Documents & Media

These node types extract content from documents and images, generate or manipulate PDFs, spreadsheets, and images, and convert between text and audio. Generated files are written to the workspace file store for the run when supported; otherwise content is returned inline as base64.

document

Extracts content from a PDF, image, or HTML source via PDF text extraction, OCR, HTML stripping, or LLM vision analysis.

{
  "id": "read_invoice",
  "type": "document",
  "config": {
    "operation": "extract_pdf",
    "source_type": "url",
    "source": "{{steps.fetch_invoice.output.url}}"
  }
}

Field	Type	Default	Description
operation	string	"extract_pdf"	One of extract_pdf, ocr, html_to_text, vision_analyze.
source_type	string	"url"	One of url, base64, file_path — how source is interpreted.
source	string (expression)	—	URL, base64-encoded content, or absolute file path, depending on source_type.
mime_type	string	"application/pdf"	MIME type to assume for base64/file_path sources.

operation: extract_pdf

Extracts text from a PDF using pdftotext (poppler), with pdfinfo for page count. Requires poppler-utils on the worker.

Output:

Field	Type	Description
text	string	Extracted text (layout-preserving).
pages	number	Page count.
char_count	number	Character count of text.
word_count	number	Word count of text.

operation: ocr

Runs Tesseract OCR over the fetched image/document. Requires tesseract-ocr on the worker.

Field	Type	Default	Description
language	string	"eng"	Tesseract language code passed to -l.

Output: { "text", "char_count", "word_count", "language" }.

operation: html_to_text

Strips HTML tags/whitespace from source (treated as raw HTML, not fetched).

Output: { "text", "char_count", "word_count" }.

operation: vision_analyze

Sends the fetched image/document to a multimodal LLM with a prompt and returns its response.

Field	Type	Default	Description
prompt	string (expression)	"Describe and extract all content from this document."	Instruction sent to the vision model alongside the image.
provider	string	"anthropic"	"anthropic" or "openai". Unrecognized values fall back to anthropic.
model	string	"claude-sonnet-4-6" (anthropic) / "gpt-4o" (openai)	Vision-capable model to call.
api_key	string (expression)	—	Required. API key for the chosen provider, e.g. {{secret.ANTHROPIC_API_KEY}}.

Output: { "text": "<model response>", "model": "<model used>", "provider": "anthropic" | "openai" }.

pdf

PDF manipulation utilities — page count, text extraction, merge, split, rotate, and metadata. PDF input/output is base64-encoded inline (not workspace files).

{
  "id": "merge_reports",
  "type": "pdf",
  "config": {
    "operation": "merge",
    "pdfs": ["{{steps.report_a.output.pdf_base64}}", "{{steps.report_b.output.pdf_base64}}"]
  }
}

Field	Type	Default	Description
operation	string	"count_pages"	One of count_pages, extract_text, merge, split, rotate, metadata.
pdf	string (expression, base64)	—	Source PDF, required for count_pages, extract_text, split, rotate, metadata.
pdfs	array of strings (expression, base64)	—	Two or more source PDFs, required for merge.
degrees	number	90	Rotation angle for rotate.

Operations & outputs

Operation	Output fields
count_pages	page_count
extract_text	text, char_count, word_count (parsed from raw Tj/TJ PDF operators — best-effort, not OCR)
merge	pdf_base64, size (bytes) — concatenation of pdfs in order
split	pages (array of { page, pdf_base64, size }, one per page), count
rotate	pdf_base64, degrees, size
metadata	page_count, author, creator, title, subject, keywords, producer

spreadsheet

Parses or generates CSV data. No xlsx or Google Sheets support — for spreadsheet APIs use the relevant connector.

{
  "id": "parse_export",
  "type": "spreadsheet",
  "config": {
    "operation": "parse_csv",
    "value": "{{steps.download.output.text}}",
    "has_header": true
  }
}

Field	Type	Default	Description
operation	string	"parse_csv"	"parse_csv" or "generate_csv".
delimiter	string	","	First character used as the field delimiter for both operations.

operation: parse_csv

Field	Type	Default	Description
value	string (expression)	—	Raw CSV text to parse.
has_header	boolean	true	If true, the first row is used as column names; otherwise columns are named col1, col2, ...

