TL;DR
- Before jumping to AI, ask: is this a structure problem or a familiarity problem?
- When creators ≠ searchers (marketing creates, sales searches), semantic search helps immediately
- Lindy gives full prompt visibility — you can see why it works and tweak it when it doesn't
- The agent builder gets you 85% there; you course correct and refine from there
- Use smaller models for retrieval — it's not reasoning, it's matching
What I've Learned
The first question I always ask: is this actually an AI problem?
At 600 people with 20 years of collateral, you might assume yes. But the real diagnostic is simpler: are the people searching the same people who created the content? If yes, better folders solve it. If no — and at Murali's scale, marketing creates while sales searches — semantic search creates immediate leverage.
I used Lindy for this build because of prompt visibility. Many knowledge-retrieval platforms are black boxes. Lindy shows you the actual prompts the agent uses. When results aren't what you expect, you can go in and tweak — I changed the default from a one-sentence explanation to a paragraph, and immediately saw better results.
The agent builder is useful too. You describe what you want in natural language, and it builds the workflow. It's not perfect — in my tests it got about 85% there — but you can course correct. That's faster than building from scratch.
The setup took about 5 minutes: connect Google Drive, point to your docs folder, wire it to Slack. The demo query ("fintech company interested in API integrations") returned three relevant case studies with deep links. From there, you could extend it — calendar triggers that auto-fetch docs before meetings, enrichment steps that research who's attending.
One thing I discovered while researching: real-time knowledge base access is emerging. Tools like Fireflies now have live assistants that query your knowledge base during calls — not just prep beforehand, but support in the moment.
Tools & Resources
- Lindy AI — automation platform with agent builder and full prompt visibility
- Clone this workflow — the exact Lindy template from this episode
- Fireflies — AI note-taker with knowledge base and live assistant features
- Agent builder pattern — describe what you want, get 85% there, course correct
- Model selection heuristic — retrieval doesn't need expensive reasoning models
- Familiarity diagnostic — "Did the searcher create the content?" determines your approach