Skip to Content
TemplatesHybrid

hybrid template

BM25 + vector retrieval with contextual enrichment — for teams hitting recall ceilings with vector-only search.

What you get

project/
├── src/
│   ├── pipeline.py       # Hybrid ingest (enrich=True) and query (alpha=0.6)
│   └── config.py         # Pipeline configuration
├── eval/
│   ├── golden_set.json   # Evaluation dataset
│   └── config.yaml       # Evaluation thresholds and CI gate settings
├── pyproject.toml        # Python dependencies including rag-forge-core[cohere]
└── README.md

Default configuration

The hybrid template blends dense vector scores with BM25 sparse scores using a weighted alpha parameter. Contextual enrichment (enrich = true) annotates each chunk with surrounding context before indexing. The Cohere reranker is available via the cohere extra but is set to none by default — enable it when you need a second-pass ranking pass.

# pyproject.toml [tool.rag-forge] section
[tool.rag-forge]
template = "hybrid"
chunk_strategy = "recursive"
chunk_size = 512
overlap_ratio = 0.1
vector_db = "qdrant"
embedding_model = "BAAI/bge-m3"
retrieval_strategy = "hybrid"
retrieval_alpha = 0.6
reranker = "none"
enrich = true

alpha = 0.6 weights results 60% toward dense and 40% toward sparse. Tune this value based on your domain: lower alpha for keyword-heavy corpora, higher for semantic queries.

To enable Cohere reranking at query time:

rag-forge query "your question" --strategy hybrid --alpha 0.6 --reranker cohere
  1. Run rag-forge index --source ./docs --enrich --sparse-index-path .rag-forge/sparse to build both the dense and sparse indexes.
  2. Experiment with --alpha between 0.3 and 0.8 to find the recall/precision sweet spot for your corpus.
  3. Set COHERE_API_KEY in your environment and test --reranker cohere to see whether a second-pass rerank improves answer quality on your golden set.

When to upgrade

Move to the agentic template when users ask multi-hop questions that require decomposing a single query into sub-questions and synthesising answers across multiple retrieval passes.

BasicAgentic