Output: { "rows": [{ "<header>": "<cell>", ... }], "headers": ["..."], "count": <number of rows> }.

operation: generate_csv

Field	Type	Default	Description
headers	array of strings, JSON array string, or comma-separated string (expression)	—	Column headers, written as the first row if non-empty.
rows	array (expression)	—	Each element is either an array of cell values (positional) or an object keyed by header name.

Output: { "csv": "<csv text>", "rows": <number of rows written, including header> }.

image-gen

Generates images from a text prompt via OpenAI, Stability AI, or fal.ai. Returns image URLs or base64 — not written to the workspace file store.

{
  "id": "generate_hero",
  "type": "image-gen",
  "config": {
    "provider": "openai",
    "api_key": "{{secret.OPENAI_API_KEY}}",
    "prompt": "A futuristic city at sunset, cinematic lighting",
    "size": "1024x1024"
  }
}

Field	Type	Default	Description
provider	string	"openai"	"openai", "stability", or "fal".
api_key	string (expression)	—	Required. Provider API key.
prompt	string (expression)	—	Required. Image description.
negative_prompt	string (expression)	—	Stability and fal only — content to avoid.
model	string	provider-specific (below)	Model/engine identifier.
size	string	"1024x1024"	Image dimensions, "<width>x<height>".
seed	number	-1 (random)	Stability and fal only.

provider: openai

Field	Type	Default	Description
model	string	"dall-e-3"	"dall-e-3" or "dall-e-2".
size	string	"1024x1024"	"1024x1024", "1792x1024", or "1024x1792" (dall-e-3).
quality	string	"standard"	"standard" or "hd" (dall-e-3 only).
style	string	"vivid"	"vivid" or "natural" (dall-e-3 only).
n	number	1	Number of images (1-4). dall-e-3 always generates 1.

Output: { "image_url", "image_urls": [...], "revised_prompt", "model", "provider": "openai" }.

provider: stability

Field	Type	Default	Description
model	string	"sd3"	"sd3", "sd3-turbo" (v2beta API), or an SDXL v1 engine id (e.g. "sdxl-1.0").
steps	number	25	Inference steps (SDXL v1 only).
cfg_scale	number	7	Guidance scale (SDXL v1 only).

Output: { "image_url": "", "image_urls": [], "image_base64": "<base64 PNG>", "model", "provider": "stability" } — Stability returns image bytes directly, not a URL.

provider: fal

Field	Type	Default	Description
model	string	"fal-ai/flux/schnell"	fal.ai model slug, e.g. "fal-ai/flux/dev".
steps	number	4	Inference steps.
n	number	1	Number of images.

Output: { "image_url", "image_urls": [...], "seed", "model", "provider": "fal" }.

image-helper

Image manipulation using the Go standard library (PNG/JPEG only). All inputs/outputs are base64-encoded — not workspace files.

{
  "id": "thumbnail",
  "type": "image-helper",
  "config": {
    "operation": "resize",
    "image": "{{steps.generate_hero.output.image_base64}}",
    "width": 256
  }
}

Field	Type	Default	Description
operation	string	"get_metadata"	One of get_metadata, to_base64, from_base64, resize, crop, rotate, grayscale.
image	string (expression, base64 or data URL)	—	Source image, required for get_metadata, resize, crop, rotate, grayscale. A data:...;base64, prefix is stripped automatically.
output_format	string	source format	"jpeg"/"jpg" or "png" (default) for the encoded result.
quality	number	85	JPEG quality (1-100), used when output_format is "jpeg"/"jpg".

