Support for logprobs in Vertex AI (ChatVertexAI) and Gemini API (ChatGoogleGenerativeAI) integrations

[persas]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

It would be valuable for LangChain to expose logprobs support when using both the Vertex AI models and the Gemini API directly.

Both Google APIs now allow retrieving token-level log probabilities (logprobs) alongside completions. This is important for downstream tasks like uncertainty estimation, calibration, forced decoding, or model-based reasoning steps.

Currently, LangChain’s interfaces for Vertex and Gemini do not surface this information, and even if the underlying APIs support it, LangChain users cannot access it natively without writing custom wrappers.

Documentation:

• https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference
• https://discuss.ai.google.dev/t/get-logprobs-at-output-token-level/54418
• http://ai.google.dev/api/generate-content

Motivation

Many advanced applications need token-level information, not just final completions. For example:

• Programmatic scoring of completions
• Implementing self-consistency sampling
• Fine-grained reranking or answer verification
• Custom prompting strategies where low-confidence tokens are treated differently

Without logprobs access, LangChain users must bypass LangChain itself and directly call the Google APIs.

I really love langchain, but this really is a stopper for me to use it.

Proposal (If applicable)

Update the internal handlers (VertexAI and ChatVertexAI, ChatGoogleGenerativeAI and related models) to:

• Accept a logprobs parameter as done in OpenAI or AzureOpenAI - https://python.langchain.com/docs/how_to/logprobs/

• Return the logprobs in the generation_info or equivalent output object

• Ideally, include an option to control how many top logprobs (top_logprobs) are returned, since both Vertex and Gemini support a top_k setting. as done in OpenAI or AzureOpenAI - https://python.langchain.com/docs/how_to/logprobs/

This could follow the style already used by other integrations (like OpenAI models via LangChain) where logprobs is optional but cleanly handled if requested.