Reducto

Abraham frames document processing as more than converting PDFs to text. Merged cells, charts, and handwriting encode information that character extraction can miss. In question answering, a bad parse can spoil an answer; in an agent workflow, it can propagate through multiple decisions and modifications. He cites a document-decision benchmark with a frontier-model result of about 30%, a benchmark-specific claim rather than a general PDF accuracy rate.

The talk recommends efficient computer-vision models for layout detection, including models that run on CPUs, and vision-language models for difficult regions such as handwriting. Efficiency and determinism matter at large document volumes. Abraham describes agentic OCR as verification and targeted token correction after an initial parse: fix zero versus O or comma versus period. Regenerating the page can introduce a different error—a model may recompute an incorrectly printed total, producing better arithmetic but a less faithful transcription. This is a speaker-described approach; the official site separately presents its r-1 preview as a unified parsing model that replaces complex multi-tool pipelines.

Simple tables can use Markdown; complex tables may require HTML to retain relationships expressed by merged cells. HTML adds token overhead, and its tags and numbers can be poor retrieval material. Abraham’s example is a user asking how revenue changed over time without naming individual values. His approach creates a natural-language rendering for embedding-based retrieval while retaining the structured HTML table for reasoning. The distinction matters because a model cannot reason over a table that retrieval fails to supply.

Abraham reports that supplying structured parse results alongside original PDFs improved tested models’ answers and reduced reasoning-token use and latency. Some exceeded the cited frontier model’s unassisted result, but his team could not test that model with the same improved inputs after losing access. Classification and splitting address another problem: excessive context can reduce quality as well as increase cost. Paper-mail packets hundreds of pages long illustrate why sorting documents and selecting relevant portions should support, rather than distract from, extraction and decision making.

For line-chart extraction, Abraham describes an agent using a code interpreter, plotting proposed data, inspecting the reconstruction, and repeatedly correcting mistakes to produce a Markdown table. The supplied material includes no chart visuals or numerical error tolerance, so it supports the method without establishing exact recovery. Structured extraction extends verification through a parent agent setting criteria for subagents, with forms containing tens of thousands of fields as an example. He describes a precision–recall tradeoff: frontier models return accurate rows but silently omit content, while document services recover more content with lower precision. He reports improving both through a harness, without supplied scores quantifying the gain.

The ending connects stage-level evaluations and production monitoring to completed agent work: accurate parsing cannot compensate for irrelevant retrieval or unusable final formatting. Abraham describes customers giving agents a file system and CLI to choose document operations, separating readable content from metadata such as citation bounding boxes. Editing is only briefly mentioned; document generation is presented as a future direction. His six closing recommendations are to decompose parsing around accuracy, cost, and latency; verify outputs; format data for each consumer; classify and split; evaluate every stage; and reconsider workflow design as agent capabilities change.