All case studies
06 — Experience · public sector · code not public
A RAG assistant over documents
A proof of concept of an assistant that answers questions from dozens of PDF documents: vector search, cloud and local models, streaming chat.
Problem
Long regulations and guidelines — dozens of PDF files. Staff and applicants searched them for answers by hand.
What I built
- A proof of concept of a RAG assistant in a Turborepo monorepo: a Next.js front end and a NestJS API.
- Vector search in Qdrant, application data in PostgreSQL.
- A choice of cloud models (Claude, OpenAI) and local models through Ollama.
- Chat with streamed answers.
- Containerised deployment.
Outcome
- A working proof of concept presented to stakeholders.
- Experience in choosing models (cloud or local) to match data requirements.
Facing a similar problem?
Describe your process — I'll tell you whether and how it can be solved.