A

Annie, the RAG router

Knowledge retrieval · Layer 3

"Ask me anything. I'll ask the right one."

What Annie does

Annie sits between a question and a model. She works out which collection of documents is relevant, retrieves the best passages, and hands them to a model so the answer is grounded in sources.

Cheapest capable model

Annie picks between local and cloud models on cost, latency and what the question needs.

The right collection

A firmware question goes to the firmware docs, a policy question to the legal files. Collections are config, so you define your own.

Cited answers

Retrieval happens first, so every answer traces back to the passage it came from.

Session-only uploads

Need to ask about one PDF? Upload it, ask, and it's gone when the session ends.

Knowledge that compounds

Add a document to a collection and every future question benefits.

Reach it from anywhere

Use her from Marco, over the HTTP API, or from any tool that speaks MCP.

How it works

Annie doesn't answer questions herself. She decides where to look and who should answer.

  1. A question arrives from Marco, the API or an MCP client.
  2. Annie picks the collection that fits, for example firmware, legal, CAD/CAM or general.
  3. She retrieves the best passages from the vector store.
  4. The carousel sends the question and passages to the cheapest capable model, and the answer comes back with citations.

Where it fits in the four layers

1 Infrastructure & roles Vera, Grant, the bus, the device registry
2 Data pipeline & state Flo, DynaRAG, the carousel
3 Reasoning & inference Rick, Annie, Max, Hank
4 UI & collaboration Marco, Noted, Norm, the graph

Annie is layer 3. Her collections are fed by Norm and the data pipeline, and she hands inference to the carousel.