Developing a GraphRAG Research Chatbot with Neo4j

This project started with a simple question: how do we make an LLM answer from a research corpus without hallucinating? The answer was to combine graph structure, semantic retrieval, and strict evidence grounding.

Team Context

This work was built collaboratively. Team members included:

• Peter Mangoro
• Bekithemba Nkomo
• Masheia Dzimba
• Tafadzwa
• Sharman

Project Focus

The chatbot is designed for Q&A over a PDF corpus using a GraphRAG pattern:

• papers and chunks are stored in Neo4j
• chunks are embedded for vector retrieval
• graph relationships support structured filtering (gene, author, topic paths)
• LLM answers are generated from retrieved evidence with citations

What I Built

I implemented and documented an end-to-end workflow across ingestion and serving:

• extraction pipeline (PDF -> structured JSON)
• graph schema and ingestion into Neo4j
• embedding generation and vector indexing
• retrieval routing (semantic, gene-focused, author-focused, aggregates)
• streaming chat outputs with trace/progress support in UI

I also maintained architecture and operations guides to make the system reproducible for others.

Key Findings

• Graph + vector hybrid retrieval is much stronger than either strategy alone for research Q&A.
• Routing matters: classifying query intent before retrieval avoids noisy context.
• Grounded citations significantly improve trust in responses.
• Operational clarity (envs, run order, schema contracts) is as important as model quality for team maintainability.

Lessons Learned

The biggest lesson was that building a useful chatbot is mostly systems engineering, not just prompt writing. Data contracts, ingestion repeatability, retrieval quality, and traceability determine whether answers stay reliable over time.

I also learned the value of documenting architecture in plain language so both technical and non-technical collaborators can contribute effectively.

Skills I Gained

• GraphRAG architecture design
• Neo4j schema design and graph ingestion pipelines
• Retrieval orchestration (route-based + tool-based patterns)
• Streaming response UX with retrieval traceability
• Technical documentation for long-lived project maintenance