// the find
docker/genai-stack
Langchain + Docker + Neo4j + Ollama
A Docker Compose stack that wires together Ollama (local LLM), Neo4j (graph + vector store), and LangChain into five runnable demo applications. It's a learning scaffold for RAG patterns, not a production foundation. The target is developers who want to see vector search and knowledge graphs working together without setting up each piece by hand.
The multi-platform Ollama handling is well thought out — Linux gets a containerized instance with GPU profile, Mac/Windows use host Ollama, and the .env config makes switching straightforward. Using Neo4j for both vector embeddings and the knowledge graph is genuinely interesting; you get graph traversal and semantic search from one store rather than bolting on a separate vector DB. The five apps cover meaningfully different patterns (streaming API, PDF ingestion, KG-backed ticket generation) so there's something concrete to dissect in each. Docker Compose watch mode for hot-reload is a nice touch for iterative development.
The LLM config still references 'claudev2' via AWS Bedrock — a model that's been deprecated for years — which signals the repo is not actively maintained at the integration layer despite recent pushes. There are no tests anywhere; the whole stack is run-it-and-see. The LangChain dependency is pinned loosely, and LangChain's API has broken backwards compatibility repeatedly, so expect version-pinning pain when you try to upgrade. The apps share a flat `requirements.txt` and a tangle of top-level `.py` files rather than any package structure, which means you can't reuse the chains logic in your own project without copy-pasting.