A dense text index stores vectors of text. A chart inside a PDF is a raster region, and nothing in the embedding step turns one into the other, so the values printed in a figure are absent from the searchable space unless some earlier stage writes them out [5]. A caption fixes the topic and not the numbers. Databricks separates the two failures that follow: the retriever returns the wrong page, or it returns the right page carrying no usable values [5]. It says it first saw the same gap in its own OfficeQA Pro results, where chart and multimodal questions scored worse than the rest [4].
The evaluation is built to tell those two failures apart. Hit Rate@10 over a fully parsed 16,000-page corpus means the top ten results are one page in 1,600 [9][12][1]. At that ratio the retrieval half is the harder half, and a caption cannot rescue it, because the caption never contains the figure the question asks about [5].
The protocol is more careful than the demo. 310 ViDoRe V3 questions and 114 Chart-RAG questions, run three times each, come to 1,272 graded answers, each scored Correct, Partially Correct or Incorrect by a gemini-3-flash judge against a gold answer [8][10][11][13][2]. Two indexes from identical PDFs, the same chunker and the same 300M BGE embedder, differing only in figure representation, is the comparison you actually want to read [7].
The part I would copy is Chart-RAG. Databricks noticed that many of its chart-selected ViDoRe questions could still be answered from surrounding prose, and therefore were not testing chart reading, so it wrote 114 questions against 378 pages of BIS, IMF and J.P. Morgan reports [10]. That is one question per 3.3 pages against one per 52 pages in ViDoRe, about sixteen times the density [3]. Confound control by construction rather than by caveat.
The opening comparison is one chart, one question, an unnamed frontier model, and a miss of exactly one count [2][3][4]. Counting offers no partial credit, though the judge does [11].
Where this design breaks is the seam. Once ai_parse_document has emitted a figure as JSON and the answering agent reads that JSON, the agent has no pixels left to check it against [6]. A misread axis or a dropped series arrives downstream indistinguishable from good data, so pipeline accuracy becomes parser accuracy, visible only through end-to-end scores [6][15].
For the win to transfer you need charts whose series are recoverable as values, questions that turn on those values rather than on topic, and a corpus big enough that wrong-page retrieval is your dominant failure mode rather than a rounding error. Financial reporting, which is where Databricks says it sees the gap in customer systems, fits all three [5]. A slide deck of hand-drawn diagrams fits none of them.
The summary line in the material available states that structured chart JSON improved both answer quality and retrieval on both datasets, and Databricks says the approach is competitive with large single-vector and multi-vector multimodal embedding models [14][15]. The excerpt stops before the table, so for now the design is more inspectable than the result [16].
