Validate citations

Prepare files and profile

Prerequisites: Demo180–190, corpus eligibility and separate chat/embedding profiles. Use the complete distribution built in Release and a working copy; files are shared/Maf.Rag/corpus.json, shared/Maf.Rag and tests/Maf.Rag.Tests. These are complete programs with no unfinished Start. rebuild/probe need the configured embedding service; answer also needs chat. Tests use controlled vectors without a model service.

• Demo180 — Vector Ingestion

• Demo190 — Retrieved Evidence

Build then query the same index

$env:MAF_STATE_ROOT = Join-Path $env:TEMP ('maf-citations-' + [guid]::NewGuid().ToString('N'))
dotnet build BaseTechnique.slnx -m:1 -c Release
dotnet run --project demos/Demo180_VectorIngestion/Demo180.App.csproj -c Release --no-build -- --Rag:Action rebuild --Rag:AsOf 2026-10-23
dotnet run --project demos/Demo190_RetrievedEvidence/Demo190.App.csproj -c Release --no-build -- --Rag:Action probe --Rag:AsOf 2026-10-23 --Rag:Question "Where can I export CSV in Reporting?"

Keep MAF_STATE_ROOT between the two processes. Record ID, passage, score, canonical source and date. Then compare an answer quoting an actual returned passage with an answer containing INVENTED#0 in a copy of the AnswerGate test. Do not change the reference corpus to make a fabricated citation true.

What changes and what stays blocked

Expected observation: the new process reads the index and manifest; an unknown citation or quotation that differs from its passage is rejected. A high cosine does not validate an answer. Eligible source availability does not prove the answer claim is supported. An incompatible manifest requires an explicit rebuild, not lexical fallback.

Progressive hints

First compare diagnostic IDs with those passed to AnswerGate. Then compare quotation text without inventing a passage. If retrieval fails before the gate, inspect the embedding model, dimensions and corpus fingerprint first. The fixed date separates corpus evolution from code evolution.

Accessible, verifiable correction

dotnet test tests/Maf.Rag.Tests/Maf.Rag.Tests.csproj -m:1 -c Release --no-build --no-restore

• Vector index and retrieved evidence

• Exercise: integrated documented action

The correction combines executable counterexamples in Maf.Rag.Tests with Demo470: an eligible citation before approval, a fabricated citation rejected before a request is created. Explain eligibility, ranking, citation and claim validation separately. Each demo page links to the full program.

Scope of evidence

Microsoft.Data.Sqlite stores real vectors; exact cosine is computed in C#. Tests qualify neither the real embedding model nor general answer relevance. No Qdrant or Docker is needed.

Return to mechanisms

• Corpus governance

• Business retrieval and structured sources

• Exercise index