Include 3.5 Sonnet as a tool calling LLM #3904
Open
+146
−78
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎
|
evan-danswer
force-pushed
the
tools-evan-make-sonnet-tool-calling
branch
from
ea2e9aa to
0214867
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
https://linear.app/danswer/issue/DAN-1411/35-sonnet-tool-calling
Technically, all Claude 3 models are tool-calling enabled. However, we've specifically noticed an issue where in our "non explicit tool calling" flow, 3.5 Sonnet responds with a textual decision not to use the search tool, then emits some explicit search tool calls during the following .stream() call (which for NON-explicit-tool-calling LLMs is intended to emit an LLM-only answer). So, we're adding 3.5 Sonnet to our list of explicit tool callers so we can skip that initial failed step.
In addition, Claude models LOVE making sequential tool calls. We tried several prompts (including "don't call tools) that Sonnet simply ignored. One solution is to allow sequential tool calls (see tools-evan-sequential), which honestly didn't look too bad in the UI. The one thing I'm concerned about is citations; I don't think they work with repeated searches at the moment.
This PR introduces an Anthropic-specific workaround where we use the non-tool-calling message construction (i.e. add to the user message instead of adding explicit tool messages) for anthropic models.
How Has This Been Tested?
Correct Behavior through UI + Unit tests
Backporting (check the box to trigger backport action)
Note: You have to check that the action passes, otherwise resolve the conflicts manually and tag the patches.