API integration for Gemini
We build your Gemini connector
We integrate Gemini to extract, classify and call your tools, with EU residency via Vertex when the contract requires it. The Studio API is not Europe.
- Senior product team
- LLM connectors in production
- from scoping to monitoring
What does the Gemini API do and why integrate it in a European product?
Gemini is Google's family of AI models, available via API for text generation, image analysis, PDF processing and function calling in your application. You integrate it to automate document reading, email classification, knowledge base question-answering, or triggering business actions from natural language input. For projects that require data to remain in Europe, Gemini is available through Vertex AI in European regions, satisfying data residency requirements for regulated sectors.
What our clients build on the Gemini API
Invoice PDF extraction into the ERP
Vertex eu, minimised prompt. The full customer file does not leave. Quality measured in French, not assumed.
Support assistant with ticket tools
The model queries internal docs and opens a ticket. It does not invent stock: it calls your function.
Support email classification
Urgency, theme, routing. A business label, not a free reply pasted into the queue.
Product copy from the spec sheet
Bounded generation, pinned model ID, prompt regression tests on every bump.
What this changes in your product
Engineering in service of a measurable outcome: EU when the contract requires it, tools over invention, measured FR quality.
Residency is no longer a slogan
Studio API is not Google Cloud Europe. When the contract requires it, we go through Vertex in an EU region, with audit proof.
The model calls your functions
Stock, price, ticket: the model queries your business instead of inventing. Fewer costly hallucinations.
The engine stays replaceable
An internal layer (provider, model, region). You switch to Mistral without rewriting the business.
French evals are a deliverable
Accents, SIRET, addresses: we replay cases on every model bump. FR quality is measured, not assumed.
How we ship your Gemini connector
Scoping
Studio (prototype) or Vertex eu (regulated prod). models.list on the chosen endpoint. Written decision, not a Europe slogan.
Evals
French corpus: SIRET, addresses, liaisons. The model is pinned after measurement, not on a latest alias.
Development
LLM abstraction, hostname logged, prompt minimisation, tokens / day caps, tool-call idempotence.
Monitoring
Tokens, regional 429s (quotas separate from global), endpoint traces. A model switch is a regression test.
What the Gemini API allows
- generateContent and streaming
- Text, PDF, image, audio in a single model. Delivery note, ID, invoice extraction, subject to quality measured in French.
- Tools / function calling
- The model calls your functions (stock, price, ticket) rather than inventing. Business idempotence on side effects.
- Vertex EU multi-region
- aiplatform.eu.rep.googleapis.com. Paris europe-west9 is locational, model-dependent matrix. EU is not France only.
- Two auths, not interchangeable
- AI Studio key on generativelanguage.googleapis.com. Vertex service account. Studio keys do not work on Vertex, and the reverse.
Gemini API vocabulary
- Developer API
- AI Studio, generativelanguage.googleapis.com, x-goog-api-key. No region. Google AI Developers staff points to Vertex to pin processing to the EU.
- Vertex eu
- Host aiplatform.eu.rep.googleapis.com, or europe-west* region. GCP project, IAM, service account. The documented path for a residency commitment.
- Global endpoint
- Google's phrase: doesn't support data residency requirements. Using it and saying this is Google Cloud Europe is false.
- ML processing vs storage
- Data at rest in the chosen location. ML processing follows the jurisdictional or locational endpoint. The model × country matrix changes.
- Separate quotas
- Global vs regional. A move to eu can 429 while global still passed. Anticipate, do not discover at go-live.
- ZDR / training
- Zero-data-retention and training opt-out: not asserted here. Paid contracts, pages to reread at scoping. We do not promise them on the strength of a slogan.
The real constraints of the Gemini API
Studio is not Europe
generativelanguage.googleapis.com and the global endpoint do not offer residency. Vertex eu is Google's documented path. The hostname is logged.
EU is not France
The eu multi-region is not europe-west9. The model-dependent matrix is reread at scoping. An EU DPA does not say France only, and we write that down.
