spruvly
Ingrid Lehto

Internal-docs RAG for a 50-person agency

Ingested 4k internal documents, vector store on pgvector, semantic search + Claude answer generation. Slack bot front-end.

Reduced "where do I find X" Slack messages by 65%.
0

Problem

Agency had 4k internal documents in Drive + Notion + Confluence. New hires drowned. Old hires kept getting pinged "where do I find X?" multiple times a day.

Users

Internal — 50-person agency, all departments

Key flow

  1. 1Anyone asks question in Slack with /ask
  2. 2Bot replies with answer + 2-3 cited document links
  3. 3User opens citation in source app for deeper context

What I built

  • Drive + Notion + Confluence ingestion pipelines with incremental sync
  • pgvector index with semantic + keyword hybrid retrieval
  • Slack bot — answers in-channel with linked sources
  • Per-team scoping so HR docs don't bleed into engineering channels

Limits

  • Best for documents — for live data (CRM, billing) it punts to specialists
  • Re-indexing lags 6 hours on document edits — known trade-off for cost
Internal-docs RAG for a 50-person agency

AI tools used

  • LangChain
  • Supabase
  • Anthropic API
Ingrid Lehto
@ingrid.lehto
View profile