🚀 RoadMap

[zzhangpurdue]

Description

This is a tentative roadmap for ModelScope-Agent, We will update it as things evolve. Some items might be discarded, others will be added later. It would be more beneficial for the community to know all the ideas and objectives.

On the other hand, answering questions, issues, and PRs is also intermixed with this file, but as in some cases one of these features would solve several issues or questions, or potentially solve something done by one or more PRs, in many cases we focus on this a bit more than on answering specific issues, questions, PRs.

Maintenance

ModelScope-Agent is currently maintained by ModelScope team, in particular @zzhangpurdue, @wenmengzhou, @lylalala, @suluyana, @Zhikaiiii, @xyliugo.

We also receive a lot of help from community contributors like @mushenL, @wangyijunlyy. They help answering questions, reviewing PRs, making sure things are in order and actually ready to be reviewed, etc.

We hope that with the growthing of Modescope-Agent the maintenance work could be done by the amazing community.

Roadmap

Now, here's the current high-level tentative roadmap:

• Allow Modelscope-Agent running in a distributed multi-agent mode. (Motivation: [multi-agent] needed #335, Code: [WIP]add multi-agent logic based on ray #319)

• Add assistant api for user who want to modify and deploy ModelScope-Agent or Agentfabric on their own local system. ([service] Add assistant api and docker file for deploying a standalone agent system #342 )

  • Add docker deployment file for environment control for assistant api
  • Allow cancel and resend user query for assistant api

• Demos for finetuning agent LLM, evaluating agent LLM and running agent llm in Modelscope-Agent([<Demos>] Demos for finetuning agent LLM, evaluating agent LLM and running agent llm in Modelscope-Agent #349)

• With the increasing of tools in system, tool retrieval module should add back. ([framework] Add tool retrieval logic #340)

• Token count module, and related api for checking the token consuming. ([framework] Add token counting logic in agent task process #341)

• Skip summarization step when observation step has already generated user needed result for some tools to avoid duplicated output, and to reduce time/token consumption on user side.

• With the increasing number of tools, the api keys for the tools should be managed in one place

• All returns in Modelscope-Agent should be in the format with message, string is not solid.

• RAG based on mulitple sources, such as embedding database, document database and knowledge-graph database, etc., could be designed by using llamaindex

• Async refactor

• Speed up ModelScope-Agent during loading time and reduce memory consumption.

  • tool lazy load logic
  • embedding indexing module not loading on no knowledge tasks.

• Rountine Works.

  • Add more tools
  • Support more LLM
  • Support more Agent methods
  • Add More Demos(Add demo about local tool, fix #360 #361 , Add demo about memory history ,fix #353 #368 )

• New Apps Exploration

  • devin on Modelscope-agent #370

Note

This issue is just to keep anyone interested updated. For individual conversations, comments, requests, questions, please create a separate discussion question. Otherwise, we would end up with a single issue to cover everything and that's not the idea.