Whisper ASR

README.md

@@ -75,6 +75,7 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Instructions
 1. Install the package: `npm i @ztimson/ai-utils`
+2. For speaker diarization: `pip install pyannote.audio`
 
 </details>
 
@@ -90,8 +91,9 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Instructions
 1. Install the dependencies: `npm i`
-2. Build library: `npm build`
+2. For speaker diarization: `pip install pyannote.audio`
-3. Run unit tests: `npm test`
+3. Build library: `npm build`
+4. Run unit tests: `npm test`
 
 </details>
 

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.5.3",
+	"version": "0.7.11",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,14 +25,14 @@
 		"watch": "npx vite build --watch"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.67.0",
+		"@anthropic-ai/sdk": "^0.78.0",
 		"@tensorflow/tfjs": "^4.22.0",
 		"@xenova/transformers": "^2.17.2",
-		"@ztimson/node-utils": "^1.0.4",
+		"@ztimson/node-utils": "^1.0.7",
-		"@ztimson/utils": "^0.27.9",
+		"@ztimson/utils": "^0.28.13",
 		"cheerio": "^1.2.0",
-		"openai": "^6.6.0",
+		"openai": "^6.22.0",
-		"tesseract.js": "^6.0.1"
+		"tesseract.js": "^7.0.0"
 	},
 	"devDependencies": {
 		"@types/node": "^24.8.1",

src/ai.ts

@@ -8,26 +8,22 @@ export type AbortablePromise<T> = Promise<T> & {
 };
 
 export type AiOptions = {
+	/** Token to pull models from hugging face */
+	hfToken?: string;
 	/** Path to models */
 	path?: string;
-	/** Embedding model */
+	/** Whisper ASR model: ggml-tiny.en.bin, ggml-base.en.bin */
-	embedder?: string; // all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5
+	asr?: string;
+	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
+	embedder?: string;
 	/** Large language models, first is default */
 	llm?: Omit<LLMRequest, 'model'> & {
 		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
 	}
-	/** Tesseract OCR configuration */
+	/** OCR model: eng, eng_best, eng_fast */
-	tesseract?: {
+	ocr?: string;
-		/** Model: eng, eng_best, eng_fast */
+	/** Whisper binary */
-		model?: string;
+	whisper?: string;
-	}
-	/** Whisper ASR configuration */
-	whisper?: {
-		/** Whisper binary location */
-		binary: string;
-		/** Model: `ggml-base.en.bin` */
-		model: string;
-	}
 }
 
 export class Ai {

src/antrhopic.ts

@@ -13,25 +13,25 @@ export class Anthropic extends LLMProvider {
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
+		const timestamp = Date.now();
-			const orgI = i;
+		const messages: LLMMessage[] = [];
-			if(typeof history[orgI].content != 'string') {
+		for(let h of history) {
-				if(history[orgI].role == 'assistant') {
+			if(typeof h.content == 'string') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
+				messages.push(<any>{timestamp, ...h});
-						history.splice(i + 1, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
+			} else {
-					});
+				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				} else if(history[orgI].role == 'user') {
+				if(textContent) messages.push({timestamp, role: h.role, content: textContent});
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
+				h.content.forEach((c: any) => {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
+					if(c.type == 'tool_use') {
-						h[c.is_error ? 'error' : 'content'] = c.content;
+						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
-					});
+					} else if(c.type == 'tool_result') {
-				}
+						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
-				if(!history[orgI].content) history.splice(orgI, 1);
+					}
+				});
 			}
-			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
 		}
-		return history.filter(h => !!h.content);
+		return messages;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -50,8 +50,8 @@ export class Anthropic extends LLMProvider {
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res, rej) => {
+		return Object.assign(new Promise<any>(async (res) => {
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -73,7 +73,6 @@ export class Anthropic extends LLMProvider {
 			};
 
 			let resp: any, isFirstMessage = true;
-			const assistantMessages: string[] = [];
 			do {
 				resp = await this.client.messages.create(requestParams).catch(err => {
 					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
@@ -119,7 +118,6 @@ export class Anthropic extends LLMProvider {
 						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
-							console.log(typeof tool.fn);
 							const result = await tool.fn(toolCall.input, options?.stream, this.ai);
 							return {type: 'tool_result', tool_use_id: toolCall.id, content: JSONSanitize(result)};
 						} catch (err: any) {
@@ -131,7 +129,7 @@ export class Anthropic extends LLMProvider {
 				}
 			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
 			history.push({role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')});
-			this.toStandard(history);
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/audio.ts

