This repo loads a Python codebase from the web, parses, splits, converts then indexes it into a graph database with an additional vector index of embeddings as node properties, then queries the graph database using a generated Cypher statement or with the vector index using semantically similar embeddings to generate an answer.
Tree-sitter - Parses code into a syntax tree using a language's grammar (e.g. Python Binding)
Code-graph - Constructs graph by Modules, Classes and Functions with respective vector embeddings as node properties
FalkorDB - Graph DB (Successor to RedisGraph EOL)
OpenAI - LLM for QA, Cypher and RAG
LangChain - Framework to build apps with LLMs
[Question]
[Initial Prompt]
[Cypher Graph Search]
[Question into Embedding]
[Cypher Vector Search]
[Retrieved Similar Embeddings]
[Search Result]
[Final Prompt]
[Answer]
[Tokens]
[Time]
In the root of this repo, create a .env file with the below keys alongside [your-values]:
NEXT_PUBLIC_MODE=UNLIMITED
FALKORDB_URL=redis://localhost:6379
OPENAI_API_KEY=[your-value]
Run FalkorDB via Docker
docker run -p 6379:6379 -it --rm falkordb/falkordb
Install node packages
npm install
Run the development server:
npm run dev
To interact with the QA web app, open:
Example questions you could ask that use different indexes include:
- For Cypher Graph Search, "Which functions are in the AppGroup Class?"
- For Cypher Vector Search, "Find a few functions which have conditional statements."
If you are rebuilding the Graph DB, first make sure FalkorDB via Docker is empty with
docker exec -it [container_name] redis-cli FLUSHDB
then inspect repo_root for the local folder in route.ts