File Analysis node

Status: Current | Last reviewed: 2026-09-09

Analyzes files and documents to extract text, tables, forms, and structured data via OCR, vision models, or both.

Behavior

Extracts content from a file or document. Choose Process with OCR for low-latency structured extraction (Textract / Azure DI / Document AI), Process with LLM for multimodal reasoning with a vision-capable model, or Both to run OCR first and then hand the extraction plus the original image to an LLM for structured interpretation.

Merge-template references in Document reference are resolved at runtime, so the common pattern inside a Foreach is to point at {{<foreach_id>.current_item.s3_uri}} or an equivalent file reference.

You can add this node from the workflow editor add-node menu when building a flow.

Configuration options

Option Required Type Values / default What it controls
Processing type (processing_type) Yes string ocr, llm, both; Default: ocr How to extract content from the document.
OCR integration (ocr_integration_id) No string Credential used to call the OCR service (AWS Textract, Azure Document Intelligence, or Google Document AI).
Document source (document_source) Yes string Where the document lives. Options change based on the selected OCR integration.
Document reference (document_reference) Yes string S3 URI / GCS URI / Azure Blob URL / HTTPS URL / base64 string - depending on Document Source. Merge templates supported.
Features (aws_features) No array Default: ["text","tables","forms"] Textract analyses to run. All three are selected by default.
Natural-language queries (aws_queries) No array Optional Textract questions, one per line. Example: "What is the total amount?" or "Who is the vendor?". Answers land under query_answers keyed by the question text.
Document AI processor ID (gcp_processor_id) No string Processor resource ID (without the projects/.../processors prefix). Example: 8a12f93b5c6d4e0a.
Processor location (gcp_location) No string us, eu; Default: us The GCP region where your Document AI processor lives.
Model ID (azure_model_id) No string Default: prebuilt-document Azure Document Intelligence model. Use a prebuilt like 'prebuilt-invoice', 'prebuilt-receipt', 'prebuilt-document', or a custom model name. The model itself determines what comes back (text / tables / KVs) - Azure DI has no per-call feature toggle.
LLM integration (llm_integration_id) No string Credential used to call the vision-capable model.
Model (llm_model) No string Vision-capable model to call.
Prompt (llm_prompt) No string Default: Extract all relevant information from this document. Instructions sent to the model. The document image is attached automatically in the right format for the selected provider - do NOT embed the image or a URL in the prompt.
LLM input (llm_input_mode) No string image, ocr_text, ocr_text_and_image; Default: ocr_text_and_image When processing_type is 'both', controls what the LLM sees alongside the prompt. Ignored for 'llm' / 'ocr'.
Structured output (llm_output_schema) No object Optional JSON Schema. When set, the model returns a structured object matching the schema, surfaced downstream as structured.<field>.

For LLM vision, Direct OpenAI remains available. Amazon integrations expose Bedrock GPT-5.6 Luna, Terra, and Sol when enabled in the chosen region; Microsoft integrations expose the Azure GPT-5.5 and GPT-5.6 Terra deployment names configured on that resource. GPT-OSS and Safeguard are not vision models.

Outputs

Output Type What it contains
raw_text string Concatenated text from all pages.
pages array Per-page text + geometry.
tables array Detected tables as cell grids.
key_values array Key-value pairs (forms).
query_answers object Textract query results keyed by question.
structured object Present when LLM analysis with an output schema ran.
provider string
page_count integer

See also