@@ -1,39 +1,163 @@
-import {spawn} from 'node:child_process';
+import {execSync, spawn} from 'node:child_process';
-import fs from 'node:fs/promises';
+import {mkdtempSync} from 'node:fs';
-import Path from 'node:path';
+import fs, {rm} from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import Path, {join} from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
 
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
+	private pyannote!: string;
 	private whisperModel!: string;
 
 	constructor(private ai: Ai) {
-		if(ai.options.whisper?.binary) {
+		if(ai.options.whisper) {
-			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
 			this.downloadAsrModel();
 		}
+
+		this.pyannote = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${ai.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${ai.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
 	}
 
-	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
+	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
-		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		let proc: any;
-		let abort: any = () => {};
+		const p = new Promise<any>((resolve, reject) => {
-		const p = new Promise<string | null>(async (resolve, reject) => {
+			this.downloadAsrModel(opts.model).then(m => {
-			const m = await this.downloadAsrModel(model);
+				let output = '';
-			let output = '';
+				const args = [opts.diarization ? '-owts' : '-nt', '-m', m, '-f', file];
-			const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
+				proc = spawn(<string>this.ai.options.whisper, args, {stdio: ['ignore', 'pipe', 'ignore']});
-			abort = () => proc.kill('SIGTERM');
+				proc.on('error', (err: Error) => reject(err));
-			proc.on('error', (err: Error) => reject(err));
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
-			proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
-			proc.on('close', (code: number) => {
+					if(code === 0) {
-				if(code === 0) resolve(output.trim() || null);
+						if(opts.diarization) {
-				else reject(new Error(`Exit code ${code}`));
+							try { resolve(JSON.parse(output)); }
+							catch(e) { reject(new Error('Failed to parse whisper JSON')); }
+						} else {
+							resolve(output.trim() || null);
+						}
+					} else {
+						reject(new Error(`Exit code ${code}`));
+					}
+				});
 			});
 		});
-		return Object.assign(p, {abort});
+		return <any>Object.assign(p, {abort: () => proc?.kill('SIGTERM')});
+	}
+
+	private runDiarization(file: string): AbortablePromise<any> {
+		let aborted = false, abort = () => { aborted = true; };
+		const checkPython = (cmd: string) => {
+			return new Promise<boolean>((resolve) => {
+				const proc = spawn(cmd, ['-c', 'import pyannote.audio']);
+				proc.on('close', (code: number) => resolve(code === 0));
+				proc.on('error', () => resolve(false));
+			});
+		};
+		const p = Promise.all<any>([
+			checkPython('python'),
+			checkPython('python3'),
+		]).then(<any>(async ([p, p3]: [boolean, boolean]) => {
+			if(aborted) return;
+			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
+			const binary = p3 ? 'python3' : 'python';
+			return new Promise((resolve, reject) => {
+				if(aborted) return;
+				let output = '';
+				const proc = spawn(binary, ['-c', this.pyannote, file]);
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
+				proc.on('close', (code: number) => {
+					if(code === 0) {
+						try { resolve(JSON.parse(output)); }
+						catch (err) { reject(new Error('Failed to parse diarization output')); }
+					} else {
+						reject(new Error(`Python process exited with code ${code}`));
+					}
+				});
+				proc.on('error', reject);
+				abort = () => proc.kill('SIGTERM');
+			});
+		}));
+		return <any>Object.assign(p, {abort});
+	}
+
+	private combineSpeakerTranscript(transcript: any, speakers: any[]): string {
+		const speakerMap = new Map();
+		let speakerCount = 0;
+		speakers.forEach((seg: any) => {
+			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
+		});
+
+		const lines: string[] = [];
+		let currentSpeaker = -1;
+		let currentText = '';
+		transcript.transcription.forEach((word: any) => {
+			const time = word.offsets.from / 1000; // Convert ms to seconds
+			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()}`);
+				currentSpeaker = speakerNum;
+				currentText = word.text;
+			} else {
+				currentText += ' ' + word.text;
+			}
+		});
+		if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
+		return lines.join('\n');
+	}
+
+	asr(file: string, options: { model?: string; diarization?: boolean | 'id' } = {}): AbortablePromise<string | null> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		const clean = () => rm(Path.dirname(tmp), { recursive: true, force: true }).catch(() => {});
+		const transcript = this.runAsr(tmp, {model: options.model, diarization: !!options.diarization});
+		const diarization: any = options.diarization ? this.runDiarization(tmp) : Promise.resolve(null);
+		let aborted = false, abort = () => {
+			aborted = true;
+			transcript.abort();
+			diarization?.abort?.();
+			clean();
+		};
+
+		const response = Promise.all([transcript, diarization]).then(async ([t, d]) => {
+			if(aborted || !options.diarization) return t;
+			t = this.combineSpeakerTranscript(t, d);
+			if(!aborted && options.diarization === 'id') {
+				if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
+				let chunks = this.ai.language.chunk(t, 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", 2: "Second Name"}', {
+					system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
+					temperature: 0.1,
+				});
+				Object.entries(names).forEach(([speaker, name]) => t = t.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
+			}
+			return t;
+		}).finally(() => clean());
+		return <any>Object.assign(response, {abort});
 	}
 
