hnswlib for an in-process nearest-neighbor index

How it works

The library implements a navigable graph index and offers Python bindings. Construction settings influence memory and build time, while query settings control search effort.

What you need

Create the index with the expected dimension and metric. Preserve a mapping from internal IDs to source records outside the index so results can be interpreted after reload.

What to try first

Generate an exact neighbor list for a sample, then compare recall and latency across query settings. Save and reload the index as part of the same test.

Increase query effort across a range of values and plot recall against latency. Choose the operating point based on your own target, since no single setting is best for all data.

Where to look for failures

The surrounding application owns metadata, persistence and concurrent update policy. Pin the dimension and distance metric used for every vector.

Related Franklin pages

Decide whether search belongs in your existing database or in an index inside the process. Compare operational needs such as persistence, filtering and updates as well as nearest-neighbor speed.

• Run exact and approximate vector queries alongside existing relational data. pgvector for embedding search in Postgres
• Add vector retrieval to a SQLite-backed local application. sqlite-vec for embedded vector search

For the role of the index in nearest-neighbor search, see Franklin’s glossary. Vector indexes

Project documentation

• Current project README
• README at the reviewed revision