feat(llm): pass llm params directly

[Pouyanpi]

Description

langchain-community models support .bind() method

• All models inherit from langchain-core Runnable interface
• .bind() is universally available across all LangChain packages

langchain-core (contains Runnable interface with .bind())
── langchain-openai (inherits .bind())
── langchain-community (inherits .bind())
── langchain-anthropic (inherits .bind())
── langchain-* (all inherit .bind() ? )

Related Issue(s)

#1408

Checklist

• I've read the CONTRIBUTING guidelines.
• I've updated the documentation if applicable.
• I've added tests if applicable.
• @mentions of the person or team responsible for reviewing proposed changes.

[Copilot]

Pull Request Overview

This PR migrates from using context managers for LLM parameter management to passing parameters directly to the llm_call function. The change leverages LangChain's universal .bind() method to pass parameters like temperature and max_tokens directly to LLM models without temporarily modifying their state.

Key changes:

• Added llm_params parameter to llm_call function for direct parameter passing
• Replaced all with llm_params(...) context manager usage with direct parameter passing
• Updated tests to cover the new parameter passing approach

Reviewed Changes

Copilot reviewed 21 out of 21 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
nemoguardrails/actions/llm/utils.py	Added llm_params parameter to llm_call and implemented LLM binding
tests/test_tool_calling_utils.py	Added comprehensive tests for new parameter passing functionality
tests/test_llm_params_e2e.py	New end-to-end tests for LLM parameter functionality with real providers
tests/test_llm_params.py	Added migration tests comparing context manager with direct parameter approach
Various action files	Updated all LLM calls to use direct parameter passing instead of context managers
docs/user-guides/advanced/prompt-customization.md	Updated documentation example to show new parameter passing syntax

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[tgasser-nv]

Looks good, 4k LOC is too large for a single PR.

I'm a little confused about a few things:

• Why did we use a context-manager to pass a dict of LLM parameters in the first place? Normally they're used to make sure we close files/DB connections so we don't forget.
• Does a context-manager break some Langchain functionality?
• Can you add some local integration tests to make sure this works calling tools with production LLMs?

[github-actions]

Documentation preview

https://nvidia-nemo.github.io/Guardrails/review/pr-1387

[codecov-commenter]

Codecov Report

❌ Patch coverage is 61.29032% with 24 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
nemoguardrails/evaluate/evaluate_hallucination.py	9.09%	10 Missing ⚠️
nemoguardrails/actions/v2_x/generation.py	63.63%	4 Missing ⚠️
nemoguardrails/actions/llm/generation.py	86.36%	3 Missing ⚠️
nemoguardrails/library/hallucination/actions.py	0.00%	3 Missing ⚠️
nemoguardrails/evaluate/evaluate_factcheck.py	0.00%	2 Missing ⚠️
nemoguardrails/eval/check.py	0.00%	1 Missing ⚠️
nemoguardrails/evaluate/evaluate_moderation.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[trebedea]

Looks good, 4k LOC is too large for a single PR.

I'm a little confused about a few things:

• Why did we use a context-manager to pass a dict of LLM parameters in the first place? Normally they're used to make sure we close files/DB connections so we don't forget.
• Does a context-manager break some Langchain functionality?
• Can you add some local integration tests to make sure this works calling tools with production LLMs?

The context manager was a hack to provide early support for running LLM calls for the same Langchain model with different parameters. I guess the binding support in Langchain did not exist back then and @drazvan used a context manager to add different parameters to each call. This worked fine when calls where serial, but I think the mechanism was braking when we implemented parallel rails.

At that point I realized that Langchain supports binding especially for running different generations with different parameters, by binding them to a call. And I talked with @Pouyanpi about doing this cleanly now, with the binding support from Langchain.

Hope this helps.

[trebedea]

Looks good!