langchain - 💡(How to fix) Fix fix(openrouter): merge fragmented reasoning_details in streaming multi-turn conversations [1 comments, 2 participants]

TL;DR

Modify the merge_lists function in LangChain to correctly handle the merging of reasoning_details entries.

Guidance

• Identify the merge_lists function in LangChain and modify it to concatenate lists of reasoning_details into a single list with merged entries, rather than fragmenting them.
• Inspect the additional_kwargs dictionary after merging chunks to verify that reasoning_details is correctly merged into a single entry.
• Consider adding a custom serialization step before passing the conversation history to the OpenRouter API to ensure that reasoning_details is properly formatted.
• Review the ChatOpenRouter implementation to ensure that it correctly handles the merged reasoning_details entries.

Example

def merge_lists(lists):
    # Custom implementation to merge reasoning_details entries
    merged_list = []
    for lst in lists:
        if lst:
            if merged_list:
                # Merge the current list with the existing merged list
                merged_list[0].update(lst[0])
            else:
                merged_list = lst
    return merged_list

Notes

The provided code snippet assumes that the reasoning_details entries can be merged by updating the first entry with the subsequent entries. This may need to be adjusted based on the actual structure and requirements of the reasoning_details data.

Recommendation

Apply a workaround by modifying the merge_lists function to correctly handle the merging of reasoning_details entries, as this will allow multi-turn streaming conversations to work without errors.