Document processing

For pipelines that turn high volumes of documents into structured data.

Cost per document sets the economics

These workloads are high volume and narrow: the same extraction, classification or flagging task repeated across millions of pages.

Long inputs multiply per-token pricing, so the bill scales with page count even when the task itself is simple.

Frontier models are expensive, and cannot handle your throughput

Caching barely helps when every document is different, and prompt trimming can't shrink the document itself. Batch discounts lower the price without changing the unit economics.

Rate limits then cap how fast you can drain a backlog. A million-page queue moves at whatever throughput your provider allocates you, and no amount of tuning on your side changes that.

Custom endpoints scale infinitely, without compromise on quality

Throughput becomes a function of the hardware you allocate rather than a quota someone else sets. A model a fraction of the size multiplies documents per GPU hour, so you scale by adding capacity instead of asking for it.

Narrow, repetitive tasks are where specialized small models do best. We train on your document distribution and measure against your evaluation set, so accuracy means your schema rather than a benchmark.

Where it applies

Field extraction to schema. Invoice numbers, dates, totals and line items, returned in the shape your database expects.
Document classification and routing. Sort incoming documents by type so each one reaches the right pipeline.
Clause and risk flagging. Surface the paragraphs a reviewer needs to read, in contracts and policy documents.
Table extraction. Recover rows and columns from documents where the layout carries the meaning.
Redaction and PII removal. Strip personal data before documents move to a system with a wider audience.
Long-document summarization. Where input length, not task difficulty, is what makes a frontier model expensive.

Find out what your workload should cost

We identify where you are overspending, evaluate the optimizations available, and show you the lowest-cost configuration that meets your quality bar. It starts with an export or one day of traffic.