Get tables and figures out of PDFs without flattening them into useless text

You are designing how a RAG ingestion pipeline handles the non-plain-text content inside a annual financial reports and product spec sheets (PDF) — tables, charts and figures, and embedded images — deciding a specific extraction strategy per content type instead of routing everything through one generic text extractor that turns a table into a stream of numbers with no structure left.

DOCUMENT PROFILE
Financial reports contain multi-row-header tables (a metric broken out by quarter and by region in the same table) and bar/line charts; spec sheets contain simple single-header comparison tables and product photos

CONTENT TYPES PRESENT
multi-header financial tables, single-header spec tables, bar/line charts with axis labels, product photography, company logos on every page header

EXTRACTION STRATEGY PER TYPE
For tables:
- State whether tables get extracted to a structured format, such as a Markdown table with column headers preserved, versus flattened into a plain-text left-to-right, top-to-bottom read order. Flattening a table with more than a couple of columns produces text where a number is disconnected from the column header and row label that actually give it meaning — decide explicitly whether this document profile's tables are simple enough for flattening to be acceptable, or complex enough, with multi-row headers or merged cells, that structure-preserving extraction is required.
- State whether a table gets embedded and retrieved as one unit, so a query never gets half a table with no header, or whether individual rows can be independently retrieved, which is only viable if each row genuinely stands alone in meaning without its header row present in the same chunk.

For charts and figures:
- State whether these get a caption-style description generated by Claude (vision) generating a structured description per chart, and if so, what that description is required to state explicitly — the actual data or trend shown, not a description so generic it would be equally true of a hundred different charts.
- State what happens to the figure's own caption or label text if the source PDF has one — it should be preserved verbatim alongside the generated description, not replaced by it, since the caption often carries information the generated description won't reproduce, such as a figure number referenced elsewhere in the document.

For embedded images with no clear semantic role, such as logos or decorative elements:
- State the rule for skipping these rather than generating a description for every image indiscriminately — a vision-model description of a company logo is pure noise in a retrieval index and costs a real vision-model call for zero retrieval value.

RETRIEVAL IMPLICATIONS
Analysts ask both direct lookup questions ('what was APAC revenue in Q3') and interpretive questions ('why did the chart show a dip in Q2') — the pipeline needs to support both — for each content type above, state explicitly how a real query would need to reach it: does a question about a specific figure need the table's numeric content directly retrievable, or does it need the surrounding paragraph that already interprets the table in prose? Sometimes the answer is genuinely both, and the pipeline needs to keep them linked, not just co-located.

VALIDATION
Run 10 real analyst questions against the extracted content end to end and confirm the correct number/trend is retrievable and correctly attributed, not just that a chunk containing roughly the right area of the page was returned — state how you'll verify the chosen strategy actually preserved enough structure to answer a real question against a real extracted table, not just that extraction ran without an error.

OUTPUT FORMAT
1. The extraction strategy per content type, stated as a concrete rule.
2. How each extracted piece gets linked back to its surrounding text and page number.
3. One worked example: a specific table from the document profile, its chosen extraction format, and a sample query it needs to answer correctly.