What can you do with Generative Actions?

Generative Code Actions enable you, for example, to:

• Extract information from images (e.g., data from a document photo, image classification, inventories generated from a photograph, etc.).
• Analyze documents (summarize content, identify risks in contracts, extract totals or items from PDF invoices, etc.).
• Classify and synthesize user messages (create summaries, detect user intentions, categorize tickets based on customer-provided free-form descriptions, etc.).
• Process responses from external APIs (use AI to analyze and simplify large JSON responses, for instance).

Flexible and multimodal: supports text, images, and documents (PDF, PNG, JPG, etc., up to 20 MB).

Differences compared to other features

• It is not an intelligent agent: It does not maintain conversations or pursue long-term objectives. It executes a function ("tool") and returns a result in a variable.
• It does not primarily rely on a Knowledge Base or Content Bases: Incorporating them is recommended only when genuinely beneficial and for very large fragments/documents. For specific rules or shorter lists, it is more efficient to include them directly in the prompt.
• It does not replace the classic Prompt or Agent functions. It's a specific tool ideal for handling "one-shot" processing tasks.

How to set up a Generative Code Action?

1. Receive the input (user, image, file, API)

It can be:

• A user message (text)
• An image (file)
• A document (PDF, etc.)
• An API response (e.g., a JSON previously processed in the flow)

2. Set up the Code Action

You should structure the Code Action by including:

• Instructions (Prompt): Clearly specify the desired task (e.g., “You are an assistant specialized in categorizing household issues. Read the options and classify the user's message into these categories…”).
• Dynamic input (Query): The information provided by the user or the content you want analyzed (text, image, file).
• Optional - Specify Format (Schema): If you want the result in JSON, explicitly define the structure ("I want you to return a JSON object with the following keys: problem, summary..."). Recommended for automated processing scenarios.
• Choose Generative Model: Select from various available models (e.g., Flash for fast and cost-efficient answers, Thinking/Reasoning for higher quality and depth).
• Assign Output Variable: Determine where to store the result for subsequent use in the flow.

Minimal structure example

Basic code example:

https://gist.github.com/hernanliendo/a899224343d967e1fa09d2437579a156

Example with multimedia, we send an image and describe what is there:

https://gist.github.com/hernanliendo/4a0f9745354863570f773b7c63a20b29

Example where it parses the user's message and extracts information in variables in a JSON:

https://gist.github.com/hernanliendo/011595222e2ebd3a1117fdbaade7bacc