Hybrid retrieval and reranking

Hybrid retrieval combines dense vector similarity with sparse keyword signals in Qdrant. Optional Cohere reranking reorders candidates before the engine injects context into the agent prompt. Tuning weights and top_k / top_n balances recall, latency, and token cost.

Pipeline overview: Vector RAG pipeline. Memory types: Memory types.

When to use hybrid

Hybrid requires retrieval.mode: "hybrid" and ARCFLOW_QDRANT_HYBRID=true at runtime.

Retrieval config

{
 "retrieval": {
 "mode": "hybrid",
 "top_k": 8,
 "dense_weight": 0.65,
 "sparse_weight": 0.35,
 "rerank": {
 "provider": "cohere",
 "model": "rerank-english-v3.0",
 "top_n": 4
 }
 }
}

Weights should reflect your corpus. Start with 0.65 / 0.35 and adjust using trace chunk_count and answer quality on a fixed eval set.

Environment

export ARCFLOW_QDRANT_URL=http://localhost:6333
export ARCFLOW_QDRANT_HYBRID=true
export ARCFLOW_EMBEDDING_PROVIDER=openai
export OPENAI_API_KEY=sk-...
export COHERE_API_KEY=...

Without ARCFLOW_QDRANT_HYBRID=true, the engine may fall back to dense behavior even when config says hybrid. Verify in integration tests.

Dense-only baseline

{
 "retrieval": {
 "mode": "dense",
 "top_k": 6
 }
}

Compare latency and MemoryRetrieved.total_bytes against hybrid on the same queries before committing weights.

Rerank stage

Cohere rerank receives query + chunk texts server-side (not in trace data policy traces). Config:

{
 "rerank": {
 "provider": "cohere",
 "model": "rerank-english-v3.0",
 "top_n": 3
 }
}

Disable rerank during dev to reduce dependencies:

{
 "retrieval": {
 "mode": "hybrid",
 "top_k": 5
 }
}

(Omit rerank object entirely.)

Full memory_config example

{
 "memory_type": "vector",
 "scope": "workflow",
 "namespace": "product-docs",
 "embedding": "openai/text-embedding-3-small",
 "retrieval": {
 "mode": "hybrid",
 "top_k": 8,
 "dense_weight": 0.65,
 "sparse_weight": 0.35,
 "rerank": {
 "provider": "cohere",
 "model": "rerank-english-v3.0",
 "top_n": 4
 }
 },
 "chunking": {
 "chunk_size": 512,
 "chunk_overlap": 64
 }
}

Tuning workflow

1. Ingest representative docs (Knowledge ingestion)
2. Run fixed queries with top_k=12, no rerank; record answers
3. Enable hybrid; sweep weights in 0.1 steps on eval set
4. Add rerank with top_n=4; compare quality vs latency
5. Lock config in registry-published chat workflow

Trace metadata to watch:

{ "kind": "MemoryRetrieved", "chunk_count": 4, "total_bytes": 3100 }

Lower total_bytes after rerank usually means tighter context and faster provider turns.

Static product note

Site knowledge ingests into site kb_namespace. Published chat agents should use the same namespace in memory_config so browser runs retrieve the operator-ingested corpus. See Workflow registry.

Related pages

• Vector RAG pipeline
• Provider configuration
• Trace data policy
• Validation and testing