A Translator ARA that aims to be modular in its query expansion and ranking, among other TRAPI workflow operations.
Shepherd's tech stack consists of a FastAPI server that handles all web traffic with multiple workers, async python so it is never stuck waiting on a web request, and PostgreSQL that acts as the central state management.
- Shepherd receives a TRAPI query from the ARS.
- It contacts all of its query expansion modules and creates all of the queries.
- It sends all of the individual queries to the Retriever.
- Once it receives those queries back, it does any message merging across the different expanded queries.
- It then sends the merged message on to the other specified workflow operations, i.e. different scoring modules.
You can run everything through Docker, just uncomment the shepherd section in the compose.yml file.
Run the postgres through Docker:
- docker compose up If you want to run Shepherd through local python: It is recommended that you use a python virtual environment
- python -m venv <path_to_venv>
- source <path_to_venv>/bin/activate Install dependencies
- pip install -r requirements.txt Install Shepherd
- pip install -e . Start
- python main.py