2025 · AI / RAG
Rocky: Domain-Specific RAG Assistant
PythonLangChainGPT-4o miniRAGChromaDBOllamaDockerNomic Embed
Role
ML Engineer
Year
2025
Category
AI / RAG
Context
EPFL Rocket Team
01.
Problem
The wiki is large and specific. A generic LLM does not have it, and searching by hand is slow.
Offline indexer, 2000-character chunks with overlap. Query expansion, Nomic embeddings through Ollama, ChromaDB. An LLM reranks hits, then GPT-4o mini answers. Docker, with a daily sync against the wiki.
02.
Approach
03. Outcome
Answers come from the docs. If nothing matches, it says so. Complex queries took about 7 seconds in our tests.
Gallery
01
02
03
04