DeniableUse case · retrieval and grounding
RAG and LLM evaluation with document-grounded cases.
RAG evaluation needs more than a plausible answer. It needs a repeatable question, source material with meaningful context, and a way to inspect whether the answer stayed grounded.
The testing job
Keep the answer attached to its evidence.
Use documents with definitions, exceptions, references, and competing passages to test retrieval and generation together. A useful evaluation case makes it possible to see why a system selected or missed a source.
- Retrieval across long-form documents
- Citation and evidence checks
- Conflicting or incomplete source context
- Fine-tuning and regression material with a defined scope
Useful checks
Inspect the whole retrieval path.
Review query interpretation, chunk selection, reranking, answer composition, and citation coverage as one workflow. Keep acceptance criteria specific to the application rather than reducing quality to a single score.
Deniable approach
A person defines what the case is testing.
Human input and creative scenario design shape the document relationships and failure modes. Deniable provides prepared conditions for evaluation; it does not claim a universal benchmark or model ranking.