parallel embedding cap

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.7.0",
+	"version": "0.7.2",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/audio.ts

@@ -40,9 +40,11 @@ export class Audio {
 			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;
-				const names = await this.ai.language.json(transcript, '{1: "Detected Name"}', {
+				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: 0.2,
+					temperature: 0.1,
 				});
 				Object.entries(names).forEach(([speaker, name]) => {
 					transcript = (<string>transcript).replaceAll(`[Speaker ${speaker}]`, `[${name}]`);

src/llm.ts

@@ -255,11 +255,12 @@ 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, parellel?: number} opts Options for embedding such as chunk sizes and parallel processing
 	 * @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, parallel?: number} = {}) {
+		let {maxTokens = 500, overlapTokens = 50, parallel = 1} = opts;
+
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
 				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
@@ -279,13 +280,19 @@ class LLM {
 				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),
-		})));
+		let i = 0, chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+		const next: Function = () => {
+			const index = i++;
+			if(index >= chunks.length) return;
+			const text = chunks[index];
+			return embed(text).then(embedding => {
+				results.push({index, embedding, text, tokens: this.estimateTokens(text)});
+				return next();
+			})
+		}
+
+		await Promise.all(Array(parallel).fill(null).map(() => next()));
+		return results.toSorted((a, b) => a.index - b.index);
 	}
 
 	/**