Bio Chat

An LLM using retrieval augmented generation over a biomedical knowledge graph to answer questions. Served as a website.

Set Neo4j password:

export NEO4J_PASSWORD=...

Run the app:

docker compose up -d

Enter docker container:

docker exec -it bio-chat-devcontainer-1 bash

Setup env without docker:

conda create -n bio_chat python=3.11 -y

conda activate bio_chat

pip install -r requirements.txt

Also download Ollama

Process text and upload to neo4j:

python ingest.py

Run a prompt:

python starter.py

TODO: Citations

• ingest should add a new edge property
• retrieve should return edge property

TODO: Improve prompt?

TODO: Create NER evaluation dataset

TODO: Mess with CLIP/soft prompting?

TODO: Decide on predicates

• chemical
• source
• disposition
• exposure_route
• food
• health_effect
• organoleptic_effect
• process
• role

Note to self: plan:

• VectorIndex (VectorIndexRetriever) -> interfaces with embeddings in neo4j
• GraphIndex (KGRAGRetriever) -> interfaces with triplets and subgraph in neo4j
• GRetriever -> combines the two
• will use both neo4jstores but have them point to the same db