Search

Role: optional self-hosted local inference gateway for text/audio/vision/embeddings without changing NCA PRIMARY SPECIALTY or authority.

const OPENAI_EMBEDDINGS_MODEL = "text-embedding-3-small";
const PINECONE_API_VERSION = "2024-10";
async function embedQuery(query: string, apiKey: string): Promise<number[]> {
  const res = await fetch("https://api.openai.com/v1/embeddings", {
    method: "POST",
    },
    body: JSON.stringify({ model: OPENAI_EMBEDDINGS_MODEL, input: query }),
  });
  if (!res.ok) {
    throw new Error(`OpenAI embeddings request failed (${res.status}): ${await res.text()}`);
  }

  /^\/api\/v1\/chat\/completions$/,
  /^\/api\/v1\/embeddings$/,
  /^\/api\/v1\/models$/,

  const data = await resp.json();
  return data.embeddings.map((e: any) => e.values);
}
  const allTexts = [query, ...candidates];
  const embeddings = await embedBatch(allTexts);
  const queryVec = embeddings[0];
  const candVecs = embeddings.slice(1);

      text:
        "NVIDIA embeddings help convert text into vector representations for similarity search",
      metadata: { category: "ai", tags: ["nvidia", "embeddings"] },
    },

import { GpuContext } from "./gpu-core.ts";
import { Embeddings } from "./model-embeddings.ts";
import { TransformerBlock } from "./model-block.ts";
export class Transformer {
  emb: Embeddings;
  blocks: TransformerBlock[];
    this.cfg = cfg;
    this.emb = new Embeddings(cfg.vocabSize, cfg.dim, cfg.maxSeqLen);
    this.blocks = Array.from(

export class Embeddings {
  token: TokenEmbedding;

  // Embeddings: tok + pos
  const x0: number[][] = [];

  // Embeddings: tok + pos
  const x0: number[][] = [];

  // Embeddings
  const x0: number[][] = [];

A simple interface for making and querying Pinecone vector databases. Use OpenAI
embeddings to vectorize and search

async function generateQueryEmbedding(query: string): Promise<number[]> {
  const response = await openai.embeddings.create({
    model: EMBEDDING_MODEL,

const PINECONE_INDEX_NAME = Deno.env.get("PINECONE_INDEX_NAME") ||
  "aws-ring-embeddings-small";
const PINECONE_DIMENSIONS = parseInt(
);
const OPENAI_EMBEDDINGS_MODEL = Deno.env.get("OPENAI_EMBEDDINGS_MODEL") ||
  "text-embedding-3-small";
async function embedQuestion(question: string): Promise<number[]> {
  const response = await openai.embeddings.create({
    model: OPENAI_EMBEDDINGS_MODEL,
    input: question,
    // Step 1: embeddings
    const { result: embedding, ms: embedMs } = await timeStep(

import type { Page, SearchOptions, SearchResult } from "./types.ts";
// import { searchStrategy, generateEmbeddings } from "./placeholder.ts";
// import { searchStrategy, generateEmbeddings } from "./jigsawstack-orama.ts"; // ~550ms query embeddings (no API key needed, works on Val Town)
// import { searchStrategy, generateEmbeddings } from "./openai-orama.ts"; // ~100-200ms query embeddings (API), ~29ms DB overhead
// import { searchStrategy, generateEmbeddings } from "./openai-cosine.ts"; // ~100-200ms query embeddings (API), <1ms cosine calc
// import { searchStrategy, generateEmbeddings } from "./mixedbread-embeddings-cosine.ts"; // Mixedbread Embeddings API (mxbai-embed-large-v1) with local cosine similarity
// import { searchStrategy, generateEmbeddings } from "./mixedbread.ts"; // Mixedbread Stores (managed search service, handles storage + embeddings + search)
// import { searchStrategy, generateEmbeddings } from "./hf-inference-qwen3-cosine.ts"; // HF Inference Client with Qwen3-Embedding-8B (API, requires HF_TOKEN)
import { searchStrategy, generateEmbeddings } from "./cloudflare-bge-cosine.ts"; // Cloudflare Workers AI (bge-large-en-v1.5, 1024-dim) with local cosine similarity
// import { searchStrategy, generateEmbeddings } from "./transformers-cosine.ts"; // ~10-30ms (local embeddings, ~23MB model cached after first load)
// import { searchStrategy, generateEmbeddings } from "./transformers-local-onnx.ts"; // ~10-30ms (pre-downloaded local ONNX models, NO downloads, NO isolate loading)
// Embeddings function for recalculate - uses same import as search strategy
export const getActiveEmbeddingsFunction = (): ((content: string) => Promise<number[] | null>) => {
  return generateEmbeddings;
};

// Transformers.js + Cosine Similarity Strategy: Local embeddings in browser/Deno
// Fastest option - no API calls, runs entirely client-side
      // Load feature extraction pipeline with all-MiniLM-L6-v2 model
      // This is a small, fast model optimized for embeddings (~86MB, ~384 dims)
      // Downloads from Hugging Face on first run, then uses cache (~150ms vs several seconds)
// Generate embeddings using transformers.js (runs locally, no API calls)
export const generateEmbeddings = async (content: string): Promise<number[] | null> => {
  try {
    
    // Generate embeddings - returns a tensor, we need to extract the array
    const output = await pipeline(content, {
    const errorMessage = error instanceof Error ? error.message : String(error);
    console.error("Transformers.js embeddings failed:", errorMessage);
    return null;
  name: "transformers-cosine",
  description: "Semantic search using Transformers.js local embeddings with cosine similarity (fastest, no API calls)",
  search: async (query: string, pages: Page[], options: SearchOptions = {}): Promise<SearchResult[]> => {
    const queryEmbedStart = performance.now();
    const queryEmbedding = await generateEmbeddings(query);
    if (enableTiming) {
    for (const page of pages) {
      const pageEmbedding = page.embeddings;
      if (!pageEmbedding || pageEmbedding.length !== queryEmbedding.length) {
        continue; // Skip pages without embeddings or wrong dimension
      }