Add SelfConsciousReplies component for context aware message response ignore

[AbdurrehmanSubhani]

This PR adds a SelfConsciousReplies to the DefaultComponents of an Agent.

The SelfConsciousReplies component is a PerceptionModifier which is aimed to provide the agent with the capability of choosing not reply to an incoming "say" perception based on the following parameters:

• Conversation chat history context
• Conversation Members
• Prev 6 messages to check if the agent has been back and forth (helps to reduce agent's interest in the conversation)
• Agent's personality (bio)
• Set of guidelines for the Agent to consider provided the context above

Currently getting in response to the thinking process above the following schema:

• shouldRespond: boolean
• reason: string (this is for helping debugging cases in the future to see what the thought process was and how the prompt can be improved later on)
• confidence: number (the overall confidence score the Agent thinks it has to send a reply to this message)

If the agent has been back and forth in the conversation a lot, a backAndForthPenalty is applied to the agent's reply reasoning to lower down its overall confidence for replying to this message

[vercel[bot]]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
chat	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 25, 2024 1:22pm
docs	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 25, 2024 1:22pm

[avaer]

Additional thoughts/questions:

• How much latency does gpt-4o-mini add to this pipeline?
• Can we run an OODA loop in the main action pipeline to one-shot this action decision to eliminate the latency?
• Can we make the OODA loop asynchronous from the action process to reduce latency?

[AbdurrehmanSubhani]

Additional thoughts/questions:

• How much latency does gpt-4o-mini add to this pipeline?
• Can we run an OODA loop in the main action pipeline to one-shot this action decision to eliminate the latency?
• Can we make the OODA loop asynchronous from the action process to reduce latency?

• The average latency with gpt-4o-mini is ~3 seconds
• That has been implemented in https://github.com/UpstreetAI/upstreet-core/pull/699 removing the latency however maintaining the complete cost of the main action pipeline inference. This cost is cut-off as the perception is aborted before the action pipeline inference is triggered hence saving around ~26 upstreet credits for users
• would be possible, the only downside being 2 inference costs for a chat message (~0.26 + ~26-30 upstreet credits)