fix(prompt): clarify tool-calling instructions to prevent fabricated Observation in lightweight LLMs #3468
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Outdated suggestions cannot be applied.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.
Summary
Lightweight LLMs sometimes generate
Observation
text without invoking a tool. This PR clarifies the system prompt to distinguish tool-use vs. final-answer cases, reducing hallucinated tool results.Problem
The old prompt mixed
Observation
into the output template without clear separation. Smaller models often filled it in prematurely, fabricating results even when no tool was called (see #3154).Solution
Update
en.json
prompt to:Thought
,Action
,Action Input
, then blankObservation
until tool returns.Thought
,Final Answer
only.Final Answer
when calling a tool; no non-emptyObservation
unless it is the actual tool output.This removes ambiguity and enforces correct sequencing.