Expand benchmarks

Reproducible public benchmarks published by Expand, with the harness, the per-task run records and the methodology for each. Each benchmark documents its protocol, its adaptations and the commands that regenerate its tables from the stored records. Repository: https://github.com/expandai/benchmarks.

Benchmark	What it measures	Directory
Search APIs for agents	Web-search APIs (Expand, Exa, Perplexity, Parallel, Firecrawl, DeepSeek, Tavily, Brave) on WebCode RAG, SimpleQA Verified, SealQA, FRAMES and AA-Omniscience	search/

Running a benchmark

Each directory is self-contained: cd into it and follow its README. For the search benchmark:

git clone https://github.com/expandai/benchmarks && cd benchmarks/search
make upstream venv test     # pinned harness clone, venv, offline tests (no keys needed)
make report paired index    # recompute harness reports and comparison tables from stored records

The root Makefile forwards any target to search/ (make test here is make -C search test). The GitHub Actions workflow in .github/workflows/ runs each benchmark's offline tests on every push.

License

MIT for everything written here (LICENSE), which also lists the third-party material each benchmark pins or redistributes.