@daydreamsai/genai v0.3.7

@daydreamsai/genai

This package provides a DaydreamsAI SDK extension with general-purpose Generative AI (GenAI) capabilities and actions that can be easily added to any Daydream agent.

Overview

The primary goal of this package is to encapsulate reusable AI functionalities, making them independent of specific communication channels (like Discord, Telegram, Web UI, etc.). By adding this extension to an agent, the agent gains access to these core AI actions.

Communication-specific extensions (e.g., @daydreamsai/discord) are responsible for handling the specifics of their platform and preparing data in a format that these general AI actions can consume.

Current Features

Actions

• analyzeImage:

  • Description: Analyzes provided text and accompanying image attachments, then generates a relevant textual response. This action can be used to describe images, answer questions about them, or perform other vision-related tasks.
  • Input Schema:
    • text (string): The text prompt accompanying the image(s).
    • attachments (array of objects, min 1): Image attachments to analyze. Each object should contain:
      • url (string, URL): URL of the image.
      • filename (string, optional): Filename of the image.
      • contentType (string, optional): MIME type of the image (e.g., image/jpeg).
      • data (Buffer, optional): Pre-fetched image data as a Buffer. If provided by an input extension (like @daydreamsai/discord with PROCESS_ATTACHMENTS=true), this will be used directly, avoiding another fetch.
  • How it works: This action leverages the agent's main configured Language Model (e.g., a Gemini model via @ai-sdk/google). It constructs a multimodal prompt with the provided text and image data (either pre-fetched data or by fetching the url) and uses the AI SDK's generateText function to get a response.

• analyzeVideo:

  • Description: Analyzes provided text and accompanying video attachments, then generates a relevant textual response. This action can be used to describe videos, answer questions about them, summarize content, or perform other video-related tasks.
  • Input Schema:
    • text (string): The text prompt accompanying the video(s).
    • attachments (array of objects, min 1): Video attachments to analyze. Each object should contain:
      • url (string, URL): URL of the video.
      • filename (string, optional): Filename of the video.
      • contentType (string, optional): MIME type of the video (e.g., video/mp4).
      • data (Buffer, optional): Pre-fetched video data as a Buffer. If provided by an input extension, this will be used directly, avoiding another fetch.
  • How it works: Similar to analyzeImage, this action uses the agent's main configured Language Model. It constructs a multimodal prompt with the text and video data. The underlying AI SDK (ai package) and the language model provider (e.g., @ai-sdk/google) are responsible for processing the video content. Note: The exact capabilities depend on the LLM's support for video understanding.

Usage

1. Installation (assuming it's part of your monorepo/workspace): Ensure it's correctly set up in your project's package management (e.g., pnpm workspaces).

2. Add to Agent Configuration: In your agent's main setup file (e.g., where you use createDreams):

   import { createDreams } from "@daydreamsai/core";
   import { genai } from "@daydreamsai/genai";
   import { createGoogleGenerativeAI } from "@ai-sdk/google";
   
   const agent = createDreams({
     model: createGoogleGenerativeAI({
       apiKey: env.GEMINI_API_KEY,
     })("gemini-2.5-flash-preview-04-17"),
     logger: new Logger({ level: LogLevel.DEBUG }),
     extensions: [genai],
   });
   
   await agent.start();

3. Invoking Actions: Once the agent is running with this extension, the Language Model (LLM) can choose to use the analyzeImage action when appropriate based on the user's input and the action's description. For instance, if a user on Discord sends a message like "What is in this picture?" along with an image, the LLM should be able to identify that the analyzeImage action is suitable.

   The input extension (e.g., @daydreamsai/discord) is responsible for providing the text and attachments (ideally with pre-fetched data if PROCESS_ATTACHMENTS=true is set for the Discord extension) to the agent. The agent's core logic, guided by the LLM, then routes this to the analyzeImage action.

Future Possibilities

This package can be expanded to include other general AI actions, such as:

• Advanced video analysis.
• Complex document summarization or Q&A.
• Voice transcription and analysis (if an input extension provides audio data).

By keeping these actions in a dedicated genai package, they remain reusable across different agent types and communication platforms.