Knowledge Core
Knowledge Core is a retrieval engine where role law governs what gets retrieved, ranked, and cited. It is the knowledge subsystem of Role OS.
The Problem
Section titled “The Problem”Generic RAG treats all queries the same. Every role gets the same chunks, the same ranking, the same context window stuffing. This produces role blur — a security reviewer and a product strategist sound identical because they received identical evidence.
The Solution
Section titled “The Solution”Knowledge Core introduces role overlays — declarative per-role configurations that control:
- What sources to prefer or forbid
- How to weight lexical vs. semantic vs. metadata scores
- What vocabulary to boost or suppress
- How to handle weak, stale, or conflicting evidence
The result: same corpus, same task, materially different evidence per role.
Key Concepts
Section titled “Key Concepts”| Concept | Description |
|---|---|
| Corpus | A single document store with metadata, embeddings, and versioning |
| Overlay | A JSON config that defines how one role retrieves and ranks evidence |
| Bundle | The governed output of a retrieval operation — selected, rejected, provenance, warnings |
| Fallback | Explicit governance for degraded scenarios (weak, stale, conflicted, none) |
Proven Claims
Section titled “Proven Claims”- 5 pilot roles retrieve measurably different evidence from the same corpus on the same task
- Every posture (strong/weak/stale/conflicted/none) produces distinct prompt and artifact behavior
- Forbidden sources are enforced, drift is detected, posture is visible to operators
- The full chain holds: overlay → bundle → prompt law → artifact behavior