🔀 Hybrid Retrieval

Combine sparse (BM25) and dense retrieval for state-of-the-art results using Reciprocal Rank Fusion (RRF).

Quick Start

from rankify.retrievers.hybrid_retriever import HybridRetriever

# Combine BM25 + BGE with RRF fusion
retriever = HybridRetriever(
    sparse="bm25",
    dense="bge",
    fusion="rrf",
)

documents = retriever.retrieve(documents)

---

Why Hybrid Retrieval?

Retrieval Type	Strengths	Weaknesses
Sparse (BM25)	Exact matches, keywords	Misses semantic meaning
Dense (BGE, DPR)	Semantic understanding	May miss exact keywords
Hybrid	Best of both!	Slightly slower

---

Fusion Strategies

1. Reciprocal Rank Fusion (RRF) - Recommended

from rankify.retrievers.hybrid_retriever import HybridRetriever

retriever = HybridRetriever(
    sparse="bm25",
    dense="bge",
    fusion="rrf",
    rrf_k=60,  # Higher = more weight on top ranks
)

How RRF Works:

RRF_score(doc) = sum(1 / (k + rank_i)) for each retriever

2. Weighted Combination

retriever = HybridRetriever(
    sparse="bm25",
    dense="bge",
    fusion="weighted",
    weights=[0.3, 0.7],  # 30% BM25, 70% BGE
)

3. Interleave

retriever = HybridRetriever(
    sparse="bm25",
    dense="bge",
    fusion="interleave",  # Alternate results
)

---

Multiple Dense Retrievers

Combine BM25 with multiple dense retrievers:

retriever = HybridRetriever(
    sparse="bm25",
    dense=["dpr", "bge", "colbert"],  # Multiple dense
    fusion="rrf",
    weights=[0.2, 0.3, 0.3, 0.2],  # Optional weights
)

---

Complete Example

from rankify.retrievers.hybrid_retriever import HybridRetriever
from rankify.dataset.dataset import Document, Question, Answer

# Create hybrid retriever
retriever = HybridRetriever(
    sparse="bm25",
    dense="bge",
    fusion="rrf",
    n_docs=100,
)

# Prepare query
doc = Document(
    question=Question("What are transformers in NLP?"),
    answers=Answer([]),
    contexts=[],
)

# Retrieve with hybrid fusion
results = retriever.retrieve([doc])

# Get fused results
print(f"Retrieved {len(results[0].contexts)} documents")
for ctx in results[0].contexts[:5]:
    print(f"Score: {ctx.score:.3f} - {ctx.text[:50]}...")

Expected Output:

Retrieved 100 documents
Score: 0.0323 - Transformers are a type of neural network architecture...
Score: 0.0312 - The transformer model was introduced in "Attention Is All You Need"...
Score: 0.0298 - BERT uses the transformer encoder for NLP tasks...
Score: 0.0287 - GPT models are based on the transformer decoder...
Score: 0.0275 - Self-attention is the key mechanism in transformers...

---

Configuration

Parameter	Default	Description
sparse	"bm25"	Sparse retriever method
dense	"bge"	Dense retriever(s)
fusion	"rrf"	Fusion strategy
weights	Equal	Weights for each retriever
n_docs	100	Documents per retriever
rrf_k	60	RRF parameter

---

Best Practices

1. Start with RRF - Works well without tuning
2. Use BGE for dense - Best balance of speed and accuracy
3. Tune weights for your domain - Test with your data
4. Consider ColBERT for precision - Best accuracy but slower