JustHandled Labs
// AI Agents & LLM Ops

Retrieval Trace Auditor

Audit real RAG evidence traces for missing expected sources, impossible citations, unsupported claims, stale evidence, unused context, and weak routing.

What problem does Retrieval Trace Auditor solve?

A RAG answer can sound plausible while the expected source was never retrieved, a cited or declared support ID never appeared in context, stale evidence displaced its replacement, or a temporal question followed an inadequate route.

Use it to

What it returns

A representative input and result

fixture-backed sample
input {"case_id":"policy-17","query_class":"temporal","route":"vector_only","retrieved_evidence":[{"id":"policy-2024","superseded_by":"policy-2026"}],"expected_evidence_ids":["policy-2026"],"cited_evidence_ids":["citation-ghost"],"claims":[{"claim_id":"claim-1","text":"The 2026 policy replaced the rule.","supporting_evidence_ids":["support-ghost"]}]}
result Release gate: FAIL Critical: citation-ghost was cited but never retrieved. Critical: claim-1 declares support-ghost, which was never retrieved. High: policy-2026 was expected but absent; temporal query used vector_only. Boundary: semantic entailment and answer correctness remain unverified.

Access and approval boundaries

Known limitations

Questions

Does this score answer correctness or faithfulness?

No. It checks the declared evidence path and routing metadata. Semantic correctness and entailment require a separate qualified or disclosed evaluator.

What is a stable evidence ID?

It is an identifier that lets the same source or chunk be recognized across a trace, claim, citation, and regression run without relying on mutable display text.

Can it replace LangSmith or Ragas?

No. It is a portable deterministic audit layer that can complement those systems, especially when you need an offline case record or explicit trace-integrity gate.

Why flag unused retrieved evidence?

Unused context may indicate retrieval noise or wasted context, but it is a Low review item rather than an automatic error.

Does vector-only routing always fail temporal questions?

No. The flag requires review because temporal, causal, relational, and multi-hop questions often need more than similarity retrieval. It is not proof of failure by itself.

Does it upload private traces?

No. The included helper runs locally and requires no browser, network, model, vector database, or observability-platform access.

Retrieval Trace Auditor keeps proof and approval boundaries visible.

The listing includes the tested package, realistic samples, declared permissions, and known limitations.

Get Retrieval Trace Auditor on Agensi