Reducto just launched Deep Extract, and it solves a problem anyone who's worked with document extraction knows all too well: models get lazy. Give a standard extraction model a 500-page financial statement and it'll often stop short, consolidate line items, or skip entries entirely. Deep Extract takes a different approach. It runs an agentic loop that extracts data, verifies results against the source document, identifies what's missing or wrong, and re-extracts until it meets a quality threshold.
The numbers back it up. During its production beta, Deep Extract extracted over 28 million fields across documents up to 2,500 pages long, hitting 99-100% field accuracy. Customers who were seeing 10-20% accuracy with frontier models switched to Deep Extract and got near-perfect results. The system deploys sub-agents to break complex documents into manageable pieces, and users can define correctness criteria directly in system prompts, things like "ensure line items sum to total" or "verify assets equal liabilities plus equity."
The trade-off is speed. Deep Extract takes longer than single-pass extraction because it's doing more work. But compared to having someone manually review a 500-page statement field by field, Reducto says it's still faster, cheaper, and more consistent at scale. The system also generates bounding box citations for every field, which matters for audit trails and compliance. Deep Extract is available now. You enable it by setting deep_extract: true in your extract settings.