BedrockChat advertises top_k but silently drops it (not forwarded to Converse)

[kimnamu]

First, thank you for Pathway and the recent native AWS Bedrock integration (#170) — the BedrockChat Converse wrapper is genuinely useful. I ran into a small inconsistency around the top_k inference parameter and wanted to report it before proposing any change, per CONTRIBUTING.md.

Summary
BedrockChat._accepts_call_arg advertises top_k as a supported inference argument, but BedrockChat.__wrapped__ never forwards it to the Converse call, so it is silently dropped.

Where

• python/pathway/xpacks/llm/llms.py: _accepts_call_arg lists top_k in supported_args.
• In __wrapped__, the Converse inferenceConfig is built only from max_tokens / temperature / top_p / stop_sequences. top_k is never read, so it does not reach client.converse(...).
• The docstring Args: and the BEDROCK_VALID_ARGS list in tests/test_llms.py both omit top_k, so the advertised support is currently a dead entry.

Why this is the expected fix
Per the AWS Converse API reference, top_k is not part of the base InferenceConfiguration (only maxTokens, stopSequences, temperature, topP). Model-specific parameters such as top_k (e.g. Anthropic Claude) must be passed via additionalModelRequestFields. The AWS docs even show this exact case (Example 3): "additionalModelRequestFields": { "top_k": 200 } against anthropic.claude-3-sonnet.

Behavior

Call	Before	After (proposed)
chat(..., top_k=200)	top_k dropped; never sent to Converse	additionalModelRequestFields={"top_k": 200} sent
_accepts_call_arg("top_k")	True (but no effect)	True (now honored)
other args (max_tokens/temperature/top_p/stop_sequences)	unchanged	unchanged

Proposed change (minimal)
Forward top_k via additionalModelRequestFields when provided — completing the support _accepts_call_arg already declares — plus a docstring line and a credential-free regression test that captures the kwargs passed to client.converse. This is framed as filling a gap in the existing convention, not adding new behavior.

I already have this implemented and verified locally (formatting via the repo-pinned black==24 and isort pass). Following CONTRIBUTING.md, I'm filing this issue first and will open a PR once a Pathway core-team engineer acks the approach.

This contribution was prepared with the help of an AI agent (Claude Code); a human reviewed the change, reasoning, and test results before submission.