Blog Details

Document automation almost always starts with the same question: do we extract text with OCR and apply rules, or do we hand the document to a large language model and let it reason about the content directly? We've shipped both approaches in production, and the honest answer is that neither one is universally better — they solve different shapes of problem.

OCR Plus Rules: Predictable and Auditable

In ClinOps's Expense module, receipts get read by an in-house PaddleOCR microservice tuned for Turkish and English text, and the extracted fields feed straight into duplicate-detection and categorization logic. Arkem, the predecessor to TradeCheck, took the same approach for SWIFT MT700 messages: OCR extraction feeding a deterministic 14-rule UCP 600 engine. The advantage of this pipeline is predictability — the same input reliably produces the same output, every decision traces back to an explicit rule, and there's no model non-determinism to explain to an auditor. The cost is rigidity: a document layout the rules didn't anticipate, or a field the OCR model misreads, needs a code change or a manual fallback.

LLM Vision: Flexible, but You Have to Verify It

TradeCheck's newer pipeline uses cloud LLM vision to read commercial documents directly and cross-reference them against Letter of Credit terms, with a hybrid flow — vision first, regex second, LLM structured output third — rather than trusting a single pass. LLM vision handles messy, inconsistent document layouts far better than rule-based OCR ever could, and it can generalize to document types nobody explicitly coded for. The trade-off is that outputs need structured validation, not blind trust: we constrain the model to structured output formats and cross-check its extractions against the deterministic parts of the pipeline (like MT700 parsing) wherever a hard rule exists to check against.

Picking a Pipeline

Our rule of thumb: if the document format is stable and the rules are genuinely fixed (like UCP 600's 14 core checks, or a fixed receipt category list), OCR-plus-rules gives you an auditable, cheap, fast pipeline. If document layouts vary widely, or the judgment involved is closer to "does this description reasonably match that line item" than a fixed rule, LLM vision — kept honest with structured output and cross-checks against whatever deterministic rules do exist — is the better fit. Most real systems, including TradeCheck's, end up as a hybrid: deterministic where you can be, model-based where you have to be, and validated at every handoff between the two.