fix: process Vertex AI response properly for usage

[t-miyak]

Fixed to send usage of VertexAI correctly.

I encountered the following error while using VertexAI with LangChain.

10 validation errors for UpdateGenerationBody
usageDetails -> prompt_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> candidates_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> cache_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> prompt_tokens
  field required (type=value_error.missing)
usageDetails -> completion_tokens
  field required (type=value_error.missing)
usageDetails -> total_tokens
  field required (type=value_error.missing)
usageDetails -> prompt_tokens_details -> modality
  value is not a valid integer (type=type_error.integer)
usageDetails -> input_tokens
  field required (type=value_error.missing)
usageDetails -> output_tokens
  field required (type=value_error.missing)
usageDetails -> total_tokens
  field required (type=value_error.missing)
Traceback (most recent call last):
  File "/app/.venv/lib/python3.13/site-packages/langfuse/client.py", line 2733, in update
    request = UpdateGenerationBody(**generation_body)
  File "/app/.venv/lib/python3.13/site-packages/pydantic/v1/main.py", line 347, in __init__
    raise validation_error
pydantic.v1.error_wrappers.ValidationError: 10 validation errors for UpdateGenerationBody
usageDetails -> prompt_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> candidates_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> cache_tokens_details
  value is not a valid integer (type=type_error.integer)
usageDetails -> prompt_tokens
  field required (type=value_error.missing)
usageDetails -> completion_tokens
  field required (type=value_error.missing)
usageDetails -> total_tokens
  field required (type=value_error.missing)
usageDetails -> prompt_tokens_details -> modality
  value is not a valid integer (type=type_error.integer)
usageDetails -> input_tokens
  field required (type=value_error.missing)
usageDetails -> output_tokens
  field required (type=value_error.missing)
usageDetails -> total_tokens
  field required (type=value_error.missing)

This seems to occur when using the newer version of langchain-google-vertexai, and was caused by the UsageMetadata in VertexAI API responses containing fields that are not integers.

Example: json of response (`langchain_core.outputs.LLMResult`)

{
  "generations": [
    [
      {
        "text": "",
        "generation_info": {
          "is_blocked": false,
          "safety_ratings": [],
          "usage_metadata": {
            "prompt_token_count": 4464,
            "candidates_token_count": 18,
            "total_token_count": 4482,
            "prompt_tokens_details": [
              {
                "modality": 1,
                "token_count": 4464
              }
            ],
            "candidates_tokens_details": [
              {
                "modality": 1,
                "token_count": 18
              }
            ],
            "cached_content_token_count": 0,
            "cache_tokens_details": []
          },
          "finish_reason": "STOP",
          "avg_logprobs": -0.000507740666055017
        },
        "type": "ChatGeneration",
        "message": {
          "content": "",
          "additional_kwargs": {
            "function_call": {
              "name": "xxxxx",
              "arguments": "{\"aaaaa\": \"bbbbb\"}"
            }
          },
          "response_metadata": {
            "is_blocked": false,
            "safety_ratings": [],
            "usage_metadata": {
              "prompt_token_count": 4464,
              "candidates_token_count": 18,
              "total_token_count": 4482,
              "prompt_tokens_details": [
                {
                  "modality": 1,
                  "token_count": 4464
                }
              ],
              "candidates_tokens_details": [
                {
                  "modality": 1,
                  "token_count": 18
                }
              ],
              "cached_content_token_count": 0,
              "cache_tokens_details": []
            },
            "finish_reason": "STOP",
            "avg_logprobs": -0.000507740666055017,
            "model_name": "gemini-2.0-flash-001"
          },
          "type": "ai",
          "name": null,
          "id": "run-3a8c2d0d-9f67-4e04-a58e-6dafeaeb9202-0"
        }
      }
    ]
  ],
  "llm_output": null,
  "run": null,
  "type": "LLMResult"
}

Since these fields are not necessary for Usage, I added processing to remove them.

---

my dependencies

langchain==0.3.24
langfuse==2.60.3
langchain-google-vertexai==2.0.21
google-cloud-aiplatform==1.90.0
google-cloud-core==2.4.3

---

Important

Fixes VertexAI response processing in LangChain integration by removing non-integer fields from usage data.

• Behavior:
  • Fixes processing of VertexAI responses in langfuse/callback/langchain.py by removing non-integer fields from usage data.
  • Handles fields prompt_tokens_details, candidates_tokens_details, and cache_tokens_details.
• Dependencies:
  • Removes langchain-google-vertexai dependency from pyproject.toml.

^{This description was created by for 7270a42. You can customize this summary. It will automatically update as commits are pushed.}

Greptile Summary

Disclaimer: Experimental PR review

This PR addresses validation errors when processing Vertex AI responses in the Langfuse Python SDK by modifying how usage metadata is handled.

• Modified UpdateGenerationBody to properly validate Vertex AI's non-integer usage metadata fields
• Updated langchain.py to filter out non-standard fields from Vertex AI responses before validation
• Removed langchain-google-vertexai from dev dependencies to avoid version conflicts
• Ensures proper handling of usage fields like prompt_tokens_details and candidates_tokens_details

The changes focus on maintaining compatibility with newer versions of langchain-google-vertexai while ensuring proper validation of usage data structures.

_{💡 (1/5) You can manually trigger the bot by mentioning @greptileai in a comment!}

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

LGTM

_{2 file(s) reviewed, no comment(s)
Edit PR Review Bot Settings | Greptile}

[greptile-apps]

LGTM

_{1 file(s) reviewed, no comment(s)
Edit PR Review Bot Settings | Greptile}

[t-miyak]

Sorry, I'm not used to poerty and updated lock file, but now I restored it.

[t-miyak]

I've fixed a test for VertexAI and tested if it works.

[hassiebp]

Thanks a lot for your contribution! Closing this in favor of #1181