Compatibility, route by route

Every route and parse option was tested against Reducto's live API with the same request (12 November 2026).

Route by route against Reducto's published OpenAPI schema (20 routes). "Full" = documented request fields accepted with their defaults, response validates against Reducto's schema, behaviour implemented. Unknown fields are ignored, as on Reducto.

Route	Level	Notes
POST /upload	full	Multipart file, ?extension=. Without a file: returns a presigned-style PUT URL on this server.
DELETE /upload/{file_id}	full
POST /parse	partial	See the field table below. Inputs that work out of the box: PDF (also password-protected), PNG / JPG / GIF / BMP / WEBP / TIFF (multi-page), DOCX, XLSX / XLSM, CSV, PPTX, TXT, EML, MSG. DOC / RTF / ODT need LibreOffice or macOS textutil; XLS / PPT need LibreOffice (brew install --cask libreoffice, or apt install libreoffice); HEIC needs a decoder on the machine (pillow-heif, macOS sips or heif-convert). Without the tool the answer is 415 with that advice. An async block queues the job and returns {job_id}, as the spec allows.
POST /parse_async	partial	Same options. async.priority, async.metadata, async.webhook implemented; queue_priority accepted and ignored.
POST /extract, /extract_async	full route, quality depends on readanydoc.tasks.extract	The route parses (or reuses jobid://), calls the task, returns V3ExtractResponse. Additions that Reducto's API does not have (see "Extract additions" below): a warnings list in every response, and three opt-in request fields.
POST /split, /split_async	full route, as above	SplitResponse.
POST /classify, /classify_async	full route, as above	ClassifyResponse.
POST /edit, /edit_async	full route, as above	EditResponse; document_url is a 24 h link on this server. Fillable PDFs: the file's own fields are filled. Flat and scanned PDFs: every detected (or given) box becomes a real form field named temp_<n>, text or checkbox, filled in edit_options.color, so the output stays fillable, as Reducto's does (W-2 scan: 69 fields against Reducto's 68, 24 filled against 23; CMS-1500: 180 against 167, 14 filled on both). edit_options.flatten bakes the values into the page; edit_options.flat_fields: false (ours) draws text instead of adding fields.
POST /pipeline, /pipeline_async	accepted, 501	Deprecated by Reducto. The body is validated, then a 501 error tells you to chain /parse and /extract.
GET /job/{job_id}	full	Sync calls are stored as jobs too. status uses Reducto's four values (running = Pending, cancelled = Failed with a reason) unless READANYDOC_JOB_STATUS_DETAIL=1.
DELETE /job/{job_id}	full	202 deleted, 409 still running, 410 already deleted or expired, 404 unknown. include_persisted accepted; deletion always removes everything.
POST /cancel/{job_id}	full	Pending jobs are cancelled at once; a running job finishes its current work and its result is discarded. 409 if already finished.
GET /jobs	full	cursor, limit (1-500, default 100), exclude_configs. Newest first. Reports all six statuses.
GET /version	full	Returns "readanydoc-<version>".
POST /configure_webhook	partial	There is no Svix portal. Returns (and with ?url= sets) the default webhook URL as a string.

Parse request fields. Every row was run against Reducto's API and this server with the same request on 2026-11-12. "Same" means the status code and the option's effect match; text differs wherever the two readers differ.

