faiss-server

faiss-server provides gRPC services to for similarity search using faiss. It is written in C++ and now supports only CPU environments.

Installation

Currently, installation is from source only.

$ git clone https://github.com/ynqa/faiss-server
$ cd faiss-server

To build faiss-server on your local, please use cmake.

$ mkdir build; cd build
$ cmake ..
$ make

[Beta] To build as docker image:

$ docker build -t faiss-server .

Demo

1. Create indexed file, which is composed of the word vectors. Indexed words are on text8 corpus and those vectors are trained by word2vec model of gensim.

$ python example/indexing/indexing.py

2. Select the indexed file and then execute faiss-server.

$ ./build/bin/faiss_server -file_path example/indexing/faiss.index

• If you build with docker:

  $ docker run -p 8080:8080 -v $(pwd)/example/indexing:/tmp faiss-server -file_path /tmp/faiss.index
  

3. Throw the requests for similarity search

$python3 -m pip install --user protobuf==3.20.1
$conda install grpcio

#Generate the protobuf client file faiss_pb2.py:
$build/protobuf/bin/protoc-3.21.1.0 --proto_path=protobuf \
                --python_out=example/client \
                --grpc_out example/client \
                --plugin=protoc-gen-grpc=`which grpc_python_plugin` \
                protobuf/faiss.proto

$ PYTHONPATH=python python example/client/search.py

If meet error of "TypeError: bases must be types" when running search.py, just uninstall the protobuf, and reinstall: $python3 -m pip install --user protobuf==3.20.1

Usage

faiss_server:
    -file_path (index file name) type: string default: ""
    -host (host of faiss server) type: string default: "0.0.0.0"
    -port (port of faiss server) type: string default: "8080"
    -top_k (default number of neighbor) type: uint64 default: 5