How does Milvus integrate with and manage its Faiss dependency? #53274

So far as I know, Milvus does not use facebookresearch/faiss directly — it uses Faiss through a vendored fork inside Knowhere.
Dependency chain:

Milvus (internal/core)
       └─ zilliztech/knowhere   (pinned to a fixed commit, fetched at build time)
            └─ thirdparty/faiss (full faiss source vendored in-tree, Zilliz's patched fork)
            └─ thirdparty/hnswlib, thirdparty/DiskANN

Possible approach to benchmark your Faiss within Milvus:

1. Fork zilliztech/knowhere and replace thirdparty/faiss with your modified faiss (keep the APIs Knowhere relies on, e.g. HNSW in knowhere/src/index/hnsw/faiss_hnsw.cc, and index serialization compatible).
2. Point Milvus at your fork by changing KNOWHERE_VERSION (and GIT_REPOSITORY if you forked) in internal/core/thirdparty/knowhere/CMakeLists.txt.
3. Build milvus and benchmark.

Build milvus from source code:

1. clone from github repository

git clone https://github.com/milvus-io/milvus.git

2. install third-party dependencies.

cd milvus/
./scripts/install_deps.sh

3. compile Milvus.

make

Benchmark tool: https://github.com/zilliztech/VectorDBBench

@ThatikondV Hey. Yes, Milvus relies on Faiss with some minor exceptions (check https://github.com/zilliztech/knowhere/tree/main/thirdparty/faiss/faiss/cppcontrib/knowhere). We periodically update the clone of the Faiss OSS in our codebase

@ThatikondV

Is it possible to build/benchmark Milvus against a newer Faiss version or a specific Faiss commit for evaluation purposes?

Technically, it is possible. You need to replace thirdparty/faiss directory in the knowhere repo with the required Faiss version. There are several exceptions: 1) thirdparty/faiss/faiss/MetricType.h file, which is a bit different from the baseline. This is the only file. 2) thirdparty/faiss/faiss/cppcontrib/knowhere directory contains knowhere-specific changes that are not present in the baseline Faiss; , 3) cmake is managed via cmake/lib/libfaiss.cmake. Usually, any AI assistant tool can handle such a modification.

Hi @alexanderguzhva, @yhmo,

Thank you both for the detailed explanations. This is very helpful in understanding how Milvus integrates Knowhere and how Faiss is maintained within Knowhere. I now have a much clearer idea of how to set up the custom Knowhere/Faiss build for benchmarking. Really appreciate your help!