Field	Level	Against Reducto's real response
input: reducto:// id, upload object, http(s) URL, jobid://<id>, list of those	full	a list is parsed in order into one response with pages renumbered
retrieval.chunking.chunk_mode = disabled, page, block, section, page_sections	full	same rules: fed Reducto's own blocks, our shaper returns its chunk text character for character (35 of 35 cases; the 34 filter, marker, page-range and merge cases too). Chunks repeat the headings they sit under with " (cont.)", as Reducto's do
chunk_mode = variable, chunk_size	full	same grouping in 16 of 19 cases (null size = about 1000 characters, upper bound size + 25%, tables never split). Reducto also cuts by "visual context" (in front of page footers); that is not reproduced
chunk_overlap	full	N characters of the previous and of the next chunk, as Reducto returns it
retrieval.filter_blocks	full	same: filtered types leave content / embed, the blocks stay listed. Depends on block typing
retrieval.embedding_optimized	full	embed carries a model-written summary of each table instead of its HTML (one text call per table, about 1,400 input + 950 output Haiku tokens). Without a live model: header: value lines
formatting.add_page_markers	full	same: [[START OF PAGE n]] / [[END OF PAGE n]] in content and as Page Number blocks with a zero box
formatting.table_output_format = html, md, json, csv, dynamic	full	same formats in chunk and block content. Markdown is unpadded (\| a \| b \|); Reducto pads the columns
formatting.table_output_format = jsonbbox	full	differs: Reducto rejects it with r-1 (400); this server answers it
formatting.merge_tables	full	same rule (consecutive tables with the same column count merge; fragments listed in merged_tables). The outcome depends on how many columns each reader saw: equal on 5 of 7 test files
formatting.include	partial	hyperlinks, comments, change_tracking for DOCX / PPTX / PDF annotations; DOCX hyperlinks show without it, as on Reducto. highlight, signatures, ignore_watermarks accepted, ignored (r-1 ignores them too)
settings.page_range (object, list of objects, list of ints, sheet names)	full	same pages and original_page; a range outside the document is 400 INVALID_CONFIG on both. Reducto ignored a numeric range on a workbook and on a multi-page TIFF; this server honours it
settings.document_password	full	same: right password parses, wrong or missing is 442 DOCUMENT_PASSWORD_PROTECTED
settings.return_images = figure, table	full	same: image_url on Figure and Table blocks (a PNG crop in this server's file store, link valid for READANYDOC_RESULT_TTL_HOURS). Figure boxes come from the page image, mean overlap with Reducto's boxes 0.74 on 7 figures. Pictures the reader did not describe (logos, stamps, seals, QR codes) get a Figure block with the shape option layout_figures (off by default; needs the engine profile; turn on with READANYDOC_STAGE_OPTIONS='{"shape": {"layout_figures": true, "layout_figures_describe": true}}'). On the 22 business pages that hold Reducto's 23 Figure blocks: 8 Figure blocks without the option, 24 with it; Reducto's figures matched by one of ours (IoU >= 0.5) 4 -> 15 of 23, mean IoU 0.91; every other block unchanged. layout_figures_describe adds a one-sentence description per new figure (one small image call each, about 120 input + 120 output tokens); without it the new block has a box, an optional crop and empty content
settings.return_images = page	accepted	Reducto's response showed no page images either
enhance.agentic[{scope: "figure", prompt}]	full	same: the Figure block shows the model's answer to the prompt (one call per figure, about 1,000 input + 150 output Haiku tokens)
enhance.agentic[{scope: "figure", advanced_chart_agent}], enhance.advanced_chart_agent	partial	chart_data (a list with one link to a JSON file in Reducto's shape), extra.is_chart, block content = legend + data table. One Haiku read per figure (about 1,300 input + 6,000 output tokens for a 144-point chart), values within 2.3 on a mean of 65 of Reducto's on the test chart, pie labels exact. Verification pass: free consistency checks on the data (values inside the axis range, pie adds up, series lengths, repeated x values); one more read only when a check fails; result in extra.chart_verified (Reducto's key) and extra.chart_verification. Reducto verifies by re-drawing the chart; no extra.chart_reconstruction image here
enhance.agentic[{scope: "table", prompt}]	full	same effect on the two test pages ("write every dollar amount without the dollar sign": 23 -> 0 and 2 -> 0 dollar signs in tables on both). One text-only Haiku call per table that returns the changed cells; cells are edited in place, so rows, columns and spans cannot change. mode is accepted; with a prompt every table is processed. Without a prompt nothing extra runs
enhance.agentic[{scope: "text", prompt}]	full	same effect on the two test pages ("write dates as YYYY-MM-DD": the same 2 and 1 dates rewritten, usage_breakdown.prompted_blocks 1 on both). One text-only Haiku call per page for its text, form, header and footer blocks; a change that rewrites a block instead of reformatting it is dropped. Word citations of a changed block keep their boxes but not exact character offsets
enhance.summarize_figures, enhance.intelligent_ordering, settings.ocr_system, settings.extraction_mode	accepted, ignored	legacy-model knobs; r-1 ignores them too
settings.return_ocr_data	partial	result.ocr.words / lines in Reducto's shape. About the same word count on digital pages (131-456 against 134-406); on the scanned test page 378 words against Reducto's 888, and fewer lines on tables
settings.return_citations	full	extra.word_citations on text blocks (72 of 73 in the test; Reducto 76 of 76) and cell_citations on every table, in Reducto's shape
settings.extract_document_properties	full	same values on PDF and DOCX
settings.force_url_result, settings.persist_results	full
settings.model	full	usage_breakdown.parse_model reports the requested model as Reducto does (R-1, R-1 Mini, Legacy; legacy counted in legacy_parse_credits). r-1-mini and legacy run a cheaper tier of the configured engine: engine -> model-only (one page read, no zoom vote, no second model, no repair calls) at the API's default effort instead of high (measured at 35% cheaper and twice as fast, equal on text pages, slightly worse on tables). usage.model_usage.profile / .reader say what ran. READANYDOC_PROFILE_R_1_MINI, READANYDOC_PROFILE_LEGACY, READANYDOC_READER_MINI override; READANYDOC_MINI_SAME_PROFILE=1 = one engine for all three
settings.embed_pdf_metadata, embed_pdf_metadata_dpi	full	pdf_url is a searchable copy: our words (the reader's text on its word boxes, OCR words elsewhere) as an invisible text layer; any script. Pages with a good text layer are untouched; a page with an old unusable layer is replaced by its picture first (at embed_pdf_metadata_dpi, default 150); image inputs become a PDF. On a scanned W-2: 431 words in our layer, 219 (line fragments without spaces) in Reducto's; 88% of the characters of Reducto's layer are in ours and 90% the other way; 84 of 85 distinct words of our response can be found in our layer. Pages turned before reading get no layer
spreadsheet.clustering, split_large_tables, include, exclude, max_cell_count	full	same table counts for split sizes and hidden sheets, same 422 CELL_COUNT_EXCEEDED, cell colours and formulas as cell attributes. clustering=accurate uses the rule-based clustering. Table boxes are page-relative here, cell coordinates on Reducto
settings.timeout, tenant_throttling, force_file_extension, hybrid_vpc	accepted, ignored
v2-only fields (options, advanced_options, experimental_options, top-level webhook / priority)	accepted, ignored	v2 document_url is read as input
invalid enum values (chunk_mode, filter_blocks, table_output_format)	full	422 on both

File types are compared in File types.