Skip to main content

Documentation Index

Fetch the complete documentation index at: https://parabola.io/docs/llms.txt

Use this file to discover all available pages before exploring further.

Parabola’s PDF parser pulls structured data out of invoices, packing lists, BOLs, rate cards, scanned forms, and even handwritten notes using OCR and large language models. The parsed data lands in your flow as a normal table you can clean, join, and route to other systems.

​
Three ways to bring a PDF into a flow

  • Upload directly with the Extract from PDF step — best for one-off files.
  • Pull from inbound email with Extract from email — best when partners or vendors send PDFs by email.
  • Process in bulk with Pull from file queue — best for batch processing many PDFs accessible via URL.
The parsing configuration is the same across all three sources.

​
Columns vs. keys

The parser returns data as columns or keys:
  • Columns are values that repeat down the document — line items in an invoice, rows in a packing list.
  • Keys are document-level values that appear once and apply to the whole document — invoice number, PO date, total.
  • The AI sometimes flips the two. If a value isn’t pulling correctly, try the other option.
  • Both accept extra context (descriptions, examples, instructions) to improve accuracy.

​
Pick a parsing method

Auto-detected table (default). Parabola scans the PDF, finds candidate tables, and labels likely columns. Quickest setup; works best when the document has clear, headered tables. Detected tables show up under “Use an auto-detected table.” You can add columns or keys manually after. Custom table. Define the table structure by hand if auto-detect didn’t catch it. Name the table and add columns with + Add Column. Best for multi-table documents and tables that span pages — more setup, more control. Extract all data (OCR-first mode). Returns all text from the PDF using OCR. Use only when the first two methods don’t return what you need, or when you plan to feed the result into another AI step downstream. Return formats:
  • All data — every value, one per row
  • Table data — tables split by page, each with a table ID
  • Key-value pairs — labeled items like SKU: 12345
  • Raw text — one cell per page, useful for follow-up AI parsing

​
Extract values

For document-level fields like invoice number or PO date, add them under “Extract values” with + Add key. Each key becomes its own column with the value repeated across every row.
  • Names can be descriptive — they don’t have to match the PDF text exactly. Pick something the AI can connect to the field.
  • Examples are the highest-leverage way to improve accuracy.
  • “Additional instructions to find this value” is optional but helpful for tricky cases. Example: to split an order ID like ABC:123 into two columns, instruct the parser to “Take the order ID and extract all of the characters before the ’:’ into a new column.”
The example below shows the impact of additional instructions on a handwritten YES/NO field.
Extract values configuration with handwriting recognition example

​
Fine tuning

Add overall context and instructions in the fine-tuning text box. Specific examples and clear scenarios outperform vague guidance. The chat panel on the left can help you draft the prompt.

​
Advanced settings

Text parsing approach. Default is “Auto.” Other options:
  • OCR — slower, better for handwriting and complex scans
  • Markdown — generally faster, often works better on PDFs with nested columns
Retry step on error. On by default. Retries the step once on a transient error. Auto-update prompt versions. Off by default. Parabola occasionally updates the underlying prompts. Because that can shift output, this is opt-in. Page filtering. Off by default. Use it to parse specific pages and speed up runs.
  • Keep, Remove, or Autodetect — Autodetect lets the parser pick pages
  • The first / the last / these — set a number for “first” or “last,” or list specific pages (e.g., 1, 10, 16)

​
Child columns

Mark a column as a “Child column” when its values don’t repeat with the parent — for example, sizes within a product line. Before:
Parsed table before applying child column setting
After marking “Size” as a child column:
Parsed table after marking Size as a child column

​
Tips & limits

  • Fewer pages parsed = faster runs. Use page filtering when you only need data from a few pages.
  • Multiple tables in one file usually need multiple PDF steps — one per table.
  • File limit: under 500 MB and 30 pages.
  • PDFs cannot be password-protected.
  • Always audit AI-parsed results before relying on them.

​
Choosing the right step

For a single one-off PDF, drag a file onto the canvas to use the Extract from PDF step.
Extract from PDF step with an uploaded file
For PDFs arriving as email attachments, configure them on the Extract from email step using the same parsing options.
Extract from email step configured to parse PDF attachments
For batch-processing many PDFs at once, use Pull from file queue. It accepts files via URL and runs are added to the queue via API or via Run another Parabola flow.
Last modified on May 18, 2026