What's this about?
LLMs can increasingly answer complex questions over enterprise-scale document collections, but evaluating that is hard: companies won't share internal communications, and synthetic datasets have been unconvincing. CorporateBench is a large-scale Q&A benchmark built on temporal knowledge bases.
Why does "temporal" matter for enterprise Q&A?
Because in a company, facts expire. Who owns this service, what the policy is, what the price was — every answer has an implicit "as of when". A benchmark that ignores time rewards a model for finding *a* document that says something, when the correct behaviour is finding the document that is currently authoritative and noticing that three others are superseded. That is most of the real difficulty of enterprise search, and static benchmarks miss it entirely.
Why does this matter to you?
- Temporal correctness is the failure mode most enterprise RAG deployments actually hit, and it is rarely measured.
- The confidentiality problem it works around is why enterprise agent evaluation is so much weaker than consumer.
- The same "as of when" question applies to security findings — a vulnerability report without a timestamp and a version is not actionable.