RAG-Powered Knowledge Base

A 200-person company had documentation scattered across Confluence (engineering), Google Drive (operations), Notion (product), and three internal wikis nobody had updated since 2023. Employees averaged 25 minutes per search session, often giving up and asking a colleague in Slack, which meant the colleague lost 25 minutes too. The cost of context-switching across the company was measured in FTE-equivalents per quarter.

We built a retrieval-augmented generation system on LangChain for the document pipeline, Pinecone for vector storage, and the Claude API for answer generation. Scheduled connectors pull documents from Confluence, Google Drive, and Notion, re-indexing when content changes. The React/Next.js frontend gives employees a conversational search interface: type a question in plain English, get an answer with source citations, click through to the original document if you need more context. Permissions respect source-system ACLs: employees never see docs they wouldn't have access to in the original tool.

• Search-to-answer time fell from 25 minutes to under 4, an 85% reduction
• 78% of employees actively using it within the first month of launch
• Slack #help channel volume dropped 45% as people found answers on their own
• Three wikis decommissioned after the company realized nobody needed them anymore

LangChain orchestrates the ingestion pipeline, and chunking strategies differ per source (Confluence pages chunk differently than Google Docs). Pinecone was chosen over pgvector because the team didn't want to run vector search at scale themselves. Claude generates answers with citations baked into the prompt contract; every answer includes links to source documents. Next.js SSR is overkill for an internal tool but came free from our standard stack and pays dividends in instant load times.