Parse Document To Structured

Client POSTs a PDF to the /parse endpoint and receives a result containing the document rendered as layout-aware Markdown plus a structured list of content blocks (each with a type and bounding-box position). The integrator assum…

Consume both the Markdown rendering AND the per-block structure: each block carries a type (heading, paragraph, table, figure, list, etc.) and page-relative position. Downstream RAG indexing should preserve block boundaries rather than collapsing to one string, so retrieval can cite a specific bloc…

Integrator needs clean Markdown for an LLM prompt but configures the parse request to return the most verbose structured JSON, then re-serializes it to Markdown themselves in application code.

Request the output representation that matches the downstream consumer: a Markdown/text rendering for LLM prompting, the structured block JSON for programmatic indexing. Do not hand-roll a Markdown serializer over the JSON — that reintroduces the reading-order and layout bugs Reducto already solved…

A 40-page report has two-column layouts on some pages and single-column on others. The integrator concatenates parsed text in raw top-to-bottom pixel order without trusting Reducto's reading-order reconstruction.

Rely on Reducto's reading-order reconstruction (which sequences multi-column and interleaved layouts into logical reading order) rather than naive top-to-bottom sort. Verify on a sample of mixed-layout pages that columns are not interleaved mid-sentence before trusting the order in bulk indexing.