// The local text-embedding model behind semantic search: bge-small-en-v1.5
// (quantized, ~34 MB) run with transformers.js. Nothing is sent anywhere; the model
// is downloaded once into MODELS_DIR (the Docker image ships with it).
const path = require('path');

const MODEL = 'Xenova/bge-small-en-v1.5';
const DIMENSIONS = 384;
const CACHE_DIR = process.env.MODELS_DIR || path.join(__dirname, '..', 'models');
// bge models match short queries against passages best with this instruction.
const QUERY_PREFIX = 'Represent this sentence for searching relevant passages: ';
const IDLE_MS = 10 * 60 * 1000;

let pipelinePromise = null;
let idleTimer = null;

// Loaded on first use and released after a while unused (it holds ~150 MB).
function load() {
  if (!pipelinePromise) {
    pipelinePromise = (async () => {
      const { pipeline, env } = await import('@huggingface/transformers');
      env.cacheDir = CACHE_DIR;
      return pipeline('feature-extraction', MODEL, { dtype: 'q8' });
    })();
    pipelinePromise.catch(() => { pipelinePromise = null; });
  }
  clearTimeout(idleTimer);
  idleTimer = setTimeout(release, IDLE_MS);
  idleTimer.unref();
  return pipelinePromise;
}

async function release() {
  const loaded = pipelinePromise;
  pipelinePromise = null;
  try { await (await loaded).dispose(); } catch { /* already gone */ }
}

// texts -> unit-length vectors (arrays of numbers).
async function embed(texts) {
  if (!texts.length) return [];
  const extractor = await load();
  const output = await extractor(texts, { pooling: 'cls', normalize: true });
  return output.tolist();
}

module.exports = { MODEL, DIMENSIONS, QUERY_PREFIX, load, embed };