// Embeddings locaux via transformers.js — multilingual-e5-small (384 dimensions).
// Gratuit, privé, hors-ligne après le premier téléchargement (OpenRouter n'offre pas d'embeddings).
// Convention E5 : préfixes "query: " et "passage: ".

import { resolve } from "node:path";

export const EMBEDDING_DIM = 384;
const MODEL_ID = "Xenova/multilingual-e5-small";

type FeatureExtractor = (texts: string[], opts: { pooling: "mean"; normalize: boolean }) => Promise<{
  tolist(): number[][];
}>;

let _extractor: Promise<FeatureExtractor> | null = null;

async function extractor(): Promise<FeatureExtractor> {
  if (!_extractor) {
    _extractor = (async () => {
      const { pipeline, env } = await import("@huggingface/transformers");
      env.cacheDir = resolve(process.cwd(), "data/models");
      const p = await pipeline("feature-extraction", MODEL_ID, { dtype: "fp32" });
      return p as unknown as FeatureExtractor;
    })();
  }
  return _extractor;
}

export async function embedPassages(texts: string[]): Promise<Float32Array[]> {
  const ex = await extractor();
  const out: Float32Array[] = [];
  const BATCH = 16;
  for (let i = 0; i < texts.length; i += BATCH) {
    const batch = texts.slice(i, i + BATCH).map((t) => "passage: " + t.slice(0, 2000));
    const res = await ex(batch, { pooling: "mean", normalize: true });
    for (const v of res.tolist()) out.push(Float32Array.from(v));
  }
  return out;
}

export async function embedQuery(text: string): Promise<Float32Array> {
  const ex = await extractor();
  const res = await ex(["query: " + text.slice(0, 2000)], { pooling: "mean", normalize: true });
  return Float32Array.from(res.tolist()[0]);
}

// --- conversions blob <-> vecteur (SQLite stocke des BLOB) ---
export function vecToBlob(v: Float32Array): Uint8Array {
  return new Uint8Array(v.buffer.slice(0), 0, v.length * 4);
}
export function blobToVec(b: Uint8Array): Float32Array {
  const buf = b.buffer.slice(b.byteOffset, b.byteOffset + b.byteLength);
  return new Float32Array(buf);
}

/** Similarité cosinus — les vecteurs E5 sont déjà normalisés → produit scalaire. */
export function cosine(a: Float32Array, b: Float32Array): number {
  let s = 0;
  for (let i = 0; i < a.length; i++) s += a[i] * b[i];
  return s;
}