We need to add an agentive abstraction to the copilot chat. Currently it format a large prompt and sends that to the LLM, including recent conversation history (which might error out when it gets too long..), relevant snippets from the embeddings DB, and any selected text in the editor.
We should abstract this to have multiple tools, and then an agent that can execute tools and stream out thoughts to the user until the message "resolves" to what the user likely wanted.
These tools could include:
Translation memory (drafts, queried by keywords or by similarity depending on user's input prompt)
Long-term conversation memory (previous chats, sometimes embedded or perhaps enriched with metadata)
Resources
embedded downloaded resources)
Aquifer API
...?
We need to think through some tasks or concrete use cases for development.
We need to add an agentive abstraction to the copilot chat. Currently it format a large prompt and sends that to the LLM, including recent conversation history (which might error out when it gets too long..), relevant snippets from the embeddings DB, and any selected text in the editor.
We should abstract this to have multiple tools, and then an agent that can execute tools and stream out thoughts to the user until the message "resolves" to what the user likely wanted.
These tools could include:
We need to think through some tasks or concrete use cases for development.