README.md

STAPI: Sentence Transformers API

OpenAI compatible embedding API that uses Sentence Transformer for embeddings

Container Image: ghcr.io/substratusai/stapi

Support the project by adding a star! ❤️
Join us on Discord:

Install

There are 2 options to install STAPI: Docker or local Python install.

Install (Docker)

Run the API locally using Docker:

docker run -p 8080:8080 -d ghcr.io/substratusai/stapi

Install (Local python)

Install and run the API server locally using Python. Only supports python 3.9, 3.10 and 3.11.

Clone the repo:

git clone https://github.com/substratusai/stapi
cd stapi

Install dependencies:

pip3 install -r requirements.txt

Run the webserver:

uvicorn main:app --port 8080 --reload

Usage

After you've installed STAPI, you can visit the API docs on http://localhost:8080/docs

You can also use CURL to get embeddings:

curl http://localhost:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Your text string goes here",
    "model": "all-MiniLM-L6-v2"
  }'

Even the OpenAI Python client can be used to get embeddings:

import openai
openai.api_base = "http://localhost:8080/v1"
openai.api_key = "this isn't used but openai client requires it"
model = "all-MiniLM-L6-v2"
embedding = openai.Embedding.create(input="Some text", model=model)["data"][0]["embedding"]
print(embedding)

Supported Models

Any model that's supported by Sentence Transformers should also work as-is with STAPI. Here is a list of pre-trained models available with Sentence Transformers.

By default the all-MiniLM-L6-v2 model is used and preloaded on startup. You can preload any supported model by setting the MODEL environment variable.

For example, if you want to preload the multi-qa-MiniLM-L6-cos-v1, you could tweak the docker run command like this:

docker run -e MODEL=multi-qa-MiniLM-L6-cos-v1  -p 8080:8080 -d \
  ghcr.io/substratusai/sentence-transformers-api

Note that STAPI will only serve the model that it is preloaded with. You should create another instance of STAPI to serve another model. The model parameter as part of the request body is simply ignored.

Integrations

It's easy to utilize the embedding server with various other tools because the API is compatible with the OpenAI Embedding API.

Weaviate

You can use the Weaviate text2vec-openai module and use the STAPI OpenAI compatible endpoint.

In your Weaviate Schema use the following module config, assuming STAPI endpoint is available at http://stapi:8080:

  "vectorizer": "text2vec-openai",
  "moduleConfig": {
    "text2vec-openai": {
      "model": "davinci",
      "baseURL": "http://stapi:8080"
    }
  }

For the OpenAI API key you can use any key, it won't be checked.

Read the STAPI Weaviate Guide for more details.

Creators

Feel free to contact any of us:

• Sam Stoelinga aka Samos123
• Nick Stogner