GraphAsk — AI Chat Assistant

Plain vector-search RAG can retrieve semantically similar but structurally wrong answers for domain-specific questions. GraphAsk is an AI chat assistant backed by a knowledge graph, grounding domain Q&A in a curated entity/relationship graph instead of relying on free-form retrieval alone — the goal was more explainable, structured answers by having the assistant traverse a modeled graph of entities and relationships before generating a response.

What I Built

• Domain knowledge graph modeling entities and their relationships
• Graph-grounded retrieval — queries traverse the graph before reaching the LLM
• Conversational chat interface with source citations for every answer
• Graceful fallback handling — an honest "I don't know" when the graph lacks coverage, instead of a confident wrong answer

Tech Stack

• AI: LLM (OpenAI/Anthropic), knowledge-graph query layer
• Graph DB: Neo4j (or similar)
• Backend: Python
• Frontend: React chat UI

Key Decisions

• Grounded responses in a curated knowledge graph rather than relying on vector-search RAG alone, since plain similarity search can surface semantically close but structurally wrong answers for domain-specific questions
• Required the assistant to traverse the graph before generating a response and cite sources, making answers explainable and verifiable rather than a confident black box
• Built in a graceful "I don't know" fallback for gaps in graph coverage, preferring an honest non-answer over a plausible-sounding wrong one

Outcome / Impact

Delivered more accurate, explainable answers than plain vector-search RAG for structured domain questions, with citations that let users verify the assistant's reasoning.