// S795 — SemanticSearch.ts: retrieval semantico con embedding pre-calcolati (P31-RI)
//
// ARCHITETTURA:
//   Dev-time: scripts/gen-repo-embeddings.mjs genera public/repo-embeddings.json
//             usando @xenova/transformers + all-MiniLM-L6-v2 (Node.js).
//   Runtime:  SemanticSearch.ts carica il JSON statico + embeds la query con
//             Transformers.js (WebAssembly, browser-native, ~23MB download cached IDB).
//             Fallback: Jaccard se embedding non disponibile o timeout.
//
// Integrazione: esportare `semanticGetRelevantFiles` come drop-in per `GetRelevantFilesFn`
// in relevantFilesLoader.ts e agentLoop.ts.

import type { RelevantFileHit } from "./relevantFilesLoader";

/** Struttura di un'entry nel JSON pre-calcolato */
export interface RepoEmbeddingEntry {
  /** path relativo alla root del repo */
  path:        string;
  /** descrizione sintetica del file (da commento header) */
  description: string;
  /** keywords per fallback Jaccard */
  keywords:    string[];
  /** embedding float32 all-MiniLM-L6-v2 (dim=384), omesso se not computed */
  embedding?:  number[];
}

// ── Cache in-memory ────────────────────────────────────────────────────────────
let _entries: RepoEmbeddingEntry[] | null = null;
let _embedderP: Promise<((text: string) => Promise<number[]>) | null> | null = null;

// ── Carica embeddings JSON dal bundle statico ─────────────────────────────────
const EMBEDDINGS_URL = "/repo-embeddings.json";

async function _loadEntries(): Promise<RepoEmbeddingEntry[]> {
  if (_entries) return _entries;
  try {
    const r = await fetch(EMBEDDINGS_URL, { cache: "force-cache" });
    if (!r.ok) throw new Error(`HTTP ${r.status}`);
    _entries = await r.json() as RepoEmbeddingEntry[];
    return _entries;
  } catch (e) {
    console.warn("[SemanticSearch] repo-embeddings.json non disponibile:", e);
    _entries = [];
    return [];
  }
}

// ── Carica l'embedder Transformers.js (lazy, singleton) ──────────────────────
async function _getEmbedder(): Promise<((text: string) => Promise<number[]>) | null> {
  if (_embedderP) return _embedderP;
  _embedderP = (async () => {
    try {
      // @xenova/transformers funziona nel browser via WebAssembly (ONNX Runtime).
      // Il modello all-MiniLM-L6-v2 (quantized) è ~23MB, cached in IndexedDB dopo il primo download.
      // @ts-ignore — CDN URL import, no type declarations available
      // eslint-disable-next-line @typescript-eslint/ban-ts-comment
      const { pipeline } = await import(/* webpackIgnore: true */ "https://cdn.jsdelivr.net/npm/@xenova/transformers@2.17.2/dist/transformers.min.js" as any) as any;
      const extractor = await pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2", {
        quantized: true,
        progress_callback: undefined,
      });
      return async (text: string): Promise<number[]> => {
        const out = await extractor(text, { pooling: "mean", normalize: true });
        return Array.from(out.data as Float32Array);
      };
    } catch (e) {
      console.warn("[SemanticSearch] Transformers.js non disponibile:", e);
      return null;
    }
  })();
  return _embedderP;
}

// ── Cosine similarity ─────────────────────────────────────────────────────────
function _cosine(a: number[], b: number[]): number {
  let dot = 0, na = 0, nb = 0;
  const len = Math.min(a.length, b.length);
  for (let i = 0; i < len; i++) { dot += a[i] * b[i]; na += a[i] ** 2; nb += b[i] ** 2; }
  if (na === 0 || nb === 0) return 0;
  return dot / (Math.sqrt(na) * Math.sqrt(nb));
}

// ── Fallback Jaccard (keyword overlap) ───────────────────────────────────────
function _jaccard(query: string, entry: RepoEmbeddingEntry): number {
  const qTokens = new Set(query.toLowerCase().split(/\W+/).filter(t => t.length > 2));
  if (qTokens.size === 0) return 0;
  const eTokens = new Set([
    ...entry.keywords,
    ...entry.path.toLowerCase().split(/[/._-]+/),
    ...entry.description.toLowerCase().split(/\W+/).filter(t => t.length > 2),
  ]);
  let inter = 0;
  qTokens.forEach(t => { if (eTokens.has(t)) inter++; });
  return inter / (qTokens.size + eTokens.size - inter);
}

// ── API pubblica ──────────────────────────────────────────────────────────────

/**
 * Cerca i file più semanticamente rilevanti alla query.
 * Drop-in per `GetRelevantFilesFn` di relevantFilesLoader.ts.
 *
 * @param query      testo della query utente
 * @param topK       numero max di risultati (default 6)
 * @param minScore   soglia minima di score (default 0.15)
 * @param timeoutMs  timeout per embedding (default 3000ms, poi fallback Jaccard)
 */
export async function semanticGetRelevantFiles(
  query:    string,
  topK      = 6,
  minScore  = 0.15,
  timeoutMs = 3000,
): Promise<RelevantFileHit[]> {
  const entries = await _loadEntries();
  if (entries.length === 0) return [];

  // Controlla se ci sono entries con embedding
  const hasEmbeddings = entries.some(e => e.embedding && e.embedding.length > 0);

  let scores: Array<{ entry: RepoEmbeddingEntry; score: number }>;

  if (hasEmbeddings) {
    // Tenta embedding della query con timeout
    const embedder = await Promise.race([
      _getEmbedder(),
      new Promise<null>(r => setTimeout(() => r(null), timeoutMs)),
    ]);

    if (embedder) {
      // Embedding semantico
      const qEmb = await embedder(query);
      scores = entries
        .filter(e => e.embedding && e.embedding.length > 0)
        .map(e => ({ entry: e, score: _cosine(qEmb, e.embedding!) }));
    } else {
      // Timeout → fallback Jaccard sull'embedding delle keywords
      console.info("[SemanticSearch] timeout embedder → fallback Jaccard");
      scores = entries.map(e => ({ entry: e, score: _jaccard(query, e) }));
    }
  } else {
    // JSON senza embedding → Jaccard puro (modalità keyword-only)
    scores = entries.map(e => ({ entry: e, score: _jaccard(query, e) }));
  }

  return scores
    .filter(s => s.score >= minScore)
    .sort((a, b) => b.score - a.score)
    .slice(0, topK)
    .map(s => ({
      path:    s.entry.path,
      content: s.entry.description,  // placeholder — il contenuto vero arriva da VFS
      score:   s.score,
    }));
}

/**
 * Precaricare l'embedder in background (chiama subito dopo il mount dell'app
 * per non pagare il cold-start durante la prima query utente).
 */
export function warmupEmbedder(): void {
  _getEmbedder().catch(() => { /* non-blocking */ });
}

/**
 * Stato del sistema embedding per diagnostica.
 */
export function getSemanticSearchStatus(): {
  entriesLoaded: number;
  hasEmbeddings: boolean;
  embedderReady: boolean;
} {
  return {
    entriesLoaded: _entries?.length ?? 0,
    hasEmbeddings: !!(_entries?.some(e => e.embedding && e.embedding.length > 0)),
    embedderReady: _embedderP !== null,
  };
}