mirror of
https://github.com/Significant-Gravitas/AutoGPT.git
synced 2026-02-07 21:35:34 -05:00
b603ed2d2a
## Task The SmartDecisionMakerBlock is a specialized block in a graph-based system that leverages a language model (LLM) to make intelligent decisions about which tools or functions to invoke based on a user-provided prompt. It is designed to process input data, interact with a language model, and dynamically determine the appropriate tools to call from a set of available options, making it a powerful component for AI-driven workflows. ## How It Works in Practice - **Scenario:** Imagine a workflow where a user inputs, "Send an email to John about the meeting." The SmartDecisionMakerBlock is connected to tools like send_email, schedule_meeting, and search_contacts. - **Execution:** 1. The block receives the prompt and system instructions (e.g., "Choose a function to call"). 2.It identifies the available tools from the graph and constructs their signatures (e.g., send_email(recipient, subject, body)). 3. The LLM analyzes the prompt and decides to call send_email with arguments like recipient: "John", subject: "Meeting", body: "Let’s discuss...". 4. The block yields these tool-specific outputs, which can be picked up by downstream nodes to execute the email-sending action. ## Changes 🏗️ - Add the Smart Decision Maker (SDM) block. - Break circular imports in integration code.  ## Work in Progress ⚠️ **Important note this is a temporary UX for the system - UX will be addressed in a future PR** ⚠️ ### Current Status I’m currently focused on the smart decision logic. The main additions in the ongoing PR include: - Defining function signatures for OpenAI function-calling schemas based on node links and the linked blocks. - Adding tests for function signature generation. - Force all tool calls to be made via an agent. (Need to uncomment) - Restrict each tool call entry to a single node. - simplify the output emission process, to emit each parameter one at a time. - Change test to use agents and hardcode output how I think it should work to test it does actually work - Hook up openai, in a simplified way, to test the function calling (mock for testing) - Once all the above is working, use credentials system and build of llm.py ### What’s Next - Review Process ### Reviewers Phase 1 This PR is now ready for review, during the first phase of reviews I'm looking for comments on approach and logic. Out of scope: code style and organization at this stage ### Reviewers Phase 2 Once we are all happy with the approach and logic. We can open the review process to general code quality and nits, to be considered. --------- Co-authored-by: Zamil Majdy <zamil.majdy@agpt.co>