Customer Knowledge Base

What it does

Most of what you know about a customer lives across shared docs and Slack channels. To make that knowledge more accessible to agents, this tool ingests those sources into a vector index and exposes the index via an MCP server, so any agent can search and cite the source material.

How it works

It runs Slack data exports and docs (as markdown) through an embedding script, which writes vectors into pgvector using Gemini-embedding. The MCP server exposes search over the index. Agent queries are embedded the same way and matched against the documents.

What we've used it for

Surfacing concerns and reproducing them

We asked for a summary of the latest concerns the customer had raised. We chained that bug report into an investigation: the agent pulled logs, reproduced the issue, and opened a PR with acceptance tests that catch it.

Answering questions with citations

Asking "how does customer do X" returned answers with citations to the source Slack threads.

Where we'd take it next

We pulled the Slack export manually for this project. The next step is a bot that reads channels and re-runs embeddings periodically.