Project Overview
We built an internal AI knowledge assistant that lets employees ask questions in plain language and get accurate, cited answers drawn from scattered wikis, docs, and tickets — ending the daily hunt for information.
The Challenge
Institutional knowledge was fragmented across wikis, drives, and chat history. Employees wasted hours searching, and the same questions were answered over and over.
- Knowledge scattered across many disconnected systems
- Keyword search returned noise, not answers
- Employees lost hours hunting for information
- Experts interrupted by repeat questions
Our Strategic Approach
We connected the assistant to all content sources, indexed them with semantic embeddings, and used retrieval-augmented generation to answer questions with citations and permission awareness.
The Solution We Delivered
Employees get instant, source-cited answers in a chat interface that respects access permissions and improves as content grows.
- Natural-language Q&A over all internal content
- Source citations on every answer
- Permission-aware retrieval and access control
- Connectors to wikis, drives, and ticketing
- Automatic re-indexing as content changes
- Usage analytics revealing knowledge gaps
Technologies Used
- LLM — Answer generation with grounding
- pgvector — Semantic search over content
- LangChain — Retrieval and citation pipeline
- Next.js — Chat interface and admin tools
- Node.js — Connectors and indexing service
- OAuth / SSO — Permission-aware access
Development Process
- Source inventory — Catalogued knowledge sources and access models.
- Indexing pipeline — Built connectors and semantic indexing with permissions.
- Grounded answering — Implemented RAG with mandatory citations.
- Access control — Enforced per-user permission-aware retrieval.
- Rollout & analytics — Launched company-wide with gap analytics.
Results & Impact
Employees found trustworthy answers in seconds, reclaiming time and easing the load on experts.
- Time spent searching for information cut by ~50%
- Repeat questions to experts dropped sharply
- Answers delivered with verifiable citations
- Surfaced gaps where documentation was missing
🎯 Key Takeaway
A grounded, permission-aware knowledge assistant turned fragmented institutional knowledge into an instant, trustworthy resource for the whole company.
Project Gallery