 	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
 		if(!model.endsWith('.bin')) model += '.bin';
 		const p = Path.join(<string>this.ai.options.path, model);
 		if(await fs.stat(p).then(() => true).catch(() => false)) return p;

src/embedder.ts

@@ -1,11 +1,13 @@
 import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
 
-let embedder: any;
+const [modelDir, model] = process.argv.slice(2);
 
-parentPort?.on('message', async ({ id, text, model }) => {
+let text = '';
-	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true});
+process.stdin.on('data', chunk => text += chunk);
+process.stdin.on('end', async () => {
+	const 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 });
+	console.log(JSON.stringify({embedding}));
+	process.exit();
 });

src/index.ts

@@ -1,7 +1,6 @@
 export * from './ai';
 export * from './antrhopic';
 export * from './audio';
-export * from './embedder'
 export * from './llm';
 export * from './open-ai';
 export * from './provider';

src/llm.ts

@@ -4,9 +4,9 @@ import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
+import { spawn } from 'node:child_process';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
 export type OllamaConfig = {proto: 'ollama', host: string};
@@ -75,22 +75,10 @@ export type LLMRequest = {
 }
 
 class LLM {
-	private embedWorker: Worker | null = null;
+	defaultModel!: string;
-	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
+	models: {[model: string]: LLMProvider} = {};
-	private embedId = 0;
-	private models: {[model: string]: LLMProvider} = {};
-	private defaultModel!: string;
 
 	constructor(public readonly ai: Ai) {
-		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-		this.embedWorker.on('message', ({ id, embedding }) => {
-			const pending = this.embedQueue.get(id);
-			if (pending) {
-				pending.resolve(embedding);
-				this.embedQueue.delete(id);
-			}
-		});
-
 		if(!ai.options.llm?.models) return;
 		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
@@ -196,7 +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}`)]);
@@ -262,25 +255,57 @@ 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 {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
-	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<any[]> {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
-				const id = this.embedId++;
+				if(aborted) return reject(new Error('Aborted'));
-				this.embedQueue.set(id, { resolve, reject });
+
-				this.embedWorker?.postMessage({ id, text, model: this.ai.options?.embedder || 'bge-small-en-v1.5' });
+				const args: string[] = [
+					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
+					<string>this.ai.options.path,
+					this.ai.options?.embedder || 'bge-small-en-v1.5'
+				];
+				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
+				proc.stdin.write(text);
+				proc.stdin.end();
+
+				let output = '';
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(aborted) return reject(new Error('Aborted'));
+					if(code === 0) {
+						try {
+							const result = JSON.parse(output);
+							resolve(result.embedding);
+						} catch(err) {
+							reject(new Error('Failed to parse embedding output'));
+						}
+					} else {
+						reject(new Error(`Embedder process exited with code ${code}`));
+					}
+				});
+				proc.on('error', reject);
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
+
-		return Promise.all(chunks.map(async (text, index) => ({
+		const p = (async () => {
-			index,
+			const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
-			embedding: await embed(text),
+			for(let i = 0; i < chunks.length; i++) {
-			text,
+				if(aborted) break;
-			tokens: this.estimateTokens(text),
+				const text = chunks[i];
-		})));
+				const embedding = await embed(text);
+				results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+			}
+			return results;
+		})();
+		return Object.assign(p, { abort });
 	}
 
 	/**
@@ -312,12 +337,16 @@ 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> {
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(message, {system: 'Respond using a JSON blob matching any provided examples', ...options});
+		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;

src/open-ai.ts

@@ -68,7 +68,7 @@ export class OpenAi extends LLMProvider {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
 			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -133,7 +133,7 @@ export class OpenAi extends LLMProvider {
 				}
 			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
 			history.push({role: 'assistant', content: resp.choices[0].message.content || ''});
-			this.toStandard(history);
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/vision.ts

@@ -3,7 +3,7 @@ import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
-	constructor(private ai: Ai) { }
+	constructor(private ai: Ai) {}
 
 	/**
 	 * Convert image to text using Optical Character Recognition
@@ -13,7 +13,7 @@ export class Vision {
 	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
 		const p = new Promise<string | null>(async res => {
-			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
+			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);

vite.config.ts

@@ -1,6 +1,5 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {