PROJECT
Local RAG for government-document corruption detection
A fully local retrieval-augmented pipeline for surfacing signals in public documents.
A retrieval-augmented-generation pipeline that runs entirely on a local server — Ollama for the model, ChromaDB for the vector store, and LangChain to wire it together — aimed at surfacing corruption signals in government documents.
Problem
Public-document corpora are large and dull to read. The interesting patterns are buried. Sending sensitive documents to a hosted API is also the wrong move — this needed to stay local.
Approach
- Local model via Ollama, so nothing leaves the machine.
- ChromaDB as the vector store for the document chunks.
- LangChain to orchestrate retrieval and querying.
TODO(JP): describe the document set, how you chunked/embedded it, and what a “corruption signal” looked like in practice.
What broke / what I learned
TODO(JP): retrieval quality, chunking, hallucination control — add the real lessons.
Outcome
TODO(JP): what shipped and what it could do by the end.