System · the brain, it knows
Cortex
The knowledge engine.
Cortex turns approved interviews and documents into typed, cited Knowledge Cards, retrieves them with permission awareness, and refuses honestly when there is no grounded source.
Core running todayHow it flows
- Typed Knowledge Cards: topic, criticality, audience, source, confidence.
- Tenant isolation enforced in the database itself (Postgres row-level security).
- Answers carry citations, or return ∅ instead of inventing one.
How it fits
Fed by Pharynx and Larynx, tuned by Thalamus, spoken through Lens, shown by Atlas.
Under the hood
Two pipelines: knowledge in, grounded answers out.
The diagram above is the shape. Here is what the code actually does: one LLM call where it earns its keep, deterministic everywhere else.
Forge · text → typed Knowledge Cards
Six stages. Only extract calls a model to pull structured cards; cleaning, chunking, de-duplication and classification are deterministic rules, and the embedding vector is optional.
Ask · question → grounded answer, or ∅
Dense, keyword and graph search are fused (reciprocal-rank fusion), then a sufficiency gate refuses when the question names something the cards never mention. In strict mode a verify→refine loop checks that every claim is supported; otherwise it refuses rather than guess.
The rest of the constellation
This is one part of a system built to keep one promise: every answer is grounded and permission-aware, or it doesn’t come at all.