Discord and/or Pinecone for primary congregation places?

[daveshap]

Since we'll have swarms of agents talking to each other, it occurs to me that established chat platforms, like Discord, might be a great place to start. This will create a layer of transparency so that humans can watch and participate easily. But it also has RBAC, channels, groups, and roles.

Conversely, Pinecone might also be good as it's got intrinsic search capabilities plus metadata.

Thoughts?

[mar0der]

I think this is awesome idea

[rasalAhmCoding]

I like the Discord idea because it can be used to debug. I'm not too sure how easy Pinecone is to debug, if things goes wrong

[ben-AI-cybersec]

I've been looking through the chatdev source to see how they solved the communication problem, and pulled out a couple of key ideas:

• SOB to define roles based on an initial 'final task/end product' prompt
• define phases for development which will each have their own prompt eg:
  • ideation
  • specification
  • (tool building, QA)?
  • (coding, QA) * number of iterations
  • documentation
• possibly use their CompanyConfig json files as a starting point

A cool snippet I think we could use:

self_task_improve_prompt = """I will give you a short description of a software design requirement, 
please rewrite it into a detailed prompt that can make large language model know how to make this software better based this prompt,
the prompt should ensure LLMs build a software that can be run correctly, which is the most import part you need to consider.
remember that the revised prompt should not contain more than 200 words, 
here is the short description:\"{}\". 
If the revised prompt is revised_version_of_the_description, 
then you should return a message in a format like \"<INFO> revised_version_of_the_description\", do not return messages in other formats.""".format(task_prompt)```

[ElioDonato]

We could use Discord for everyday chat and keeping an eye on things because it's easy to use and good for quick fixes. Pinecone could be for the heavy-duty data stuff, like searching and sorting information, because that's what it's good at. This way, we use the best parts of both tools and if one has a problem, we've still got the other.