DeniableUse case · RAG evaluation
Evaluate retrieval with document-grounded test cases.
RAG systems depend on the relationship between a question, a document, and the answer produced from it. A useful test set gives you cases where that relationship is possible to inspect.
The evaluation job
Keep retrieval and grounding visible.
A RAG test can ask more than whether an answer sounds plausible. It can help you inspect whether the right source was retrieved, whether conflicting material was handled, and whether the final response stayed within the available evidence.
- Relevant and irrelevant source material
- Conflicting or incomplete documents
- Metadata and filter decisions
- Answers that need to stay grounded in the source
Scenario design
The document is part of the test condition.
Human input and creative scenario design shape the conditions around each package. That makes the document a deliberate test input, rather than an example selected only because it looks realistic.
Use your own evaluation method
Fit the cases into the process you already run.
Ingest a pack into the same retrieval workflow you use for internal evaluation, then apply your own questions, labels, or review method. Deniable provides the documents and scenario context. Your team decides how to score its system.
Delivery
Download a pack or retrieve through the API.
A fixed dataset pack gives you a complete ZIP download. Prepaid API credits let an authenticated workflow retrieve documents from the curated catalogue and track usage. They are separate products, with no monthly subscription required.
Internal use
Training and fine-tuning can be part of the job.
The intended internal license permits development, QA, evaluation, commercial internal use, model training, and fine-tuning. Redistribution, resale, public hosting, and sublicensing remain outside that internal license.
Scope
A test set informs decisions. It does not promise a score.
Deniable does not claim that a pack improves retrieval or answer accuracy, or that it represents every knowledge base. The useful question is whether the cases help your own evaluation process inspect the behavior you care about.