hybrid-retrieval（混合检索）

Hybrid Retrieval for RAG

Combine dense (semantic) and sparse (keyword) retrieval for superior results.

When to Use

• Vector search misses exact keyword matches
• Domain-specific terminology needs exact matching
• Users search with both natural language and specific terms
• Need to balance semantic understanding with precision

The Problem with Vector-Only Search

Query: "Error code E-4521 troubleshooting"

Vector search returns:
- "Common error handling patterns" (semantically similar)
- "Debugging techniques for applications" (related topic)

Missing:
- "E-4521: Database connection timeout" (exact match needed!)

Hybrid Architecture

┌─────────────────────────────────────────────────┐
│                   User Query                     │
└─────────────────────┬───────────────────────────┘
                      │
         ┌────────────┴────────────┐
         │                         │
         ▼                         ▼
┌─────────────────┐      ┌─────────────────┐
│  Dense Search   │      │  Sparse Search  │
│  (Embeddings)   │      │  (BM25/TF-IDF)  │
└────────┬────────┘      └────────┬────────┘
         │                         │
         └────────────┬────────────┘
                      │
                      ▼
              ┌───────────────┐
              │    Fusion     │
              │  (RRF/Linear) │
              └───────┬───────┘
                      │
                      ▼
              ┌───────────────┐
              │   Reranker    │
              │  (Optional)   │
              └───────┬───────┘
                      │
                      ▼
              ┌───────────────┐
              │ Final Results │
              └───────────────┘

Implementation

Basic Hybrid with LangChain

from langchain.retrievers import EnsembleRetriever
from langchain_community.retrievers import BM25Retriever
from langchain_community.vectorstores import Chroma

# Dense retriever (vector search)
vectorstore = Chroma.from_documents(docs, embeddings)
dense_retriever = vectorstore.as_retriever(search_kwargs={"k": 10})

# Sparse retriever (BM25)
bm25_retriever = BM25Retriever.from_documents(docs)
bm25_retriever.k = 10

# Combine with ensemble
hybrid_retriever = EnsembleRetriever(
    retrievers=[dense_retriever, bm25_retriever],
    weights=[0.5, 0.5]  # Adjust based on your data
)

results = hybrid_retriever.invoke("Error code E-4521")

Reciprocal Rank Fusion (RRF)

def reciprocal_rank_fusion(results_list: list, k: int = 60) -> list:
    """
    Combine multiple ranked lists using RRF.
    k=60 is the standard constant from the original paper.
    """
    fused_scores = {}

    for results in results_list:
        for rank, doc in enumerate(results):
            doc_id = doc.metadata.get("id", hash(doc.page_content))
            if doc_id not in fused_scores:
                fused_scores[doc_id] = {"doc": doc, "score": 0}
            fused_scores[doc_id]["score"] += 1 / (k + rank + 1)

    # Sort by fused score
    reranked = sorted(
        fused_scores.values(),
        key=lambda x: x["score"],
        reverse=True
    )
    return [item["doc"] for item in reranked]

# Usage
dense_results = dense_retriever.invoke(query)
sparse_results = bm25_retriever.invoke(query)
final_results = reciprocal_rank_fusion([dense_results, sparse_results])

With Pinecone (Native Hybrid)

from pinecone import Pinecone
from pinecone_text.sparse import BM25Encoder

# Initialize
pc = Pinecone(api_key="...")
index = pc.Index("hybrid-index")

# Sparse encoder
bm25 = BM25Encoder()
bm25.fit(corpus)

# Query with both dense and sparse
def hybrid_query(query: str, alpha: float = 0.5):
    # Dense vector
    dense_vec = embeddings.embed_query(query)

    # Sparse vector
    sparse_vec = bm25.encode_queries([query])[0]

    # Hybrid search
    results = index.query(
        vector=dense_vec,
        sparse_vector=sparse_vec,
        top_k=10,
        alpha=alpha,  # 0 = sparse only, 1 = dense only
        include_metadata=True
    )
    return results

With Weaviate (Native Hybrid)

import weaviate

client = weaviate.Client("http://localhost:8080")

result = client.query.get(
    "Document",
    ["content", "title"]
).with_hybrid(
    query="Error code E-4521",
    alpha=0.5,  # Balance between vector and keyword
    fusion_type="rankedFusion"
).with_limit(10).do()

Adding a Reranker

from sentence_transformers import CrossEncoder

# Load reranker model
reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')

def rerank_results(query: str, docs: list, top_k: int = 5) -> list:
    """Rerank documents using cross-encoder."""
    pairs = [[query, doc.page_content] for doc in docs]
    scores = reranker.predict(pairs)

    # Sort by reranker scores
    scored_docs = list(zip(docs, scores))
    scored_docs.sort(key=lambda x: x[1], reverse=True)

    return [doc for doc, score in scored_docs[:top_k]]

# Full pipeline
hybrid_results = hybrid_retriever.invoke(query)  # Get 20 results
final_results = rerank_results(query, hybrid_results, top_k=5)  # Rerank to top 5

Weight Tuning Guidelines

Evaluation

def evaluate_retrieval(queries: list, ground_truth: dict, retriever) -> dict:
    """Calculate retrieval metrics."""
    metrics = {"mrr": 0, "recall@5": 0, "precision@5": 0}

    for query in queries:
        results = retriever.invoke(query)
        result_ids = [doc.metadata["id"] for doc in results[:5]]
        relevant_ids = ground_truth[query]

        # MRR
        for i, rid in enumerate(result_ids):
            if rid in relevant_ids:
                metrics["mrr"] += 1 / (i + 1)
                break

        # Recall & Precision
        hits = len(set(result_ids) & set(relevant_ids))
        metrics["recall@5"] += hits / len(relevant_ids)
        metrics["precision@5"] += hits / 5

    # Average
    n = len(queries)
    return {k: v/n for k, v in metrics.items()}

Best Practices

1. Start with 50/50 weights - then tune based on evaluation
2. Always add a reranker - significant quality improvement
3. Index sparse vectors - BM25 on raw text, not chunks
4. Use native hybrid - when available (Pinecone, Weaviate, Qdrant)
5. Monitor both paths - log which retriever contributed to final results