Document Processing

Flow-Like can read, transform, and extract structured information from document collections. Prefer deterministic readers and converters first, then add schema extraction or a configured AI model when layout or language makes rules insufficient.

Choose a processing path

Input	Start with	Add when needed
Digital PDF	Text extraction or page rendering	AI extraction for complex visual layouts
Scanned PDF	Render pages to images	Vision-capable extraction and validation
Excel workbook	Cell, worksheet, or table nodes	AI table extraction for unusual layouts
CSV	Buffered reader or database registration	Batching and schema validation
Image	Read, inspect dimensions, crop, resize, convert	Barcode reading, annotation, or AI extraction
DOCX or PPTX	Native extraction and editing nodes	Template-specific replacement or generation
HTML	Convert to Markdown	Section or keyword extraction

PDF processing

Inspect, render, and extract

Need	Node
Count pages	PDF Page Count
Render one page	PDF Page To Image
Render every page	PDF To Images
Extract selectable text	Extract Text
Split or extract page ranges	Split PDF, Extract Pages
Rotate pages	Rotate Pages
Merge files	Merge PDFs

Use text extraction for digitally generated PDFs. Render pages when downstream work depends on the visual layout or when the source is scanned.

AI-assisted extraction

AI Extract Document can describe images and recover content from visually complex documents. AI Extract Documents handles multiple files in parallel.

Pass a configured model that supports the source format. Do not hard-code a vendor-specific model name into reusable boards. Model availability and capabilities depend on the configured provider.

For structured output, define a schema and validate the extracted values before writing them. An invoice schema might look like:

{
  "vendor": "string",
  "invoice_number": "string",
  "date": "date",
  "line_items": [
    {
      "description": "string",
      "quantity": "number",
      "price": "number"
    }
  ],
  "total": "number"
}

Use deterministic checks for totals, dates, identifiers, and required fields. Route low-confidence or invalid records to review instead of silently accepting them.

Spreadsheet processing

Cells and worksheets

Need	Node
Read a cell	Excel Read Cell
Write a cell	Excel Write Cell
List sheets	Get Sheet Names
Create a sheet	New Worksheet
Copy a sheet	Copy Worksheet

Tables

Use Extract Tables (Excel) for predictable workbook layouts. Use Extract Tables AI (Excel) when tables have irregular headers, spacing, or multiple regions that deterministic extraction cannot identify reliably.

For either path:

1. inspect sheet names and choose the intended worksheet;
2. define expected columns and types;
3. normalize headers;
4. validate row counts and required fields;
5. preserve the source workbook or a stable reference to it.

CSV processing

Buffered CSV Reader reads large CSV files in batches. Keep the batch size appropriate to row width and downstream work, and validate the header before processing the first batch.

CSV files can also be registered in a DataFusion session and queried with SQL:

SELECT
  c.name,
  SUM(s.amount) AS total
FROM sales AS s
JOIN customers AS c
  ON s.customer_id = c.id
GROUP BY c.name
ORDER BY total DESC;

SQL is useful for joins, aggregation, and filtering, but it does not replace source validation. Confirm delimiters, quoting, encoding, and numeric or date conventions when files come from multiple systems.

Image processing

Need	Node
Load an image	Read Image
Read dimensions	Get Dimensions
Resize	Resize Image
Crop	Crop Image
Convert color representation	Color Convert
Adjust contrast	Contrast
Read a QR code or barcode	Read QR-/Barcode
Draw review annotations	Draw Boxes
Save an image	Write Image

Resize large scans before model-based extraction when the reduced image still preserves the required text. Keep the original file for audit, reprocessing, or a higher-resolution retry.

DOCX and presentation files

The document catalog includes native operations for office files:

• Extract Text from DOCX, replace text or images, merge documents, and build documents from paragraphs, tables, images, and links.
• Extract Text from PPTX, replace slide content, merge presentations, and add slides, tables, charts, shapes, or speaker notes.

Use placeholder and replacement operations for controlled templates. Use native creation nodes when the workflow needs to assemble a new document from structured data.

Text and template processing

Task	Node
Convert HTML to Markdown	HTML to Markdown
Extract content sections	Extract Content Sections
Extract deterministic keywords	RAKE Keywords, YAKE Keywords
Extract semantic keywords	AI Keywords
Summarize a document	Summarize Document
Render a text template	Render Template

Choose deterministic keyword extraction when reproducibility and cost matter most. Use a model when the task depends on meaning rather than surface terms, and record the provider and model configuration with the run when reproducibility matters.

Batch-processing pattern

For a folder or upload collection:

1. enumerate the input files;
2. identify or validate each file type;
3. send each type through its dedicated reader;
4. normalize all results into a shared schema;
5. validate required fields and business rules;
6. store the structured result and source reference;
7. route failures or uncertain extractions to review;
8. emit a summary with processed, skipped, reviewed, and failed counts.

Limit concurrency for large documents and external model calls. A large collection should be restartable, so save progress or make each file operation idempotent.

Quality and safety checklist

• File type is validated instead of trusted from the extension alone
• Original files or stable source references are retained
• Deterministic extraction is preferred where it is sufficient
• Model choice is configurable and supports the input format
• Required fields and business rules are validated
• Large collections are batched and concurrency-limited
• Low-confidence results have a review path
• Sensitive document content is not exposed in logs
• Output records include provenance back to the source

Related guides

• Summarization strategies
• Data pipelines
• API integrations
• Node catalog