker, name]) => {
+					transcript = (<string>transcript).replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
+				});
+				return transcript;
+			})
+		}
+
 		return Object.assign(p, { abort });
 	}
 

src/embedder.ts

@@ -3,9 +3,9 @@ import { parentPort } from 'worker_threads';
 
 let embedder: any;
 
-parentPort?.on('message', async ({ id, text, model, modelDir }) => {
+parentPort?.on('message', async ({text, model, modelDir }) => {
 	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	parentPort?.postMessage({ id, embedding });
+	parentPort?.postMessage({embedding});
 });

src/llm.ts

@@ -75,8 +75,8 @@ export type LLMRequest = {
 }
 
 class LLM {
-	private models: {[model: string]: LLMProvider} = {};
-	private defaultModel!: string;
+	defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
 
 	constructor(public readonly ai: Ai) {
 		if(!ai.options.llm?.models) return;
@@ -184,7 +184,12 @@ class LLM {
 		const system = history[0].role == 'system' ? history[0] : null,
 			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
-		const summary: any = await this.json(`Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn't already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI. Match this format: {summary: string, facts: [[subject, fact]]}\n\n${process.map(m => `${m.role}: ${m.content}`).join('\n\n')}`, {model: options?.model, temperature: options?.temperature || 0.3});
+
+		const summary: any = await this.json(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+			model: options?.model,
+			temperature: options?.temperature || 0.3
+		});
 		const timestamp = new Date();
 		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
 			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
@@ -250,11 +255,11 @@ class LLM {
 	/**
 	 * Create a vector representation of a string
 	 * @param {object | string} target Item that will be embedded (objects get converted)
-	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
-	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
+	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}) {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
 				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
@@ -271,16 +276,16 @@ class LLM {
 				worker.on('exit', (code) => {
 					if(code !== 0) reject(new Error(`Worker exited with code ${code}`));
 				});
-				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', path: this.ai.options.path});
+				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', modelDir: this.ai.options.path});
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
-		return Promise.all(chunks.map(async (text, index) => ({
-			index,
-			embedding: await embed(text),
-			text,
-			tokens: this.estimateTokens(text),
-		})));
+		const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+		for(let i = 0; i < chunks.length; i++) {
+			const text= chunks[i];
+			const embedding = await embed(text);
+			results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+		}
+		return results;
 	}
 
 	/**
@@ -312,12 +317,16 @@ class LLM {
 
 	/**
 	 * Ask a question with JSON response
-	 * @param {string} message Question
+	 * @param {string} text Text to process
+	 * @param {string} schema JSON schema the AI should match
 	 * @param {LLMRequest} options Configuration options and chat history
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
-	async json(message: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(message, {system: 'Respond using a JSON blob matching any provided examples', ...options});
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
+		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
+\`\`\`json
+${schema}
+\`\`\``});
 		if(!resp) return {};
 		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
 		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;