Research Chat Architecture

How the "Research Chat" Works

A technical deep-dive into the Retrieval-Augmented Generation (RAG) pipeline enabling accurate, cited answers from my PhD research.

Data Flow Pipeline

📄

1. Ingestion

Python parses PDFs & matches BibTeX citations.

🧠

2. Embedding

Google GenAI converts text to 768-dim vectors.

🔍

3. Retrieval

Supabase finds relevant chunks via Cosine Similarity.

💬

4. Generation

Gemini 2.5 Flash answers using the retrieved context.

Why RAG?

Standard LLMs like ChatGPT are trained on general internet data and often hallucinate when asked about specific, private documents. Retrieval-Augmented Generation (RAG) solves this by forcing the AI to "read" my actual thesis before answering.

Technical Stack

🎨 Frontend

Next.js 14 (App Router), Tailwind CSS, Vercel AI SDK (React).

⚙️ Backend

Supabase (PostgreSQL + pgvector), Google Gemini API.

📚 Ingestion

Python 3.10 script with pypdf and bibtexparser running on a local Linux NUC.

The "Context Stuffing" Strategy

When you submit a query, the backend doesn't just ask the AI. It constructs a specialized prompt containing the most relevant 5,000 characters from my work:

SYSTEM: You are an academic assistant. Use ONLY the context below.

CONTEXT BLOCK 1:
[Source: "Cooperative Robots (2016)"]
"The multi-agent system uses a distributed graph..."

CONTEXT BLOCK 2:
[Source: "PhD Thesis (2017)"]
"Latency was reduced by 15% using the new algorithm..."

USER QUESTION: "How was latency reduced?"

This ensures every answer is grounded in fact and includes precise citations.

Try It Out

Ready to explore my research? Ask about cooperative robotics, POMCP, particle filters, or urban search and tracking.

Launch Research Chat →