Einstein

The agency’s knowledge base. Not even we used the first version

EPAIn productionFastAPIReactQdrantClaude · Vertex AIMCP

Problem

A knowledge base for the agency, along the lines of Obsidian or NotebookLM.

The idea was to gather what the agency knows into documents that people and AI could query, with citations, and bring it into Claude, where people already work.

How it works

  1. Agency teampeople
  2. DocumentsMarkdown · [[links]] · graph
  3. Search and chatQdrant · Claude on Vertex AI
  4. MCP connectorClaude.ai and Desktop
Deployed with Docker Compose on Compute Engine
  1. 01Every document is Markdown, with [[links]] between notes and a graph to browse them.
  2. 02Semantic search builds the context and the chat answers with citations.
  3. 03An MCP connector brings the same base into Claude.ai and Claude Desktop.

Decisions

DECISION 01

Every answer comes with citations.

The chat links the documents behind each claim, so anyone can check it.

DECISION 02

Einstein also lives inside Claude.

Besides the web app, an MCP connector lets people query the same base from Claude.ai and Claude Desktop without switching tools.

DECISION 03

We dropped the publishing paperwork.

In September 2026 we removed roles and the draft, review and publish flow. Now everything is personal unless you choose to publish it.

BEFOREDraft → in review → published.
AFTEREverything is personal unless you publish it.
If we, the people who built it, weren’t using it, our users probably wouldn’t either.
Key moment · the first version

Result

The first version went unused because we already worked in Obsidian, which does that job very well; I still use it for personal notes. Einstein shipped, with cited chat and an MCP connector, and today it’s one of EPA’s OKRs. In September 2026 roles and the approval flow were removed to simplify the permission model.

MY ROLEProduct owner.I set the architecture, permissions and roadmap; a developer on my team built it.