Model IDs move
As elsewhere. Pin an ID listed on the chosen endpoint. Global has models eu does not have yet. Prompt regression tests on every bump, not a silent latest.
No improvised ZDR
We do not promise zero-data-retention or a training opt-out without rereading them in the account DPA. Prompt minimisation, always.
Gemini Developer API or Vertex EU?
Two Google doors, two promises. The right choice is the residency contract, not how easy the AI Studio key is.
| Criterion | Developer APIAI Studio | Vertex EUResidency |
|---|---|---|
| Host | generativelanguage.googleapis.com | aiplatform.eu.rep.googleapis.com |
| Residency | No region parameter | Documented EU (not France only) |
| Auth | x-goog-api-key | Service account, IAM |
| Billing | AI Studio plan | GCP, Vertex tokens |
| Models | Studio catalogue | Subset, models.list on eu |
| Quotas | Global / Studio ones | Separate, 429 possible on the move |
| The right case | Prototype, demo | Regulated prod, tender |
Studio keys do not work on Vertex, and the reverse. We isolate provider, model and region behind an internal interface. ZDR and training: DPA at scoping, not a page sentence.
What we measure on a Gemini integration
The other AI building blocks
Gemini is combined more often than it is replaced. These options are discussed at scoping.
GeminiWe build your Gemini connectorThis page
MistralInference on api.eu.mistral.ai, when the contract wants a European lab.
HeyGenThe video avatar, when text and voice do not carry the message.
ElevenLabsWe build your ElevenLabs connectorTranscriptionWe build your transcription connector
Anthropic (Claude)A model that calls your tools, not one more chat
OpenAIThe model writes into your tools, it does not chat
CursorThe agent reaches your internal tools, not just your codeWe combine Gemini with
The stack around Gemini on our projects.
Gemini: your questions
Three steps. Pick the door: Developer API for a prototype, Vertex eu (aiplatform.eu.rep.googleapis.com) as soon as a residency commitment exists. List models on that host, pin an ID, put an LLM abstraction (provider, model, region). Minimise the prompt, cap tokens, idempotence on tool calls, log the hostname. The sensitive part is not generateContent, it is not selling Studio as Google Cloud Europe, and not promising a ZDR unread in the DPA.
Only if the call goes to Vertex on an EU endpoint (aiplatform.eu.rep.googleapis.com or a europe-west* region), not to generativelanguage.googleapis.com nor to the global endpoint, which Google writes doesn't support data residency requirements. The eu multi-region is not France (europe-west9 is locational, model-dependent matrix). We log the hostname on every request. Zero-data-retention and training opt-out: not asserted here, to reread in the account DPA at scoping.
Studio: x-goog-api-key, no region, Studio plan billing, demo path. Vertex: GCP project, IAM, service account, regional or multi-region endpoint, GCP token billing. Keys do not swap. A move to eu can 429 (separate quotas) and a global model can be missing on eu: models.list on the chosen host. For a residency tender, Vertex eu. For an internal demo, Studio is enough, provided you say so. Confusing the two loses the tender.
GDPR plays on the door (Vertex eu vs Studio), prompt minimisation, and the DPA. We do not promise ZDR on this page. AI Act: if the user talks to an AI system, information (article 50) applies, as on a voice agent. Transparency in the UI, not only in a contract. When the frame does not hold, we propose Mistral api.eu or another engine behind the same internal interface. The hostname log is part of the audit proof.
A first useful flow, extraction or classification on Vertex eu with hostname logged, ships in two to three weeks. A chain with business tools, French evals and a Mistral failover is closer to six to eight weeks. Duration depends on Studio vs Vertex, the regional catalogue, and the eval corpus. We scope the perimeter up front and give you a firm estimate before we start, including the door (Studio or Vertex).
A Gemini integration project?
Let's talk. 30 minutes to scope Studio or Vertex eu, what the DPA allows, and tell you frankly what is opposable to a customer.
Discuss my Gemini project