AI GitHub
Qdrant for filtered vector retrieval
Search embeddings alongside structured metadata in a dedicated database.
Key Takeaways
- Qdrant stores points that combine vectors with payloads.
- Choose the vector dimensions and distance metric before loading data.
- Index two overlapping corpora and add a metadata filter.
- Filtering can change recall and latency.
How it works
Qdrant stores points that combine vectors with payloads. Payload indexes help queries narrow candidates by fields such as document type or tenant.
What you need
Choose the vector dimensions and distance metric before loading data. Identify payload fields used in filters so your benchmark reflects the indexes an application will actually need.
What to try first
Index two overlapping corpora and add a metadata filter. Compare the results with a brute-force baseline on queries whose relevant items you already know.
Record recall and latency by filter selectivity. A filter that matches nearly every record can behave differently from one that matches only a few.
Where to look for failures
Filtering can change recall and latency. Use realistic field distributions and keep authorization outside model-generated filters.
Related Franklin pages
Start with a labeled query set and save each engine's candidates. Compare filtered search and ranking behavior before extrapolating throughput from a query without metadata constraints.
- Organize large embedding collections with configurable indexes and search modes. Milvus for managing vector search at scale
- Connect Python applications to turbopuffer's hosted search API. turbopuffer Python client for hosted retrieval
For the role of the index in nearest-neighbor search, see Franklin’s glossary. Vector indexes
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