Skip to content
Log in

Extract Async

POSThttps://platform.reducto.ai/extract_async

Body

  • asyncconfig__v3__AsyncConfig

    The configuration options for asynchronous processing (default synchronous).

    3 properties
    • metadataunknown

      JSON metadata included in webhook request body. Defaults to None.

    • priorityboolean

      Workers poll the priority queue ahead of the standard queue, so priority jobs start sooner when there is queued work; sync jobs are prioritized above async jobs by default.

      Default
      false
    • webhookSvixWebhookConfig | DirectWebhookConfig

      The webhook configuration for the asynchronous processing.

      2 variants
      • modestring
        Default
        "svix"
      • channelsstring[]

        A list of Svix channels the message will be delivered down, omit to send to all channels.

      • modestring
        Default
        "direct"
      • urlstringRequired
  • inputstring | string[] | UploadResponseRequired

    For parse/split/extract pipelines, the URL of the document to be processed. You can provide one of the following: 1. A publicly available URL 2. A presigned S3 URL 3. A reducto:// prefixed URL obtained from the /upload endpoint after directly uploading a document 4. A jobid:// prefixed URL obtained from a previous /parse invocation 5. A list of URLs (for multi-document pipelines, V3 API only)

            For edit pipelines, this should be a string containing the edit instructions
    
    UploadResponse properties
    • file_idstringRequired
    • presigned_urlstring
  • parsingParseOptions

    The configuration options for parsing the document. If you are passing in a jobid:// URL for the file, then this configuration will be ignored.

    5 properties
    • enhanceEnhance
      4 properties
      • agentic(TableAgentic | FigureAgentic | TextAgentic)[]

        For legacy Parse, agentic processing uses vision language models to improve text, table, or figure extraction. With r-1, use agentic processing for custom prompts or advanced chart extraction. Agentic processing adds latency.

        3 variants
        • scopestringRequired
        • promptstring

          Custom prompt for table agentic.

        • modestring

          Mode for table agentic: 'default' selectively applies enrichment only to tables likely to benefit, and 'max' runs enrichment on all tables.

          Default
          "default"
          Allowed
          defaultautomax
        • scopestringRequired
        • promptstring

          Custom prompt for figure agentic.

        • advanced_chart_agentboolean

          If True, run advanced chart extraction on figures classified as charts: an agentic extractor that returns full structured series data (chart_data) plus a reconstruction image re-drawn from that data (extra.chart_reconstruction). Higher latency. Defaults to False.

          Default
          false
        • scopestringRequired
        • promptstring

          Custom instructions for agentic text. Note: This only applies to form regions (key-value).

      • advanced_chart_agentboolean

        If True, run advanced chart extraction on figures classified as charts, without requiring a figure-scoped agentic entry. Returns full structured series data (chart_data) plus a reconstruction image re-drawn from that data. Higher latency. Defaults to False.

        Default
        false
      • summarize_figuresboolean

        For legacy Parse, if True, summarize figures using a separate vision language model. r-1 generates figure descriptions natively and ignores this setting. Defaults to True.

        Default
        true
      • intelligent_orderingboolean

        For legacy Parse, if True, use an advanced vision language model to improve reading order accuracy, with a small increase in latency. r-1 handles reading order natively and ignores this setting. Defaults to False.

        Default
        false
    • retrievalRetrieval
      3 properties
      • chunkingChunking
        3 properties
        • chunk_modestring

          Choose how to partition chunks. Variable mode chunks by character length and visual context. Section mode chunks by section headers. Page mode chunks according to pages. Page sections mode chunks first by page, then by sections within each page. Disabled returns one single chunk.

          Default
          "disabled"
          Allowed
          variablesectionpagedisabledblockpage_sections
        • chunk_sizenumber

          The approximate size of chunks (in characters) that the document will be split into. Defaults to null, in which case the chunk size is variable between 250 - 1500 characters.

        • chunk_overlapnumber

          Number of characters of overlap to include from adjacent chunks. Defaults to 0.

          Default
          0
      • filter_blocksstring[]

        A list of block types to filter out from 'content' and 'embed' fields. By default, no blocks are filtered.

      • embedding_optimizedboolean

        If True, use embedding optimized mode. Defaults to False.

        Default
        false
    • formattingFormatting
      4 properties
      • add_page_markersboolean

        If True, add page markers to the output. Defaults to False. Useful for extracting data with page specific information.

