DEMO CORPUSAGENT GOVERNANCEREV: 0.1

Agent Governance Demo

A synthetic-but-realistic corpus for showing how BaseMouse grounds AI agents in auditable context: OpenTelemetry evidence span, EU AI Act logging, policy controls, cost investigations, and incident response.

Use the demo queries View corpus JSON

retrieval quality baselinenpm run eval:retrieval:demo

35documents

20/20golden queries

100%hybrid recall

0.900hybrid MRR

20/20 golden queries pass with 100% search and context-pack recall.

PASS retrieval eval: 20/20 cases passed
mode=hybrid minRecall=100%
searchRecall=100% contextPackRecall=100%
searchMRR=0.900 contextPackMRR=0.900

demo flowwhat this proves

Corpus

Thirty-five docs model agent registry, policy controls, evidence bundles, redaction, cost attribution, and MeshAI control-plane integration.

35docs

Golden queries

Twenty realistic operator/auditor questions validate whether the expected grounding document appears in search and context packs.

20queries

Evidence

Context packs preserve document ids, citations, versions, checksums, retrieval mode, and provenance without stuffing raw bodies into traces.

OTLPready

SAMPLE QUERIESAGENT / AUDITOR LANGUAGE

Ask governance questions, get grounded context.

These are the public demo prompts that map to the golden-query suite.

what fields should an OpenTelemetry evidence span include for context retrieval
how do we prove what context grounded an agent answer during audit
EU AI Act logging controls for high risk agents
investigate agent cost spike from retrieval volume model calls and tool fanout

context-pack output shapecited + checksummed

node scripts/evaluate-retrieval.mjs \
  --corpus data/demo-corpus/agent-governance.json \
  --cases data/retrieval-eval/agent-governance-demo.json \
  --retrieval hybrid

{
  "schema": "basemouse.context_pack.v1",
  "retrieval": "hybrid",
  "entries": [
    {
      "id": "otel-evidence-span",
      "title": "OpenTelemetry Evidence Span",
      "provenance": { "checksum": "…", "version": 1 }
    }
  ]
}

WHY IT MATTERSMEASURABLE BEFORE EMBEDDINGS

Retrieval quality is now a product story, not a vibe.

The demo gives BaseMouse a repeatable proof loop before learned semantic embeddings or full GraphRAG: corpus → golden query → expected source → measured recall/MRR → regression gate. Real partner corpora can replace this synthetic set without changing the harness.

Read the demo docs Talk to us