Benchmarks run with PageIndex, with what is needed to recompute their results: questions, code, every answer and model call, bills and grades. Each folder is one benchmark.
Each benchmark has a README.md with its results and how to reproduce them, a METHOD.md, the data in data/ and
the generated tables and charts in results/.
The new PageIndex Agent against legacy PageIndex, with the same model on Financial QA's 100 questions: GPT-6.1 Terra (max) answers 99.0%, up from 96.7%, at 28% less cost.
Results, methodology and reproduction instructions →
Accuracy vs. cost of 41 configurations of 25 LLMs answering 100 questions over 40 Form 10-K filings from 8 companies, with PageIndex doing the retrieval.
Accuracy
Accuracy vs. cost


