Eight categories of technique — and what we learned from combining them.
The research community has identified dozens of techniques for reducing hallucination in AI-powered data extraction. No single technique is sufficient. The challenge is combining them into a coherent process that works across a wide range of document types — clean text, scanned pages, degraded historical records, visual documents.
We’ve studied all of these and crafted what we believe is the most rigorous combination available. Here’s what the literature describes:
Spatial and structural awareness — tables stay tables, headers stay headers. Location coordinates so every value is tied to a position on the page. Boilerplate removal to keep models focused on content. Visual understanding for documents where layout carries meaning.
Extraction-only instructions that forbid inference. Explicit “not found” directives — the model must say so rather than guess. Negative examples that teach what absence looks like. Decomposition into focused subtasks. Verbatim citation requirements.
Every extraction constrained to a defined schema. Type validation, enum restrictions, regex patterns. Explicit abstention states — null, ambiguous, not found — so the model has a valid way to express uncertainty rather than filling blanks.
Every extracted value linked to a source quote from the document. Evidence-first extraction: find the supporting text, then derive the value. If there’s no quote, there’s no value.
Document segmentation that respects structure. Overlapping windows so nothing falls between boundaries. Hybrid retrieval combining keyword and semantic search. Context budgeting so models see what matters.
Evidence-bound chain-of-thought — structured reasoning, not free-form generation. Chain-of-verification: extract, then independently verify against the source. Field-level cross-checks for internal consistency.
Multiple independent AI readers extract from every document. An arbiter compares their extractions and identifies where they diverge. Agreement is earned, not assumed. Where readers disagree, that’s information.
Deterministic validators catch impossible values. Confidence thresholds route uncertain extractions to inspection. The human remains in the loop not as a fallback, but as the authority.
Results in this field are deeply context-dependent. A technique that helps on clean English text can hurt on scanned historical records. One that worked with smaller models can be dead weight in larger ones. Move from text to vision and half the assumptions shift. Benchmarks optimize for what they measure — and that’s rarely your documents. Part of building the mint has been figuring out which techniques generalize and which need to be parameterized per task.
The craft of the mint mirrors the craft of the codebook. You refine your codebook by testing it against your documents, finding where it breaks, and making it sharper. We refine the mint the same way — testing combinations of techniques against a wide range of document types, finding what actually improves extraction and what doesn’t, and making the process more rigorous over time. The two are symbiotic: a better mint surfaces ambiguities more clearly, which helps you refine your codebook more effectively, which in turn teaches us what the mint needs to handle better.
← Back to How It Works