LLM Accuracy Testing Framework

Problem statement

An AI document-processing product extracted structured data from trade finance documents using LLMs. Each model or prompt change needed validation against thousands of documents. Humans were doing it field by field in spreadsheets.

Manual validation was slow, inconsistent between reviewers, and impossible to run on every release. Model regressions reached customers before anyone noticed.

Framework design

A Python framework with three layers: a golden dataset of documents with verified ground-truth values, an execution layer that runs documents through the model pipeline, and a comparison engine that scores extracted output against ground truth.

Comparison logic is field-type aware. Dates normalize before matching, amounts compare numerically with currency awareness, and free-text fields use similarity scoring with configurable thresholds instead of brittle exact-match.

Validation methodology and accuracy scoring

Every run produces accuracy at three levels: per field, per document type, and per model version. That granularity matters: a model can hold 95% overall accuracy while silently degrading on one critical field like LC expiry date.

Version-over-version comparison flags any field whose accuracy drops beyond a tolerance, turning model regression detection from a judgment call into a diff.

Hallucination detection

Extraction hallucinations are values the model returns that exist nowhere in the source document. The framework cross-checks extracted values against document text; unmatched values get flagged for review rather than silently passing because they 'look plausible'.

Reporting dashboard

Excel-based analytics dashboards gave product and engineering a shared view: accuracy trends per release, worst-performing fields, and document types needing more training data. Reports generated automatically at the end of every CI run.