The problem
Employees asked recurring policy and procedure questions, but the approved answers lived across documents. The assistant needed to find relevant source passages and decline when the material did not support an answer.
The system response
I built an intelligent conversational AI assistant using Retrieval-Augmented Generation technology that delivers accurate, policy-grounded responses to employee inquiries automatically.
Key capabilities:
- Semantic search via Supabase vector storage
- Automated Google Drive document ingestion
- Retrieval and response guardrails for unsupported questions
- Modern web chat with webhook API integration
- Context-aware answers from verified documents only
What shipped
The delivered chat retrieves approved document passages before drafting an answer and uses a fallback when relevant context is missing. The ingestion path keeps document updates separate from the employee-facing chat flow.
Turns approved policy documents into cited answers and declines when the supporting evidence is missing.
Architecture decisions
- Ingest approved Google Drive documents into a controlled processing path
- Store searchable chunks and embeddings in Supabase
- Retrieve source context before generating a response
- Return a fallback when the retrieval threshold is not met
Release checks
- Test questions answered directly by an approved source
- Test ambiguous and unsupported questions
- Confirm removed documents no longer appear in retrieval results