        Default
        false
      • table_output_formatstring

        The table output format. Defaults to dynamic, which returns md for simpler tables and html for more complex tables. r-1 does not support jsonbbox.

        Default
        "dynamic"
        Allowed
        htmljsonmdjsonbboxdynamiccsv
      • merge_tablesboolean

        A flag to indicate if consecutive tables with the same number of columns should be merged. Defaults to False.

        Default
        false
      • includestring[]

        For legacy Parse, the formatting details to include in the output. r-1 always includes highlights, signatures, hyperlinks, and watermark removal, and ignores those values.

    • spreadsheetSpreadsheet
      5 properties
      • split_large_tablesSplitLargeTables
        2 properties
        • enabledboolean

          If True, split large tables into smaller tables. Defaults to True.

          Default
          true
        • sizenumber | SplitLargeTableSizes

          The size of the tables to split into. Defaults to 50. Use 'row' and 'column' to independently specify the number of rows and columns to include when splitting. If you only want to split by rows or columns, set the other value to None.

          Default
          50
          SplitLargeTableSizes properties
          • rownumber

            The number of rows to include in each chunk when splitting large tables. Does not chunk rows if set to None.

          • columnnumber

            The number of columns to include in each chunk when splitting large tables. Does not chunk columns if set to None.

      • includestring[]

        Whether to include cell color, formula, and dropdown information in the output.

      • clusteringstring

        In a spreadsheet with different tables inside, we enable splitting up the tables by default. Accurate mode applies more powerful models for superior accuracy, at 5× the default per-cell rate. Disabling will register as one large table.

        Default
        "accurate"
        Allowed
        accuratefastdisabled
      • excludestring[]

        Whether to exclude hidden sheets, rows, or columns in the output.

      • max_cell_countnumber

        Maximum total non-empty cells allowed across all sheets. If exceeded, the request is rejected with a 422 error. Set to null to disable the limit. Defaults to null.

    • settingsSettings
      17 properties
      • modelstring

        The parse model to use. 'r-1' is the full-featured flagship model, with the highest accuracy and lower latency. 'r-1-mini' gives high accuracy at low cost. Both models include agentic capabilities for complex documents. 'r-1-mini' does not accept spreadsheets. 'legacy' is the previous parsing pipeline. Starting December 7, 2026, Parse defaults to 'r-1'. To keep legacy Parse, set this field to 'legacy'.

      • ocr_systemstring

        The OCR system for legacy Parse. Standard is the best multilingual OCR system. Legacy supports Germanic languages and remains available for backwards compatibility. r-1 uses native full-page processing and ignores this setting. Defaults to standard.

        Default
        "standard"
        Allowed
        standardlegacy
      • extraction_modestring

        The text extraction method for legacy Parse. OCR uses optical character recognition only. Hybrid combines OCR with embedded PDF text. r-1 uses native full-page processing and ignores this setting. Defaults to hybrid.

        Default
        "hybrid"
        Allowed
        ocrhybrid
      • force_url_resultboolean

        Force the result to be returned in URL form.

        Default
        false
      • force_file_extensionstring

        Force the URL to be downloaded as a specific file extension (e.g. .png).

      • return_ocr_databoolean

        If True, return OCR data in the result. Defaults to False.

        Default
        false
      • return_imagesstring[]

        Whether to return images for the specified block types. 'page' returns full page images. By default, no images are returned.

      • embed_pdf_metadataboolean

        If True, embed OCR metadata into the returned PDF. Defaults to False.

        Default
        false
      • extract_document_propertiesboolean

        If True, return properties embedded in the original document. Defaults to False.

        Default
        false
      • embed_pdf_metadata_dpinumber

        Render DPI used when rasterizing the source PDF before embedding the OCR text layer (only applies when embed_pdf_metadata is True). Lower values produce dramatically smaller output PDFs; higher values preserve more detail when zoomed past 200%. Defaults to 200 (good for on-screen viewing); raise toward the source scan DPI for crisper output. Min 50, max 250.

        Default
        200
      • persist_resultsboolean

        If True, persist the results indefinitely. Defaults to False.

        Default
        false
      • tenant_throttlingTenantThrottling

        Per-tenant throttling for multi-tenant applications. Tag each request with your tenant's id to bound how much of your account's concurrency a single tenant can consume. Account-level throttles still apply.

        TenantThrottling properties
        • tenant_idstringRequired

          Your identifier for the tenant (customer, workspace, organization) this request belongs to. Used only for noisy-neighbor throttling inside your account.

