Clarification on Planning Logic and Maven Central Plans #563
-
I spent some time reading through embabel’s source code. I have to say it’s a bold and high-quality project, and it aligns well with my own observations about the AI Agent Framework space. I have a few questions and I'm not sure if my understanding is correct—if you have time, I’d appreciate your clarification: Is the combination of input/output type inference and the A* algorithm essentially the underlying implementation of what you described in your blog as “uses a non-LLM AI algorithm to work out a path of actions towards a goal”? Am I missing any important details? If that's the case, when there are methods with identical inputs/outputs or overloaded methods, is cost the only criterion used for routing? Are you also planning or designing other non-LLM planning approaches at the moment? Do you have any timeline for publishing a production-ready package to Maven Central? I’m a developer based in China, and due to network conditions, it’s quite difficult for me to reliably download packages that aren’t available from Maven Central. I don’t have other questions for now—I’m still experimenting with the framework. Looking forward to your reply, and thanks in advance. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
Thank you for your comments. Yes, GOAP is the algorithm I referred to. In most cases the pre and post conditions are inferred from the flow of data types. However it's possible to add string keys for custom conditions, as in coding agent and flicker. Yes, it would be possible for >1 path to be valid due to methods with the same signatures, in which case cost would be determinant. Strictly speaking, planning would not be deterministic in that event. However, the planner would still be able to justify its decision. Planning is pluggable. We think it's important that planning is deterministic, but will consider other non-LLM algorithms if need arises. Re Maven Central, we'll take that on board and consider. |
Beta Was this translation helpful? Give feedback.
Thank you for your comments.
Yes, GOAP is the algorithm I referred to. In most cases the pre and post conditions are inferred from the flow of data types. However it's possible to add string keys for custom conditions, as in coding agent and flicker. Yes, it would be possible for >1 path to be valid due to methods with the same signatures, in which case cost would be determinant. Strictly speaking, planning would not be deterministic in that event. However, the planner would still be able to justify its decision.
Planning is pluggable. We think it's important that planning is deterministic, but will consider other non-LLM algorithms if need arises.
Re Maven Central, we'll take that on boa…