Vector search experiments with Qdrant

Lab note: rough, reproducible, work in progress.

I've been measuring how Qdrant behaves under realistic RAG workloads: many small queries with metadata filters, not just raw nearest-neighbour search.

Setup

• Embeddings stored with payload metadata (source, language, date).
• Filtered search to scope by language and recency.
• Measured p50 / p95 latency under concurrent requests.

Early observations

• Payload filters are cheap when indexed; unindexed filters fall off a cliff.
• Recall stays high with HNSW defaults; tuning ef mostly trades latency for recall.
• Co-locating Qdrant with the API container removes a surprising amount of tail latency.

Next

• Test quantization for memory pressure.
• Compare hybrid search (keyword + vector) on the same dataset.

Numbers and a reproducible script will follow once the harness is cleaned up.