BeakrGo to website
Docs/Overview/Knowledge Layer

The knowledge flywheel

Every interaction makes the knowledge base more valuable. Documents come in, agents structure them into cross-linked pages, answers improve, and new pages get captured from the answers themselves — compounding the graph over time.

The knowledge layer is a self-reinforcing system. Three always-on agents -- compiler, capture, health -- keep the knowledge base structured, captured from conversations, and structurally sound without constant human curation. Provenance is tracked at five layers so nothing is ever a black box.

The core loop

1 - INGESTfiles + multimodal2 - CONNECTconnectors3 - STRUCTUREcompiler + [[links]]4 - ANSWERask agent5 - CAPTUREnew pages6 - ENRICHuser acceptsbetter answers / richer knowledge base / tighter graph7 - BETTER ANSWERSdenser graph / more sources / fresher data

Every turn of the wheel produces provenance-linked pages from files, messages, and multimodal evidence. The same conversation that answers your question can also deposit reusable knowledge back into the knowledge base.

Design principles

1
Knowledge is structured, not stored

Raw documents go to S3. The knowledge base is the structured layer on top -- every page is agent-written, linked, and traceable back to its sources.

2
Conversations are inputs, not outputs

The capture agent treats substantive ask-agent responses as candidate knowledge. The knowledge base absorbs what users learn in chat.

3
Provenance is non-negotiable

Five layers from page source down to paragraph-level blame. Citation tokens ({{pmid:12345678}}) give every claim a stable key across messages and pages.

4
Human review for risky content changes

Content-changing suggestions enter a proposal queue for human review. Low-risk structural cleanup, such as link fixes and orphan reparenting, can be written directly with revision history.

5
Red links are features

A [[link]] without a target is not an error -- it's a queue for the health agent. Unresolved references surface gaps in coverage.

6
Extraction is guided, not generic

Extraction profiles at org / project level tell the compiler what to prioritize. Surgical teams extract different entities than regulatory teams.

In this guide