Simple RAG tool experiment.
Use docker compose up to start the services, the first time you will also need to run initial_setup.sh to download the Ollama model and create the DB schema.
No UI at the moment, visit http://localhost:8000/docs to store documents, and http://localhost:8090/docs for the QA component.
- Postgres: vanilla Postgres database, used to store the documents for RAG and the task queue
- Qdrant: qdrant server, to perform vector search
- docstore: offers an API to index and retrieve documents, responsible for calculating the embeddings and interfacing with postgres and qdrant
- crawler: ingest documents (sending them to the docstore) from the web or other sources, uses the task queue to handle the volume
- Ollama: vanilla Ollama server, used to form a single textual reply from the retrieved documents. May use llama2 or
- QA: uses the docstore and ollama to retrieve relevant documents and form a single textual answer. Includes a web interface to show the stuff nicely and also the crawler status and the insertion of new documents.
- Implement Hyde
- Try full text search or FTS/vector hybrids
- Improve ways to chunk documents (sentences, paragraphs, etc)
- Add thresholds on retrieval to not fetch more than needed
- Use "tools" (JSON calls) to make the process iterative