Research Chatbot

A scoped, evidence-grounded assistant for navigating my academic work. Try it out yourself, by clicking on the chat icon in the bottom right corner!

The chatbot embedded on this website lets visitors ask questions about my research and get answers drawn directly from my publications, datasets, and research statements - information that is not part of general model knowledge. It is a full-stack RAG implementation that I built from scratch.

How it works

A Chroma vector database holds chunked and embedded versions of all my major academic outputs (publications, dataset documentation, research statement). When a visitor submits a question, the backend retrieves the most relevant chunks from this vector database via semantic search, injects them into a structured system prompt, and calls an OpenAI completion model via API to generate a response grounded in my own writing and information tailored to my research. The system prompt enforces strict scope: the assistant declines questions unrelated to my academic work and is forbidden from inventing publications or results not present in the retrieved context.

Multi-turn retrieval is handled by concatenating prior user messages with new questions before retrieving relevant context, improving relevance in follow-up exchanges without requiring a separate query-rewriting step.

Stack

• Embeddings: OpenAI text-embedding-3-large
• Vector store: Chroma, self-hosted on a Railway persistent volume
• Completion model: gpt-4.1-nano (fast, cheap, sufficient for bounded Q&A)
• Backend: FastAPI, deployed on Railway
• Frontend: Next.js chat widget calling the Railway service via a scoped API URL

Cost and abuse controls

The backend uses a layered defence, including per-IP rate limiting via slowapi, a global in-memory daily cap, and hard input and output character limits.

Retrieval runs against the self-hosted Chroma vector database, so retrieval costs are negligible per query. Spend per question is one embedding call plus one completion, which runs cheap on current models.

What I built and learned

This implementation covers the full RAG pipeline: corpus preparation and chunking of my research output, embedding and vector DB seeding of the resulting chunks, retrieval with conversation history, prompt design for a constrained research assistant, and production deployment with cost controls. The main challenge was keeping the assistant reliably on-topic, with the system prompt going through several iterations before it consistently declined unrelated questions without being unhelpfully terse for legitimate queries.