fix(llm): make LLMReranker work with Claude Opus 5.5 on Bedrock via call_kwargs

[kimnamu]

Thanks for taking a look, and for the Bedrock wrappers this builds on (#170, #248).

LLMReranker now takes call_kwargs. The default stays {"temperature": 0}; call_kwargs={} lets the reranker run on Claude Opus 5.5 on Bedrock, which accepts only the default temperature of 1 and rejects the reranker's temperature=0.

This is a small, self-contained bug fix and non-breaking: existing callers get the same kwargs as before. CHANGELOG.md has an entry under [Unreleased].

Related issue(s) / prior discussion

Fixes #277

Context

LLMReranker(llm, ...), live us-east-1	Before (argument did not exist)	After
BedrockChat("us.anthropic.claude-opus-5-5"), call_kwargs={}	❌ ValidationException: temperature is deprecated, no way to drop it	✅ [{'score': 5.0}]
BedrockChat("global.anthropic.claude-opus-5-5"), call_kwargs={}	❌ same	✅ [{'score': 5.0}]
LiteLLMChat("bedrock/us.anthropic.claude-opus-5-5"), call_kwargs={}	❌ UnsupportedParamsError: ... does not support temperature=0	✅ [{'score': 5.0}]
BedrockChat("us.anthropic.claude-opus-5"), call_kwargs={}	❌ same ValidationException	✅ [{'score': 5.0}]
any llm, no call_kwargs	sends temperature=0	✅ unchanged, still sends temperature=0
Sonnet 4.5 / Opus 4.5 with the default	✅ [{'score': 5.0}]	✅ unchanged
other LLMReranker arguments, return type	✅	✅ unchanged

The hardcoded dict(temperature=0) in __call__ (rerankers.py:119-127) becomes self.call_kwargs, named after HFPipelineChat(call_kwargs=...). Plus one test and the CHANGELOG line. The default is kept rather than picked per model, since that would need a model list in the chat wrappers.

How has this been tested?

Run on the pathway 0.33.0 wheel with the changed files copied over it (the llm xpack files are byte-identical to main):

$ pytest -q xpacks/llm/tests/test_rerankers.py xpacks/llm/tests/test_llms.py xpacks/llm/tests/test_rag.py
77 passed, 6 xfailed, 10 xpassed, 1 warning      # main: 75 passed, 6 xfailed, 10 xpassed

black 24, isort and flake8 pass on the changed files; mypy reports nothing in rerankers.py.

New test against the old source, and the live runs

With only rerankers.py reverted:

E       TypeError: LLMReranker.__init__() got an unexpected keyword argument 'call_kwargs'
FAILED .../test_rerankers.py::test_llm_reranker_call_kwargs[None-expected_kwargs0]
FAILED .../test_rerankers.py::test_llm_reranker_call_kwargs[call_kwargs1-expected_kwargs1]
2 failed, 2 passed, 1 warning

With the change: 4 passed, 1 warning.

Live, same wheel with this change applied:

us.anthropic.claude-opus-5-5 default -> ERR ValidationException: ... `temperature` is deprecated for this model
us.anthropic.claude-opus-5-5 call_kwargs={} -> [{'score': 5.0}]
global.anthropic.claude-opus-5-5 default -> ERR ValidationException: ... `temperature` is deprecated for this model
global.anthropic.claude-opus-5-5 call_kwargs={} -> [{'score': 5.0}]
bedrock/us.anthropic.claude-opus-5-5 default -> ERR UnsupportedParamsError: ... does not support temperature=0. Only temperature=1 is supported.
bedrock/us.anthropic.claude-opus-5-5 call_kwargs={} -> [{'score': 5.0}]
us.anthropic.claude-opus-5 default -> ERR ValidationException: ... `temperature` is deprecated for this model
us.anthropic.claude-opus-5 call_kwargs={} -> [{'score': 5.0}]
us.anthropic.claude-sonnet-4-5-20250929-v1:0 default -> [{'score': 5.0}]
us.anthropic.claude-opus-4-5-20251101-v1:0 default -> [{'score': 5.0}]

Eight documents through Opus 5.5 with call_kwargs={}, both profiles: rows: 8 scores: [5.0, 4.0, 4.0, 2.0, 1.0, 1.0, 1.0, 1.0].

Why the default stays {"temperature": 0}: in one of two runs, Sonnet 4.5 with call_kwargs={} wrapped its answer in a json code fence and the parse failed (ValueError: Expected a json response, got followed by the fence); with temperature=0 it scored in both runs.

Where the docstring's "accept only the default temperature, such as Claude Opus 4.7 and newer" comes from, Bedrock Converse with inferenceConfig.temperature:

model                              0      0.5    1      1.0
us.anthropic.claude-opus-4-6-v1    OK     OK     OK     OK
us.anthropic.claude-opus-4-7       ERR    ERR    OK     OK
us.anthropic.claude-opus-4-8       ERR    ERR    OK     OK
us.anthropic.claude-opus-5         ERR    ERR    OK     OK
us.anthropic.claude-opus-5-5       ERR    ERR    OK     OK
global.anthropic.claude-opus-5-5   ERR    ERR    OK     OK
ERR = ValidationException: `temperature` is deprecated for this model.

References: Claude Opus 5.5 announcement, migration guide, models overview

[CLAassistant]

All committers have signed the CLA.

[zxqfd555]

Hi @kimnamu,
Thank you for your contribution! It solves a problem, and I agree that we should apply the fix. In order to proceed, could you please sign the CLA?

[kimnamu]

Hi @zxqfd555 , I've done. Thank you for supporting great open source :)

[zxqfd555]

Thank you, merged!