Naive RAG

# ╔══════════════════════════════════════════════════════════════╗
# ║  Naive RAG — শুধু Dense Vector Search (baseline)            ║
# ╚══════════════════════════════════════════════════════════════╝

def retrieve_naive(question: str, subject: str = "", class_no: str = "",
                   k: int = TOP_K, use_metadata_filter: bool = True) -> dict:
    """
    Naive RAG: শুধু BGE-M3 dense search।
    ❌ Multi-query  ❌ BM25  ❌ RRF  ❌ Reranker
    ✅ শুধু একটা embedding search
    """
    norm_subj = normalize_subject(subject)

    # ── Metadata filter (hybrid-এর মতোই) ─────────────────────
    where_filter = None
    if use_metadata_filter and norm_subj:
        where_filter = {"subject": {"$eq": norm_subj}}
        if class_no:
            where_filter = {
                "$and": [
                    {"subject": {"$eq": norm_subj}},
                    {"class_no": {"$eq": str(class_no)}},
                ]
            }

    # ── একটাই dense query (multi-query নেই) ──────────────────
    query_prefix = "Represent this sentence for searching relevant passages: "
    q_emb = embed_model.encode(
        [query_prefix + question], normalize_embeddings=True
    ).tolist()

    query_kwargs = {
        "query_embeddings": q_emb,
        "n_results": k,                 # সরাসরি top-k (candidate pool নেই)
        "include": ["documents", "metadatas", "distances"],
    }
    if where_filter:
        query_kwargs["where"] = where_filter

    try:
        hits = collection.query(**query_kwargs)
    except:
        query_kwargs.pop("where", None)
        hits = collection.query(**query_kwargs)

    docs  = hits["documents"][0]
    metas = hits["metadatas"][0]

    if not docs:
        return {"context": "", "sources": [], "num_chunks": 0, "texts": []}

    # ── Context build (char limit) ───────────────────────────
    final_docs, final_metas, total = [], [], 0
    for doc, meta in zip(docs, metas):
        if total + len(doc) > CONTEXT_MAX_CHARS:
            remaining = CONTEXT_MAX_CHARS - total
            if remaining > 200:
                final_docs.append(doc[:remaining] + "...")
                final_metas.append(meta)
            break
        final_docs.append(doc)
        final_metas.append(meta)
        total += len(doc)

    context = "\n\n---\n\n".join(final_docs)

    return {
        "context": context,
        "sources": final_metas,
        "num_chunks": len(final_docs),
        "texts": final_docs,            # hit rate calculation-এর জন্য
    }

print("✅ retrieve_naive() ready — dense search only")