Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.cogniagent.ai/llms.txt

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

The Ask AI node can do more than read and write text. Send it an image, a voice memo, a PDF, or a video and the AI works with the file directly. It can also generate files back — pictures, audio clips, or charts produced by code — and pass them to later steps in your workflow. This guide explains which files work with which AI models, how to pass files in, and how to capture files coming out.

​
How file inputs work

Every AI model in CogniAgent lists what kinds of files it understands. When you attach a file to an Ask AI step:
  1. CogniAgent looks at the file’s type (image, audio, video, PDF, text, other).
  2. It checks whether the model you picked accepts that type natively.
  3. If yes, the file is sent to the model as-is, with no information loss.
  4. If no, CogniAgent falls back to extracting text from the file so the model at least sees a written description of the contents.
You don’t configure any of this. Pick a model that supports the file type you’re working with, attach the file, and CogniAgent does the rest.
The model picker shows what each model can read. Look for the small badges next to the model name — they tell you whether the model accepts images, audio, video, or files.

​
What you can send in

Best for: photos, screenshots, charts, diagrams, scanned documentsModels with vision support can describe what’s in an image, extract text from a screenshot, read a chart, identify objects, or compare two images side by side.Common models that accept images: GPT-4o, Claude Sonnet, Gemini Flash, Gemini Pro.The file goes through unchanged — the AI sees the actual pixels, not a text description. That means it can pick up details like layout, color, handwriting, and visual relationships that text extraction would lose.
Best for: voice memos, call recordings, meeting audio, music samplesAudio-capable models can transcribe speech, summarize what was said, identify speakers, classify the tone, or analyze music.Common models that accept audio: Gemini Flash and Pro (including 3.x), GPT-4o audio preview.Supported formats include WAV, MP3, OGG, FLAC, WebM, M4A, Opus, and AAC. CogniAgent picks the right format automatically based on the file.
Best for: short clips, screen recordings, product demos, security footageVideo-capable models can describe what happens in a clip, find a specific moment, transcribe spoken content, or summarize a recording.Common models that accept video: Gemini Pro and Flash 2.x / 3.x.Keep clips short — most models cap at a few minutes per video and longer files cost more to process.
Best for: contracts, invoices, reports, scanned forms, research papersModels with file support read PDFs as documents — they understand page layout, embedded images, and formatting, not just the raw text.Common models that accept PDFs: Gemini, GPT-4 file mode, Claude.For models that don’t accept PDFs natively, CogniAgent automatically extracts the text using its document converter so you still get a usable result.
Best for: CSV, JSON, Markdown, XML, YAML, plain logsText-based files are inlined into the prompt directly. Every model can read them, so you don’t need a specialized vision or audio model.Use this for: data extraction from CSVs, parsing API responses, summarizing logs, working with config files.

​
Attaching files to an Ask AI step

You attach files the same way regardless of type. There are two modes:
Upload a fixed file from your computer. The same file is sent every time the workflow runs.Use this for reference materials — a product catalog PDF, a brand guidelines document, an example image.
LLM attachments field

​
When the model can’t read your file

If you attach a file type the selected model doesn’t support natively, CogniAgent falls back to text extraction: This means a PDF on a vision-only model still works (you get extracted text instead of layout-aware understanding), but an audio file on a text-only model only works if the audio converter can transcribe it locally.
Text extraction loses information. A scanned PDF becomes plain text — page layout, images, and handwriting are lost. An audio file becomes a transcript — tone and pauses are lost. For best results, pick a model that natively supports your file type.

​
Generating files with AI

Some AI models don’t just read files — they create them. Ask AI captures generated files automatically and exposes them on the node’s files output so downstream steps can use them.
Image-generation models like Gemini Flash Image produce one or more images in response to your prompt. You don’t need any special configuration — just pick the model and describe what you want.
System prompt:
You are a product photographer creating clean marketing visuals.

User message:
Generate a flat-lay photo of a leather-bound notebook, a brass pen, and a cup of coffee on a wooden desk. Soft natural light. Top-down angle.
The generated image lands in {{llm_1.files}} and can be attached to an email, posted to a Slack channel, or sent to another AI step.
Audio-generation models like Lyria produce music or speech. CogniAgent automatically streams the audio response, saves it as a file, and exposes it on {{llm_1.files}} the same way as images.Use cases: generate a voice greeting for a phone agent, create a custom ringtone, produce a short music clip for a podcast intro.
Audio generation requires streaming under the hood — CogniAgent handles this automatically when you pick an audio-output model. You don’t need to change any settings.
When you enable the Execute Python Code capability on an Ask AI node, the AI can write code that produces files — CSVs, charts, processed images, transformed documents. Any file the code writes to the sandbox is uploaded automatically and surfaced on the node’s output.Example: ask the AI to analyze sales data from a CSV and chart the results.
User message:
Read sales.csv (the file attached above). Calculate monthly revenue.
Plot a line chart and save it as `revenue.png`. Use output_data() to
return the monthly totals as JSON.
Both revenue.png and the JSON output appear in {{llm_1.files}} and {{llm_1.response}} respectively.

​
Using generated files in later steps

Files coming out of an Ask AI node are CogniAgent file descriptors — they carry the file name, size, and type, and you reference them like any other variable.
1

Pass to another Ask AI step

Use {{llm_1.files}} in the attachments field of a follow-up Ask AI step. Use this to chain steps — generate, then analyze, then describe.
2

Send via email

Most email channels and integrations accept file descriptors as attachments. Drop {{llm_1.files}} into the attachments field and the file rides along with the message.
3

Save as a variable

Use Update Variable to store a file descriptor for later in the workflow. Useful when the same file is needed in several steps.
4

Let users download it

Generated files appear in the execution history with a download button. CogniAgent mints a fresh secure link each time you click — no stale URLs.

​
Tips and gotchas

Pick the right model for the job. A vision model that costs more is cheaper than a “free” text model that fails to read your image and forces you to retry.
Mix file types in one prompt. Most multimodal models accept several attachments at once. You can hand the AI a PDF plus a chart image and ask it to cross-reference them in one step.
Generated audio and image files can be large. A high-resolution image can be several megabytes, and a minute of audio can hit ten. Keep this in mind if you plan to attach generated files to emails or store many of them per workflow run.
Model lists change frequently. The badges shown next to each model in the picker are the source of truth for what that model can read and produce — they’re synced from the upstream provider and update as new models ship.

Ask AI Node

Full reference for configuring the Ask AI step.

Read File

Pull a file into your workflow from another node or step.

Execute Code

Run Python in a sandbox — files it writes are surfaced as outputs.

Call AI Agent

For multi-step AI tasks, delegate to a dedicated agent application.