Parse, Extract, and Split Documents with Reducto in Jinba Flow

• REDUCTO_UPLOAD
  • Upload a file to Reducto and get back a reducto:// URL for use with the other Reducto tools.
  • Inputs: file_url (publicly accessible or presigned URL), optional extension hint (e.g. pdf, docx).
  • Outputs: file_id and reducto_url.
• REDUCTO_PARSE
  • Parse a document to extract text, tables, and figures with layout-aware chunking.
  • Inputs: document_url (public URL, presigned S3 URL, or reducto:// URL), plus options such as chunk_mode (variable, section, page, disabled, block, page_sections; default disabled), chunk_size (250–1500 characters), table_output_format (html, json, md, jsonbbox, dynamic, csv; default dynamic), add_page_markers, merge_tables, ocr_system (standard / legacy), extraction_mode (ocr / hybrid), page_range (e.g. "1-5", "1,3,5", "1-3,7-9"), and document_password for encrypted files.
  • Outputs: job_id, num_pages, and an array of chunks with their content blocks.
• REDUCTO_EXTRACT
  • Extract specific fields from a document into structured JSON using a user-defined JSON Schema. Ideal for invoices, forms, and contracts.
  • Inputs: document_url, schema (JSON Schema string), optional system_prompt, array_extract (repeating data such as invoice line items), include_images, optimize_for_latency, and citations_enabled.
  • Outputs: extracted_data, plus job_id, num_pages, and num_fields.
• REDUCTO_SPLIT
  • Split a document into sections using natural language category descriptions.
  • Inputs: document_url, split_description (JSON array of {"name": ..., "description": ...} categories), optional split_rules (natural language splitting rules), and table_cutoff (truncate / preserve for tables at page boundaries).
  • Outputs: num_pages and splits (each with name, pages, and confidence).

​

Connect with a Reducto API key

​

Parse a document

- id: input_file
  name: input_file
  tool: INPUT_FILE
  input:
    - name: description
      value: "PDF report to parse"

- id: upload_report
  name: upload_report
  tool: REDUCTO_UPLOAD
  needs: [input_file]
  config:
    - name: api_key
      value: "{{secrets.REDUCTO_API_KEY}}"
  input:
    - name: file_url
      value: "{{steps.input_file.result}}"
    - name: extension
      value: pdf

- id: parse_report
  name: parse_report
  tool: REDUCTO_PARSE
  needs: [upload_report]
  config:
    - name: api_key
      value: "{{secrets.REDUCTO_API_KEY}}"
  input:
    - name: document_url
      value: "{{steps.upload_report.result.reducto_url}}"
    - name: chunk_mode
      value: page
    - name: table_output_format
      value: md
    - name: add_page_markers
      value: true

- id: summarize_report
  name: summarize_report
  tool: OPENAI_INVOKE
  needs: [parse_report]
  config:
    - name: version
      value: gpt-5.2
  input:
    - name: prompt
      value: |
        Summarize the following document. Highlight key figures and tables.

        {{steps.parse_report.result.chunks}}

​

Extract fields

- id: extract_invoice
  name: extract_invoice
  tool: REDUCTO_EXTRACT
  config:
    - name: api_key
      value: "{{secrets.REDUCTO_API_KEY}}"
  input:
    - name: document_url
      value: "https://example.com/invoice.pdf"
    - name: schema
      value: '{"type":"object","properties":{"invoice_number":{"type":"string"},"total":{"type":"number"}}}'
    - name: array_extract
      value: true

​

Notes

• document_url accepts public URLs, presigned S3 URLs, and reducto:// URLs. Use REDUCTO_UPLOAD first when the file is not directly reachable by Reducto, then pass the returned reducto_url to the other tools.
• REDUCTO_PARSE defaults: chunking is disabled, table format is dynamic, OCR system is standard (multilingual; legacy supports Germanic languages only), and extraction mode is hybrid (OCR combined with native text).
• page_range is 1-indexed and accepts ranges and comma-separated lists (e.g. "1-3,7-9").
• schema (REDUCTO_EXTRACT) and split_description (REDUCTO_SPLIT) must be valid JSON strings; the step fails with an error if they cannot be parsed.
• On API errors the tools return an error object with the HTTP status and the raw response, which you can inspect in the run log.