Mycelium · an AiGNITE portfolio build
RAG that shows you what it refused to read.
Internal knowledge-base search where access control runs before retrieval scoring, not after generation. Ask as a broad principal and get a cited answer; switch to a restricted principal, ask the same question, and the restricted document stays out of the answer, out of the sources — and the exclusion is counted on screen from the same filter retrieval used.
What it actually does
Warranty labels, computed not opined. Every retrieved passage carries fresh, aging, or stale — date arithmetic against the document's timestamp, attached by deterministic code, never by the model.
Citations that resolve. Every answer sentence cites the passages it drew from; sentences the model could not ground are counted and flagged, not hidden. Click a citation, land on the passage.
Honest misses. A question the corpus cannot answer returns zero passages and says so. Synthesis is never called on an empty retrieval.
Of the tools we reviewed in a July 2026 survey, most enforce access rules after generation, if at all. The self-healing loop — gap tickets from failed queries, owner-approved patches — is the roadmap's next phase, not on this page.
Published limits
On a golden corpus of synthetic documents with planted timestamps and access rules, retrieval leaks nothing across principal boundaries (0 ACL leaks), every citation resolves to text that exists in the cited passage (1.0), every freshness label matches the planted timestamp (1.0), and the retrieved document set is stable across query paraphrases (jaccard 1.0 against a 0.60 bound); the corpus is synthetic, so it does not measure recall on a real knowledge base.
API reference · source · the LLM senses, deterministic code decides, humans approve.