This paper presents griddy-forge, a document-ingestion engine built for AI spreadsheet workflows. It was developed to replace a slower and noisier document-conversion path used in Griddy. Across product-relevant internal lanes and reproducible public benchmark lanes, griddy-forge improves latency and extraction quality relative to markitdown-core. On the Griddy workload lane, it is 18.2x faster (9.20 ms vs. 167.58 ms), improves quality from 0.853 to 1.000, and reduces benchmark output tokens by 30.1%. On the Griddy-derived document lane, it is 3.0x faster (237.14 ms vs. 710.42 ms) with quality improving from 0.913 to 0.988. On common non-PDF lanes, griddy-forge is 18.6x faster on public office files, 34.2x faster on synthetic DOCX, and 19.8x faster on synthetic spreadsheets while maintaining or improving quality. On public PDF evaluation, it is 3.56x faster than markitdown-core on the core PDF industry lane (48.80 ms vs. 173.55 ms) while improving quality from 0.782 to 0.973. On the completed 25-document standard public PDF lane, it remains faster (3.70 s vs. 4.02 s) and substantially more accurate (0.946 vs. 0.672). On the 12-document hard public PDF lane, it improves both latency (2.74 s vs. 3.43 s) and quality (0.921 vs. 0.672). The main limitation appears on the frozen 10-document public holdout lane, where griddy-forge improves quality (0.669 vs. 0.547) but is slightly slower (2.48 s vs. 2.28 s). The upload includes the manuscript PDF, LaTeX source bundle, and public benchmark-summary artifacts used for the paper-facing PDF evaluation. Additional product context is available at https://getgriddy.ai
Justin Freels (Thu,) studied this question.