[Enhancement]: Implement Google Vertex AI 'Web Grounding for Enterprise' for Data Privacy Compliance

[marlonka]

What features would you like to see added?

Problem Description

The current implementation of Google Search Grounding for Vertex AI models in LibreChat presents a significant data privacy concern for enterprise users, especially those operating under strict regulations like GDPR in the EU.

https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance

According to Google's documentation for Grounding with Google Search:

As outlined in Section 19 "Generative AI Services: Grounding with Google Search" of the Service Specific Terms, Google stores prompts and contextual information that customers may provide, and generated output for thirty (30) days... There is no way to disable the storage of this information if you use Grounding with Google Search.

This 30-day data retention policy, with no option to opt-out, is a non-starter for many businesses and makes the feature unusable in a production/enterprise environment. For our organization, this is a blocker that prevents us from updating to the latest versions of LibreChat.

Proposed Solution

Google provides a direct solution to this problem with Web Grounding for Enterprise. This service is specifically designed for compliance and data privacy.

According to the documentation for Web Grounding for Enterprise:

The indexed content is a subset of what's available on Google Search and suitable for customers in highly-regulated industries, such as finance, healthcare, and the public sector.

Key benefits of the Enterprise version include:

• No logging of customer data.
• ML processing in the US or European multi-regions.
• VPC Service Controls support.

This would resolve the data privacy blocker and make the grounding feature viable for enterprise use.

More details

Technical Implementation Details

https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/web-grounding-enterprise
The implementation appears to be a straightforward change in the API call to the Vertex AI Gemini model. Instead of the standard grounding tool, the enterprise_web_search tool should be provided.

Here is the example code provided by Google:

# The google-genai SDK needs to be configured to use Vertex AI
# export GOOGLE_CLOUD_PROJECT=...
# export GOOGLE_CLOUD_LOCATION=global
# export GOOGLE_GENAI_USE_VERTEXAI=True

from google import genai
from google.genai.types import (
    EnterpriseWebSearch,
    GenerateContentConfig,
    HttpOptions,
    Tool,
)

# Note: The docs specify api_version="v1" for this feature
client = genai.Client(http_options=HttpOptions(api_version="v1"))

response = client.models.generate_content(
    model="gemini-2.5-flash", # or other compatible models
    contents="When is the next total solar eclipse in the United States?",
    config=GenerateContentConfig(
        tools=[
            # Use Enterprise Web Search Tool
            Tool(enterprise_web_search=EnterpriseWebSearch())
        ],
    ),
)

print(response.text)

Desired Outcome / Acceptance Criteria

To enforce compliance across an entire LibreChat instance, this feature should be configured by a system administrator, not an end-user. The ideal location for this setting is in the librechat.yaml file (or .env if more appropriate for the architecture).

A new configuration key should be added under the endpoints.google section in librechat.yaml to allow administrators to select the grounding type.

Implementation Logic:

1. Introduce groundingType: Add a groundingType key to the Google endpoint configuration.
2. Define Values:
   • enterprise: When set, the backend must use the Tool(enterprise_web_search=EnterpriseWebSearch()) for all grounding requests to Google models.
   • standard: When set, the backend should use the current, standard Google Search grounding method.
3. Default Behavior: If the groundingType key is omitted or left empty, it should default to standard to ensure backward compatibility and not break existing installations.
4. No UI Change: This should be a backend-only configuration. No changes are needed in the end-user facing UI.

If thats not possible, the following would be an alternative:

For example, in the Google endpoint configuration UI, a dropdown or radio button could be added:

Grounding Type:

• Standard Google Search (30-day data logging)
• Enterprise Web Search (No data logging)

This would empower administrators to select the appropriate, compliant option for their organization's needs.

Thank you for considering this critical feature for the enterprise-readiness of LibreChat.

Which components are impacted by your request?

No response

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

I'm happy to explore this but seems odd on Google's part to offer LLM services with enterprise assurances, while also providing a feature for them that is not. It may be a simple toggle so will take the time this week.

[marlonka]

Yeah, I dont get it as well, but its more pricey than the normal option because they now enterprises will pay that to remain secure as possible. But thanks for fixing that issue then. If possible I would really love if you could make another release soon after implementing that feature. If possible, configuration via env or yaml for all users would be the safe option, otherwise users will not know that they potentially leak data.

[marlonka]

@danny-avila Could you have a look to implement that feature for 0.8.0-rc2 maybe? Would be a huge win for us

[raphaelgurtner]

A quick solution would be much appreciated as without the possibility to configure either enterprise grounding or to deactivate grounding altogether this prevents the use of gemini models in LibreChat for us... Would you @danny-avila fast track a review / pull request for either of these options for 0.8.0?

[marlonka]

@danny-avila is there anything we can help you with for implementing this feature? I am not a coder so it will probably not work If I code it

[aitp-john]

We have done more research on this and would be interested in taking on this feature in a way that it could be useable with any model/agent. It's more of a black box approach since you don't have control over how Google does it's AI summaries. The current set up has a lot of controls but requires a very specific set up to be secure (local SearXNG and Firecrawl and maybe a Cohere agreement change).

My first thought it to start by implementing this as a Tool that you can add to any agent. It would also need a small chat UI web search button handler to invoke the tool from the chat pane. Starting as a tool would be faster and have less impact on the core project.

@danny-avila thoughts?