New in Endora Commerce: a built-in CRM with sales opportunities See it →
Endora Commerce

Google Gemini integration

The admin assistant in Endora Commerce can run on Gemini models: you supply your own Gemini API key and a model ID, and the platform talks to Google directly, with no intermediary.

Which Google service the platform calls

The integration uses the Gemini API, the service whose keys are created in Google AI Studio. The platform sends requests straight to the generateContent method and authenticates with the API key in a request header. There is no SDK and no third-party relay in between.

This is worth knowing before you talk to your IT team: it is not a Vertex AI connection. If your organisation uses Gemini models only through Vertex AI and service accounts, the assistant will need a separate Gemini API key.

The admin assistant itself works the same as on other models: an instruction in the command palette, a plan of changes, your confirmation, execution.

What to prepare on the Google side

  • A Google Cloud project. Every Gemini API key belongs to a project, which handles billing and permissions. AI Studio creates a default project for new users; existing projects have to be imported into AI Studio.
  • An API key. You create it in AI Studio. Google recommends restricting keys, for example limiting a key to the Generative Language API only. See the Gemini API key documentation.
  • The paid tier. Under the Gemini API terms, only paid services may be used when making a solution available to users in the European Economic Area, Switzerland and the United Kingdom. For paid services Google states that it does not use prompts or responses to improve its products. For a company based in Poland this means, in practice, enabling billing on the project before go-live.
  • A model ID. You type it in by hand, exactly as the Gemini API documentation lists it. The model has to support function calling, because the assistant’s whole way of working depends on it.

Setup in the admin panel

  1. In the credentials module, create a new configuration of type LLM with Gemini as the provider. Enter a name, a stable code, the API key and the model.
  2. In the assistant’s settings, select that configuration as the LLM credentials and enable the assistant.
  3. On the roles screen, grant the assistant permission to the operators who should use it.

The key is encrypted in the database, cannot be read back in the panel or through the API once saved, and is redacted in the audit log. A configuration change applies from the next instruction, with no restart.

A configuration’s type and provider are locked after saving, so moving from Gemini to GPT or Claude means creating a second configuration and repointing the setting, not editing the existing one. Key rotation, on the other hand, is one change in one place.

What goes to Google and what stays in the platform

The following is sent to the Gemini API:

  • the text of the instruction typed by the operator,
  • the assistant’s system instruction,
  • descriptions of the tools that operator has permission to use,
  • results of the read-only searches the assistant uses to turn names from the instruction into concrete records, for example matching products, categories and warehouses.

The model gets no access to the database or to arbitrary endpoints. All it can do is pick a tool from a closed catalogue registered in the platform’s code. Changes are not executed while the instruction is being interpreted: they go into a plan, and everything you see on the plan card (names, record counts, current values) is computed by the server, not by the model. Each execution writes the instruction text, the provider and the model to the audit log.

The adapter absorbs the ways the Gemini API differs from other providers: Gemini returns no tool-call IDs and accepts a narrower function parameter schema. A tool added to the catalogue by a custom module therefore needs no separate Gemini version.

Limits on the Google side

Gemini API rate limits are applied per project, not per key, and are measured in three dimensions: requests per minute, tokens per minute and requests per day. Exceeding any one of them results in a 429 error. In practice:

  • If the same Google Cloud project serves other applications, they share limits with the assistant. A dedicated project for the platform keeps both limits and billing simpler.
  • Experimental and preview models have tighter limits, so a stable model is the better choice in production.

When to pick Gemini and when another provider

Gemini is the natural choice when the company already has a project, billing and procurement procedures in Google Cloud, or wants a second provider alongside its current one. What the assistant can do does not depend on the provider, because the tool catalogue belongs to the platform, not to the model.

If you need Vertex AI with service accounts, this integration does not cover it. The provider list for LLM configurations also includes DeepSeek.

Let’s find out what we can connect at your end

Book a free consultation. We will walk through your ERP, warehouse and sales processes, and show you what the integration path looks like step by step.

Michał Zabielski, founder and CEO of Endora

Michał Zabielski

Founder and CEO, Endora

We reply within one business day.