# ANNex vs FAISS

FAISS (Facebook AI Similarity Search) is a C++/Python library for efficient similarity search, widely used in research. ANNex is a Rust-native engine that adds payload filtering, hybrid dense-sparse retrieval, snapshots, and WAL on top of HNSW.

## Feature comparison

| Feature | ANNex | FAISS |
|---|---|---|
| Dense vector search | yes | yes |
| Sparse / BM25 | yes | no |
| Hybrid RRF | yes | no |
| Multivector / ColBERT | yes | no |
| Payload filters | yes | no |
| Embedded (no server) | yes | yes |
| HTTP server | yes | no |
| Rust-native library | yes | no |
| Python bindings | yes | yes |
| Snapshots + WAL | yes | no |

## Recall–latency: ANNex vs FAISS HNSW

NYT256, 290K docs, 256-D angular, Apple M2, interleaved, offset=4000, ANNex 0.1.0, FAISS 1.15.1.

| System | ef | Recall | p50 ms |
|---|---|---|---|
| annex_screen | 32 | 0.864 | 0.252 |
| annex_screen | 128 | 0.918 | 0.580 |
| annex_screen | 512 | 0.961 | 2.205 |
| faiss_hnsw16 | 128 | 0.868 | 0.356 |
| faiss_hnsw16 | 512 | 0.925 | 1.457 |
| faiss_hnsw32 | 128 | 0.904 | 0.627 |
| faiss_hnsw32 | 512 | 0.957 | 2.489 |

FAISS HNSW32 is competitive; ANNex leads at matched recall below ~95%. Full data: https://annexsearch.dev/benchmarks.md

## When ANNex makes sense

- You need payload filtering, hybrid retrieval, or multivector search. FAISS has none of these.
- You want a Rust library with persistence (snapshots, WAL) built in.
- You want Python bindings with NumPy-native batch search and thread control.

## When FAISS makes sense

- You need GPU acceleration for very large-scale search.
- You need IVF or product quantization indexes that FAISS has spent years optimizing.
- Your team already has deep FAISS expertise or an existing FAISS integration.
