MKTRCSYNTHETIC TRAINING NOTE

Your PDF Chart Became JSON. But Did the Numbers Survive?

MKTRC · data-quality note | Based on Databricks' October 5, 2026 product announcement. All Q1/Q2 values below are deliberately synthetic teaching figures, not results from a real Databricks job or our market research.

A chart in a PDF used to be difficult to turn into machine-readable data. Databricks now documents chart-value extraction by ai_parse_document, returning JSON for readable chart figures, including series values, legend labels and available axis details. That is useful. It does not make every extracted number ground truth.

What the official feature actually changes

Databricks' October 5, 2026 release notes describe structured chart extraction. Charts are represented as figure elements; when data can be read, content contains chart data as a JSON string. Some figure contents may be null. The results have a versioned structure, and support varies by runtime, compute mode and region.

A one-digit error that changes the story

Consider a fictional quarterly chart with Q1=12 and Q2=15. Now imagine a model output that incorrectly reads Q2 as 16.

Source Q1 Q2 Increase
Fictional source chart 12 15 +3
Fictional extracted values 12 16 +4
Validation Match Mismatch The conclusion changes

Perfectly formed JSON does not prove that the chart was read correctly. Confusing an axis scale, a legend or a plotted point can propagate a bad number into downstream retrieval and reporting.

A validation gate worth keeping

  1. Preserve the source — original PDF hash, version, retrieval time and page.
  2. Link the location — record the figure's page and bounding box or rendered image reference.
  3. Check structure — axis units, legend labels, series ordering, missing/null values and numeric format.
  4. Reconcile key values — independently compare decision-critical numbers with the source image, not just a model confidence field.
  5. Gate the result — mark unreadable or mismatched values for review rather than silently promoting them into an analytical table.
SOURCE_PDF → FIGURE + PAGE/BBOX → PARSED_JSON
                            ↓
                     INDEPENDENT CHECK
                       /           \
                  MATCHED       REVIEW_REQUIRED

MATCHED and REVIEW_REQUIRED are example states in our editorial validation workflow, not statuses returned by the Databricks function.

Scope and limitations

This concerns Databricks-managed ai_parse_document, not a claim about every OCR product. Official documentation requires supported compute and regions and warns about extraction/model errors and content that may be unreadable. SQL function limits and limits of a separate REST API route should not be conflated. We did not run a Databricks job or measure chart-extraction accuracy, SEO lift, trading outcomes or profitability.

The takeaway

JSON is the extraction format; verification is the evidence standard. Store enough source context to challenge the numbers before using them in research.

Official references: October 2026 Databricks release notes · ai_parse_document function