LLM Tools #2164
LLM Tools #2164
Conversation
docs/llm_tools_invocation.md
Outdated
| function_response = await llm.async_call_tool( | ||
| self.hass, tool_input | ||
| ) | ||
| except Exception as e: # pylint: disable=broad-exception-caught |
There was a problem hiding this comment.
Limit the set of exceptions caught to just those thrown by well behaving tools. While in theory it is good to be defensive here, the reason for being more specific is to avoid masking bugs.
The two types of errors I can forsee here are:
HomeAssistantErrorraised when a known problem happens e.g. transient error from a backend, etc- An input validation exception (e.g.
vol.Invalidor some kind of equivalent validation exception).
Also, the response here of returning an error with a dictionary is not a common example in the home assistant code base. Given LLM calls happen from conversation agents perhaps an example with an IntentResponse would work? e.g. https://github.com/home-assistant/core/blob/8168aff25370e5a7afcb989d83f77815cbbf6903/homeassistant/components/openai_conversation/conversation.py#L120
docs/llm_tools_invocation.md
Outdated
|
|
||
| ```python | ||
| from homeassistant.helpers import llm | ||
| ... |
There was a problem hiding this comment.
Perhaps this example should be inline in a https://developers.home-assistant.io/docs/core/entity/conversation?_highlight=conversation#process example?
|
Please take a look at the requested changes, and use the Ready for review button when you are done, thanks 👍 |
Co-authored-by: Allen Porter <allen.porter@gmail.com>
Co-authored-by: Allen Porter <allen.porter@gmail.com>
Co-authored-by: Allen Porter <allen.porter@gmail.com>
docs/llm_tools_invocation.md
Outdated
| ) -> conversation.ConversationResult: | ||
| """Process a sentence.""" | ||
|
|
||
| while True: |
There was a problem hiding this comment.
Perhaps the example should just try at most two llm calls, though i realize it will get verbose
There was a problem hiding this comment.
You mean remove the loop?
docs/llm_tools_invocation.md
Outdated
| self.hass, tool_input | ||
| ) | ||
| except (HomeAssistantError, vol.Invalid) as e: | ||
| tool_response = {"error": type(e).__name__} |
There was a problem hiding this comment.
Make a ConversationResult here?
There was a problem hiding this comment.
Too early. Need to feed the tool result back to LLM and wait for its response.
Apply suggestions from code review
|
Superseded by #2178 |
Proposed change
Add developer documentation on how to add a new LLM tool.
Type of change
Additional information