Get material in, take deliverables out

Material goes in through the composer, and deliverables come back out of the messages. Whatever arrives becomes a draft first: a chip names the conversion Big-AGI chose, you can swap it, and only then does it send. What comes back renders in place, with a toolbar for taking it away.

Attachments: every way in

Source	How	What the model receives
Files - PDF, DOCX, images, code	menu, Ctrl + Shift + F, or drag and drop	Text, images, or both - per the chip
Web - Import from web pages	menu, Ctrl + Shift + L, or paste a link	The page as text, markdown or clean HTML
Drive - Attach Google Drive files	menu, then the Google picker	The exported form of the file
Clipboard - Auto-adapts images and text	menu, Ctrl + Shift + V, or plain paste	Text, a markdown table from copied rich text, or an image
Screen - Capture tabs, apps, and screens	menu	An image, with optional OCR text
Camera - Capture photos with optional OCR	menu	An image, with optional OCR text

Clicking a chip lists the conversions available for that file.

You know it worked when a chip appears under the composer carrying the file name and the conversion in force. The size ceiling is on the whole request, not on the file: context and limits.

PDFs and scans

The default Auto converter runs a chain, so a scan works without you choosing anything:

Step	Result
1 extract the PDF's own text	≥ 160 chars/page -> sent as text, badge NNN chars/page
2 else render the pages as images (32 at most) and OCR them	≥ 160 chars/page -> sent as OCR text
3 else attach the page images	badge not a text page

Four converters override that chain. PDF Text sends the file's own text; PDF -> OCR (for scans) reads its images; PDF -> Images attaches the pages as pictures. Pick PDF -> Text + Images when the figures matter as much as the words. Handwriting takes the same OCR path. The warning Very little text extracted via OCR - this PDF may contain mostly images/diagrams. means: send the pages as images instead.

Images

Converter	What it sends
Image (high detail) · Image (low detail) · Image (original quality)
As Image (file type)	anything that is not png, jpeg, webp or gif
Add Text (OCR)	a checkbox - recognised text alongside the picture, not instead of it
AI Caption (Text)	swaps the picture for a written description
No Image

The detail chips trade reading resolution against cost. AI Caption (Text) needs a model on the Vision job (connect an AI service).

Web content

Paste a link, or Add Web Content from the menu - up to 5 links at once.

Option	What it does
Markdown (suggested) · Text · Clean HTML	the form the page arrives in
Add Screenshot	a picture of the page too

Capture and search are served for you on the hosted app; on your own instance you supply the browsing endpoint and search keys (optional services).

Sync local files and folders

Drag a file in and Big-AGI keeps track of it. Reload it to pull in edits made elsewhere and see what changed. Save the model's version back over it, behind an Overwrite File confirmation. The conversation stays in sync with a file another program is editing. The pairing does not survive a page reload - re-pair after a refresh. Files opened or dropped from a Chromium browser pair; a pasted or downloaded file cannot. Change-watching needs a browser interface Android and iOS lack. Text files past the size cap are refused (context and limits).

Google Drive

Google format	Exported as
Doc	Markdown
Sheet	CSV of the first sheet only
Slides	PDF
Drawing	SVG

Google's own formats are exported before they travel, and the chip prefix says into what; everything else downloads unchanged. All of it travels with the request to your AI service: what leaves your browser.

What renders

A code block with the right language tag renders in place instead of as source:

You get	When the block is tagged
A diagram	mermaid, or plantuml - PlantUML is also recognised from its own source
A vector image	svg, or recognisable SVG source
A chart	chartjs, with a configuration in plain JSON - the format forbids functions, and a configuration containing them renders empty
A live web page	html, or a block that is plainly a whole page
Rendered markdown	markdown, with a Show Preview / Show Code toggle
Highlighted code	any other language tag

You know a block rendered when the source is replaced by the diagram, chart or page itself. Outside code blocks, replies also render full markdown with tables and task lists, mathematical notation, word-level differences between two versions, and inline images.

Whether the model emits those blocks at all is a property of your persona: what a persona is.

Taking it out

Most attachments and media are editable in place, and downloadable. That covers images, code, and the hosted code and files Anthropic, Gemini and OpenAI models generate. Diagrams and charts download as images; a rendered result goes back into the prompt as an image.

Button	What you get
Download PNG / Copy PNG	the diagram or chart as an image, transparent by default
Attach as image	the rendered result back into the conversation, for the model to look at
Copy Code	the block's raw source, for your editor
Fullscreen / Present	the artifact filling the screen

Hover a rendered block and that toolbar appears; on Chart.js, hold Shift for the transparent version. Copy PNG needs clipboard permission; where that is withheld, Download PNG still works.

The toolbar sits at the top-right of the block.

Running HTML code

When a model writes HTML, it renders as a real page inside a sandboxed iframe in your browser, run manually or automatically. The frame carries a Show Code / Show Web Page toggle, a Reload that re-runs it, and fullscreen. Scripts in that page run, so an interactive prototype works. A page that depends on something the frame cannot fetch comes up blank; ask for a self-contained one, then Reload.

That frame is the only execution surface: no Big-AGI backend runs code. The exception is a model's own vendor-side sandbox, a model setting (model settings).

Generating an image

Two ways to a picture. Image-capable chat models draw inline, as part of the conversation: Nano Banana, or the GPT-5 family with image generation switched on in the model's settings. Image generation engines turn a prompt straight into an image - OpenAI, Azure, OpenRouter or a local server, linked to an AI service you already configured. Quality, background transparency, format and aspect ratio are set on the engine.

To draw	Where
/draw <prompt>	in any chat
Create Picture	Message > menu - that message is the prompt

The picture arrives as an image in the message. Right-click to save it, or Attach as image to keep working on it.

Related

• Context windows, upload caps, and timeouts
• What a persona is
• What leaves your browser
• Add and manage your keys
• Set up Google Drive import on your own instance