Extract values, labels, relationships and trends from charts and tables

[iliia-veselov]

What is the problem this feature will solve?

Retrieval is half the problem. The other half is extraction: capturing values, labels, relationships and trends as presented in the original source. "As presented" is the operative phrase — an answer must reflect what the chart or table actually shows, not the model's impression of it.

This is the part most likely to produce confidently wrong numbers, and a wrong figure attributed to a real source document is worse than no answer at all.

What is the proposed feature or solution?

1. Extract readable quantities from a chart or table: series, categories, units, periods, and values where they are legible.
2. Capture relationships and trends the source itself asserts — direction, inflection, comparison — without inventing analysis the source does not support.
3. Attach provenance so an extracted figure traces back to the page and figure it came from; this feeds citation behaviour downstream.
4. Flag low-confidence extractions rather than emitting them as fact. A missing number is far less damaging than a wrong one.

Acceptance criteria

• Values extracted from a chart or table match the source.
• Units and periods are captured alongside values.
• Trends stated in an answer are supported by the source, not inferred beyond it.
• Each extracted figure is traceable to its page and figure.
• Low-confidence extractions are flagged rather than silently emitted.
• Extraction accuracy is measured against hand-labelled ground truth.