The words are present, but the relationship among them was lost. A bag of extracted labels does not tell us which color belongs to which series, which point belongs to which month, what the log scale means, or whether the footnote qualifies the line we are discussing. If the assistant guesses from OCR reading order, it can cite the correct page and still get the direction wrong. This is not primarily a retrieval recall problem.

I would inspect the rendered page and every intermediate representation: page crop, parser blocks, OCR output, chart detection, legend mapping, extracted values with coordinates, and the final chunk that reached the model. Identify where series identity or axis scale was dropped. Then decide the answer path. If the underlying data table is available and authoritative, query it with the same metric definition and period. Otherwise use a chart-reading path that returns structured observations with units, scale, series and uncertainty, and require verification against the image for the material claim. DePlot is one research example of translating plots to a table-like representation before reasoning. Its existence does not make arbitrary chart extraction exact.

For this specific question, the evidence artifact should say something like: source file revision and page, chart bounding box, series key, x-axis months, y-axis units and scale, extracted points or directional comparisons, and footnote relationship. A visual crop remains linked for inspection. The index should mark the OCR paragraph and the chart artifact as different evidence types, not concatenate all labels into a single paragraph and let a text reranker call it answer-bearing. If a trend depends on tiny points near a grid line, state uncertainty or use the source table. A log-scale chart can show direction, but screen distance is not an absolute dollar difference.

The parser's “100 percent text coverage” metric is misleading here. Add a figure-level ingest gate before a new parser version becomes searchable. Test whether its structured chart artifact still maps series to legends, units and notes on a reviewed PDF set. Compare old and new representations by document revision, quarantine pages that changed interpretation, and reprocess affected figures from immutable PDF bytes. A successful OCR job should not mark a chart ready for numeric questions when chart extraction failed. Keep a visible ingest state for that figure, so retrieval can prefer a trusted table or say the visual evidence is unavailable.

An interviewer might say a vision model can simply read the page. It might, and I would compare it against the structured path on the exact failure slice. But the output still needs a source revision and a way to check the claimed series and time window. A vision model giving a confident answer is not itself the evidence. We can use it to propose an interpretation and verify high-impact claims before publication.

The chart says growth, but the axes were cropped. What can the assistant claim? presents a cropped chart directly to the model and asks what it can claim. Only newly ingested PDFs got worse. How do you find the broken stage? investigates a broad new-PDF ingestion regression. This case is the source representation and indexing contract for charts that an assistant will later retrieve. The ingestion stage must not certify an answer-bearing chart merely because it captured every word on the page.