Operations & outputs

Operation	Extra config	Output fields
get_metadata	—	width, height, format, size_bytes, aspect_ratio
to_base64	content (string, expression), format (default "png")	base64, data_url
from_base64	base64 (expression)	content (decoded bytes as string), size
resize	width, height (numbers — if only one is given, the other is computed to preserve aspect ratio)	image_base64, width, height, format
crop	x, y, width, height (numbers, default full image at 0,0)	image_base64, width, height, format
rotate	degrees (number, default 90; multiples of 90 supported, others pass through unrotated)	image_base64, width, height, degrees, format
grayscale	—	image_base64, width, height, format

tts

Converts text to speech via ElevenLabs or OpenAI. The resulting audio is written to the run's workspace files when a workspace store is configured; otherwise it's returned as base64.

{
  "id": "narrate",
  "type": "tts",
  "config": {
    "provider": "elevenlabs",
    "api_key": "{{secret.ELEVENLABS_API_KEY}}",
    "text": "{{steps.script.output.text}}",
    "format": "mp3"
  }
}

Field	Type	Default	Description
provider	string	"elevenlabs"	"elevenlabs" or "openai".
api_key	string (expression)	—	Required. Provider API key.
text	string (expression)	—	Required. Text to synthesize.
voice_id	string	"21m00Tcm4TlvDq8ikWAM" (Rachel, elevenlabs) / "alloy" (openai)	ElevenLabs voice ID, or OpenAI voice name (alloy, echo, fable, onyx, nova, shimmer).
model	string	"eleven_multilingual_v2" (elevenlabs) / "tts-1" (openai)	tts-1/tts-1-hd for OpenAI.
format	string	"mp3"	"mp3", "wav", or "opus".
speed	number	1.0	OpenAI only, 0.25-4.0.
stability	number	0.5	ElevenLabs only, 0.0-1.0 voice stability.
similarity	number	0.75	ElevenLabs only, 0.0-1.0 similarity boost.

Output:

Field	Type	Description
audio_url	string	Workspace file path (/api/v1/workspaces/runs/{exec_id}/files/audio.<ext>), present when a workspace store is configured.
audio_base64	string	Base64-encoded audio bytes, present only when audio_url is not.
format	string	"mp3", "wav", or "opus".
provider	string	Provider used.
voice_id	string	Voice used.
model	string	Model used.
char_count	number	Number of characters synthesized.

stt

Transcribes audio to text via OpenAI Whisper, Groq, NVIDIA NIM, Deepgram, or AssemblyAI.

{
  "id": "transcribe_call",
  "type": "stt",
  "config": {
    "provider": "openai",
    "api_key": "{{secret.OPENAI_API_KEY}}",
    "audio_url": "{{steps.recording.output.audio_url}}"
  }
}

Field	Type	Default	Description
provider	string	"openai"	One of openai, groq, nvidia, deepgram, assemblyai.
api_key	string (expression)	—	Required for all providers except nvidia (optional for self-hosted NIM).
audio_url	string (expression)	—	URL to fetch audio from (preferred). One of audio_url/audio_base64 is required.
audio_base64	string (expression)	—	Base64-encoded audio bytes, alternative to audio_url.
language	string	auto-detect	BCP-47 language code, e.g. "en", "es", "fr".
model	string	provider-specific (below)	Transcription model.
base_url	string	"https://integrate.api.nvidia.com/v1"	NVIDIA only — NIM endpoint (e.g. http://localhost:8000/v1 for self-hosted).
diarize	string ("true"/"false")	"false"	Deepgram and AssemblyAI only — enable speaker diarization.
punctuate	string ("true"/"false")	"true"	Deepgram only — add punctuation.

Default models per provider

Provider	Default model
openai	whisper-1
groq	whisper-large-v3-turbo
nvidia	nvidia/parakeet-ctc-1.1b (also supports nvidia/canary-1b)
deepgram	nova-2
assemblyai	best

Output:

Field	Type	Description
text	string	Full transcript.
words	array	{ word, start, end } (and confidence/speaker for Deepgram/AssemblyAI), when available.
confidence	number	Overall confidence 0-1 (Deepgram/AssemblyAI only).
duration	number	Audio duration in seconds, when available.
language	string	Detected/used language code.
provider	string	Provider used.
model	string	Model used.

Next

• AI & Agents
• Communication & Messaging
• Executions & Monitoring — Workspace files