Recoleta
GitHub
Source note

Show HN: Vitrus – the company brain that tells you what it doesn't know

Authors
ahvural
Source
Hacker News
Published
Collection
Software Intelligence
Agent MemoryCode IntelligenceKnowledge GraphMCPAPI Verification
Open GitHubMarkdown source

Summary

Vitrus is a Markdown-based company memory for humans and agents that answers questions with sources, confidence, and a deterministic gap list. It targets agent workflows where unsupported claims, stale facts, and unauthorized retrieval can cause bad actions.

Problem

  • Teams and agents often search across chat, docs, tickets, and APIs, then receive raw results without a clear source trail.
  • Company knowledge changes over time, so agents need to detect missing docs, stale decisions, contradictions, single-person ownership risk, and uncited claims before acting.
  • API-using agents can invent endpoints or wrong arguments unless calls are checked against the actual OpenAPI spec.

Approach

  • Markdown files and optional .edges.json sidecars are the canonical record; PGLite or Postgres plus pgvector builds a disposable index from that record.
  • Retrieval combines vector search, BM25, and entity match, then uses reciprocal rank fusion and graph-based re-scoring for connected or corroborated hits.
  • Gap detection is deterministic and LLM-free: it checks graph structure and explicit signals for missing nodes, contradictions, stale facts, single-point risk, and uncited events.
  • The think and verify commands return sourced answers, verdicts such as grounded/stale/contradicted/unsupported, confidence, freshness, and the exact undocumented gaps.
  • Agents connect through MCP tools and API commands; OpenAPI import/search/verify/call checks endpoint names, missing args, wrong types, unknown args, deprecated endpoints, and permissions before execution.

Results

  • The repo reports source-hit ≥90% on its eval gate.
  • Gap-Eval reports 100% gap recall and 100% gap precision on a controlled synthetic corpus; the authors state this does not prove real-world generalization.
  • ACL leak testing reports 0 unauthorized results, with fail-closed index-layer enforcement.
  • The project reports 200+ tests and four CI gates: typecheck, test, eval, and leak-test.
  • The MCP surface grew from 13 to 30 tools, including search, think, verify, gap reports, API search/verify/call, onboarding paths, quizzes, schema checks, and briefings.
  • Integration coverage includes 7 first-class live connectors, 8 one-token REST presets, 5 pagination styles, and hosted cloud support for 13 managed sources.

Link