Retrieval & RAG
Margin
A citation-only campus knowledge assistant that validates every answer against official university policies, returning 100% verbatim quotes with structured JSON schemas.
03 · Retrieval & RAG

Margin

FastAPIAsynchronous REST endpoints & inference server
OllamaLocal zero-leakage LLM execution engine
sentence-transformersBi-encoder dense semantic vector embeddings
cross-encoderToken-level joint attention reranking
What it does
- Strict citation-only RAG assistant returning 100% verbatim quotes from university policies, eliminating hallucinated answers.
- Structured output schema with automated sentence-level quote validation against indexed document chunks.
- Two-stage retrieval (MiniLM embeddings + cross-encoder reranker) delivering 0.900 Hit@1 on 30 evaluation queries.
- Integrated SHA-256 corpus hash checking and pytest suites to guarantee retrieval stability across policy revisions.
- Containerized FastAPI backend with local Ollama LLM inference, reducing response latency by 35%.
Built with
- FastAPI
- Ollama
- sentence-transformers
- cross-encoder
- pytest
- Docker