        • max_sharenumber

          Maximum fraction of your account's concurrency ceiling this tenant may use, between 0 (exclusive) and 1. Defaults to 0.5.

          Default
          0.5
      • timeoutnumber

        The timeout for the job in seconds.

      • page_rangePageRange | PageRange[] | number[] | string[]

        The page range to process (1-indexed). By default, the entire document is processed. For spreadsheets, you can also provide a list of sheet names.

        2 variants
        • startnumber

          The page number to start processing from (1-indexed).

        • endnumber

          The page number to stop processing at (1-indexed).

        • startnumber

          The page number to start processing from (1-indexed).

        • endnumber

          The page number to stop processing at (1-indexed).

      • document_passwordstring

        Password to decrypt password-protected documents.

      • hybrid_vpcHybridVpcSettings

        Hybrid VPC request-scoped settings.

        1 property
        • environmentstring

          Named Hybrid VPC environment to use for this request. Only applies when your organization has Hybrid VPC environments configured.

      • return_citationsboolean

        If True, locate every word of each text block on the page and return the per-word boxes in the block's extra['word_citations']. Each entry carries the word's text, its char_start/char_end span in that block's content (half-open, tags included), and a bbox covering that word on the page, in the same shape as the block's own bbox (including page). bbox is null when the word cannot be located on the page. Table blocks also get extra['cell_citations']: one entry per non-empty cell with its text, row_idx and col_idx (0-based lists of every row and column the cell spans, so a cell merged across rows 1 and 2 has row_idx [1, 2]), and a bbox framing the cell's words, null when none of them can be located. Requires r-1 or r-1 mini. Not supported with formatting.merge_tables. Defaults to False.

        Default
        false
  • instructionsInstructions

    The instructions to use for the extraction.

    2 properties
    • schemaunknown

      The JSON schema to use for the extraction.

    • system_promptstring

      The system prompt to use for the extraction.

      Default
      "Be precise and thorough."
  • settingsExtractSettings

    The settings to use for the extraction.

    7 properties
    • include_imagesboolean

      If True, include images in the extraction.

      Default
      false
    • optimize_for_latencyboolean

      If True, jobs will be processed with a higher throughput and priority at a higher cost. Defaults to False.

      Default
      false
    • force_url_resultboolean

      Force the endpoint result to be returned in URL form.

      Default
      false
    • array_extractboolean

      Deprecated: prefer deep_extract, which supersedes array extraction for complex and long (array-heavy) extractions via an agentic loop (at higher cost and latency). If True, use array extraction.

      Default
      false
    • deep_extractboolean

      If True, use Deep Extract, an agentic extraction mode that iteratively refines its output to achieve near-perfect accuracy. Best for complex documents where accuracy is critical.

      Default
      false
    • citationsCitations

      The citations to use for the extraction.

      3 properties
      • enabledboolean

        If True, include citations in the extraction.

        Default
        false
      • numerical_confidenceboolean

        If True, enable numeric citation confidence scores. Defaults to True.

        Default
        true
      • parent_blockstring

        How much of the source parse block to embed on each citation's parentBlock. 'full' (default) embeds the verbatim source-block HTML in parentBlock.content. 'bbox_only' suppresses parentBlock.content (returned as an empty string) while keeping parentBlock.bbox and all citation-level fields — this can drastically shrink responses on table-heavy schemas where the same source block is cited many times.

        Default
        "full"
        Allowed
        fullbbox_only
    • page_rangePageRange | PageRange[] | number[] | string[]

      The page range to extract from (1-indexed). By default, the entire document is used. For spreadsheets, you can also provide a list of sheet names.

      2 variants
      • startnumber

        The page number to start processing from (1-indexed).

      • endnumber

        The page number to stop processing at (1-indexed).

      • startnumber

        The page number to start processing from (1-indexed).

      • endnumber

        The page number to stop processing at (1-indexed).

  • queue_prioritystring

    Queue priority. 'batch' places the job in a lower-priority queue for non-urgent bulk work. 'auto' (alias: 'standard') uses the default queue.

    Default
    "auto"
    Allowed
    autostandardbatch

Response

  • job_idstringRequired
  • detailValidationError[]
    5 properties
    • loc(string | number)[]Required
    • msgstringRequired
    • typestringRequired
    • inputunknown
    • ctxobject

Type to search…

escClose