fix: Propagate model and base URL in LLMCallException; improve error handling

[JashG]

Description

Related Issue(s)

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.

[JashG]

The NeMo Guardrails MS uses a custom chat client for main models. We often face this generic error:

nemoguardrails.llm.models.langchain_initializer.ModelInitializationError: Failed to initialize model 'meta/llama-3.3-70b-instruct' with provider 'nimchat' in 'chat' mode: Could not find LLM provider 'nimchat'

However, the relevant exception was thrown by our custom class, but gets swallowed here by the exception thrown by _init_text_completion_model.

I repositioned _init_text_completion_model to come before ModelInitializer(_init_community_chat_models, ["chat"]), to ensure the relevant exception is bubbled up. Let me know if there are any issues with this.

[greptile-apps]

Greptile Summary

• Enhanced LLMCallException to propagate model name and endpoint URL from various provider-specific attributes for better debugging
• Introduced centralized exception hierarchy (ConfigurationError, InvalidModelConfigurationError, InvalidRailsConfigurationError) with improved error messages throughout config validation
• Reordered model initializers to prioritize text completion before chat/community models (FIXME comment indicates this behavior change needs verification)

Confidence Score: 3/5

• This PR has one behavioral change that requires verification before merging
• The exception handling improvements and new exception types are well-tested and safe. However, the reordering of model initializers in langchain_initializer.py changes fallback behavior (text completion now runs before chat models), which could affect existing configurations. The FIXME comment suggests the author is uncertain about this change.
• Pay close attention to nemoguardrails/llm/models/langchain_initializer.py - verify the initializer reordering doesn't break existing chat model configurations

Important Files Changed

Filename	Overview
nemoguardrails/exceptions.py	New file introducing centralized exception hierarchy with LLMCallException enhanced to include context messages
nemoguardrails/actions/llm/utils.py	Added _raise_llm_call_exception helper to enrich errors with model name and endpoint URL from multiple provider-specific attributes
nemoguardrails/llm/models/langchain_initializer.py	Reordered model initializers to try text completion before chat/community models; FIXME comment indicates uncertainty about this change

Sequence Diagram

sequenceDiagram
    participant User
    participant LLMRails
    participant ActionDispatcher
    participant LLMUtils as "llm/utils"
    participant LangChainModel as "LangChain Model"
    participant ExceptionHandler as "Exception Handler"

    User->>LLMRails: "generate_async(prompt)"
    LLMRails->>ActionDispatcher: "execute action"
    ActionDispatcher->>LLMUtils: "llm_call(llm, prompt)"
    LLMUtils->>LangChainModel: "ainvoke(prompt)"
    
    alt Success
        LangChainModel-->>LLMUtils: "response"
        LLMUtils-->>ActionDispatcher: "response"
    else Failure
        LangChainModel-->>LLMUtils: "Exception"
        LLMUtils->>ExceptionHandler: "_raise_llm_call_exception(e, llm)"
        ExceptionHandler->>ExceptionHandler: "extract model_name from llm"
        ExceptionHandler->>ExceptionHandler: "extract endpoint_url from llm attributes"
        ExceptionHandler->>ExceptionHandler: "build context message"
        ExceptionHandler-->>LLMUtils: "raise LLMCallException(e, context)"
        LLMUtils-->>ActionDispatcher: "LLMCallException"
        ActionDispatcher-->>User: "Error with model and endpoint context"
    end

Loading

[greptile-apps]

Additional Comments (2)

1. tests/test_embeddings_only_user_messages.py, line 21 (link)

   syntax: Import needs updating to use the new location

2. tests/test_tool_calling_utils.py, line 22 (link)

   syntax: Import needs updating - LLMCallException was moved

_{8 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile
React with 👍 or 👎 to share your feedback on this new summary format}

[greptile-apps]

_{10 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile
React with 👍 or 👎 to share your feedback on this new summary format}

[greptile-apps]

style: reordering changes fallback behavior - text completion now runs before chat/community models. FIXME comment suggests uncertainty. verify this doesn't break existing configurations expecting chat completion first

[codecov]

Codecov Report

❌ Patch coverage is 90.27778% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
nemoguardrails/rails/llm/config.py	78.94%	4 Missing ⚠️
nemoguardrails/rails/llm/llmrails.py	81.81%	2 Missing ⚠️
nemoguardrails/actions/llm/utils.py	96.29%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

_{12 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile
React with 👍 or 👎 to share your feedback on this new summary format}

[greptile-apps]

style: text completion now runs before chat models - verify this doesn't break existing chat configurations that expect chat completion to be tried first. the FIXME suggests uncertainty about this ordering.