Greptile Overview

Greptile Summary

Added support for extracting reasoning traces and tool calls from LangChain 1.x content_blocks format, with automatic fallback to legacy additional_kwargs and tool_calls attribute for backward compatibility.

Key changes:

• Refactored _store_reasoning_traces() to try content_blocks first, then fall back to additional_kwargs
• Refactored _store_tool_calls() to try content_blocks first, then fall back to tool_calls attribute
• Split extraction logic into separate helper functions for better modularity
• Added comprehensive test coverage with 28 new tests covering edge cases and fallback behavior

Testing:

• Tests cover both extraction methods, fallback behavior, multiple content blocks, and edge cases
• Tests verify proper prioritization (content_blocks over legacy format)
• Tests include both MockResponse and real LangChain AIMessage objects

Confidence Score: 4/5

• This PR is safe to merge with minor risk from a KeyError edge case
• Score reflects solid implementation with comprehensive tests and backward compatibility, but deducted one point for the KeyError risk in reasoning extraction when blocks have type "reasoning" but missing the key
• Pay close attention to nemoguardrails/actions/llm/utils.py line 252 for the KeyError issue

Important Files Changed

File Analysis

Sequence Diagram

sequenceDiagram
    participant LLM as LLM Call
    participant Store as _store_reasoning_traces/_store_tool_calls
    participant ExtractCB as _extract_from_content_blocks
    participant ExtractLegacy as _extract_from_legacy
    participant CtxVar as Context Variable

    LLM->>Store: response object
    Store->>ExtractCB: try content_blocks first
    alt content_blocks exists
        ExtractCB->>ExtractCB: iterate blocks
        ExtractCB->>ExtractCB: filter by type
        ExtractCB-->>Store: return extracted data
        Store->>CtxVar: set value
    else no content_blocks
        ExtractCB-->>Store: return None
        Store->>ExtractLegacy: fallback to legacy format
        alt legacy format exists
            ExtractLegacy-->>Store: return extracted data
            Store->>CtxVar: set value
        else no legacy format
            ExtractLegacy-->>Store: return None
            Store->>CtxVar: set None or original value
        end
    end

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

Greptile Overview

Greptile Summary

Refactored LLM response processing to support LangChain 1.x content blocks format while maintaining backward compatibility with the legacy format.

Key Changes:

• Extracted reasoning traces from content_blocks with type: "reasoning" (LangChain 1.x standard)
• Extracted tool calls from content_blocks with type: "tool_call" (LangChain 1.x standard)
• Implemented automatic fallback to additional_kwargs and tool_calls attribute for providers using legacy format
• Refactored extraction logic into separate functions for better modularity and testability
• Added comprehensive test coverage for both new content blocks format and legacy format

Architecture:
The implementation follows a try-first-then-fallback pattern: it first attempts to extract data from the newer content_blocks format, and only if that fails, falls back to the legacy format. This ensures compatibility with both old and new LangChain versions.

Confidence Score: 4/5

• Safe to merge with one minor defensive check already flagged in previous comments
• The refactoring is well-structured with good separation of concerns and comprehensive test coverage. The fallback mechanism ensures backward compatibility. One minor issue (missing key existence check in line 252) was already identified in previous comments and should be addressed before merging.
• No files require special attention beyond addressing the previously flagged issue in nemoguardrails/actions/llm/utils.py:252

Important Files Changed

File Analysis

Sequence Diagram

sequenceDiagram
    participant LLM as LLM Call
    participant Store as _store_reasoning_traces / _store_tool_calls
    participant CB as Content Blocks Extractor
    participant AK as Additional Kwargs Extractor
    participant CV as Context Variable

    LLM->>Store: response object
    
    alt Reasoning Extraction
        Store->>CB: _extract_reasoning_from_content_blocks
        CB->>CB: Check hasattr(response, "content_blocks")
        alt content_blocks exists
            CB->>CB: Iterate blocks
            CB->>CB: Find type="reasoning" with "reasoning" key
            CB-->>Store: Return reasoning string or None
        else no content_blocks
            CB-->>Store: Return None
        end
        
        alt No reasoning from content_blocks
            Store->>AK: _extract_reasoning_from_additional_kwargs
            AK->>AK: Check hasattr(response, "additional_kwargs")
            AK->>AK: Get "reasoning_content" from dict
            AK-->>Store: Return reasoning or None
        end
        
        Store->>CV: reasoning_trace_var.set(reasoning_content)
    end
    
    alt Tool Calls Extraction
        Store->>CB: _extract_tool_calls_from_content_blocks
        CB->>CB: Check hasattr(response, "content_blocks")
        alt content_blocks exists
            CB->>CB: Collect all type="tool_call" blocks
            CB-->>Store: Return list or None
        else no content_blocks
            CB-->>Store: Return None
        end
        
        alt No tool_calls from content_blocks
            Store->>AK: _extract_tool_calls_from_attribute
            AK->>AK: getattr(response, "tool_calls", None)
            AK-->>Store: Return tool_calls or None
        end
        
        Store->>CV: tool_calls_var.set(tool_calls)
    end

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

Looks good, please address the comments on the tests before merging. These auto-generated tests tend to bloat the line-count of tests by checking individual contents of dicts.

These tests basically store and get the same data back in each test via different permutations. Recommend re-writing test to define a dict which is both stored and checked in the assert