Skip to content

Providers and data protection

Evokai runs agents on any model provider pydantic-ai supports. Which ones an application may use is a question of law as much as of taste: where the data is processed, which company (and which country's authorities) controls the service, what the provider keeps. This page describes the provider catalog, the facts Evokai knows about every provider and every other external processor, with the official page each fact was read from.

The provider catalog

src/evokai/llm/providers.yaml (code: evokai.llm.providers) is versioned with Evokai. Per provider it states:

Fact Field Values
Who operates the service entity.name, entity.country, entity.eu_contracting_entity the company, its country, who a customer in the EU contracts with
Who controls it entity.controlled_from an ISO country code, or EU for a group owned across member states
Training on API data terms.training false, true, or unknown
What is kept terms.retention, terms.retention_days none, abuse_monitoring, stored; days
Zero data retention terms.zero_retention default, on_request, no
Prompt caching caching.style explicit, automatic, implicit_and_explicit, none

and per endpoint (a Bedrock inference profile, an Azure deployment type, a Vertex location, a regional base URL): the model names it serves (models, globs over provider:model), the models the provider lists there (offered), where processing happens (place: a geography, ISO countries, provider regions, or the deployment's own AWS region) and how Evokai connects (api, base_url, api_key_env, api_version, location, project, region).

Every fact carries its sources and the date it was read (sources, checked). A fact nobody could verify on an official page is null and listed in the provider's unknowns; it is never guessed, and an unknown fact never satisfies a data protection rule. Evokai's own list prices (prices) are in the same file; they never reach a browser.

processors describes what Evokai calls besides chat models: guard classifiers (Bedrock Guardrails, Azure Prompt Shields and Content Safety, Google Model Armor, OpenAI and Mistral moderation, the local Prompt Guard, a self-hosted Llama Guard), the AgentCore code sandbox, server speech, the OAuth connections and the observability exporter.

Coverage

Provider Model prefix Endpoints Jurisdiction
Amazon Bedrock bedrock: eu., us., apac., global. profiles; in-region ids in EVOKAI_AWS_REGION US
Azure OpenAI / AI Foundry azure: Global, Data Zone EU, Data Zone US, Regional (declared per deployment) US
Google Vertex AI google-vertex: eu and us multi-regions, EU regional endpoints, global (declared per deployment) US
Google Gemini API google-gla: global US
OpenAI API openai:, openai-eu: global; European data residency (eu.api.openai.com, eligible projects) US
Anthropic API anthropic: global (inference_geo offers us and global only) US
Mistral AI mistral:, mistral-eu:, mistral-us: global (data hosted in the EEA by default), EU, US EU (FR)
Cohere cohere: not documented CA
Groq groq: global US
DeepSeek deepseek: China CN
OpenRouter openrouter: global; EU routing (eu.openrouter.ai, EU/EEA plus UK and CH) US
Hugging Face huggingface: Inference Providers (third parties); dedicated endpoints by region US
Scaleway Generative APIs scaleway: Paris EU (FR)
OVHcloud AI Endpoints ovhcloud: "hosted and run in the European Union" EU (FR)
IONOS AI Model Hub ionos: Germany (de-txl) EU (DE)
STACKIT AI Model Serving stackit: eu01 EU (DE)
Aleph Alpha none no public API documented (sign-up only) EU (DE), pending a combination with Cohere
Ollama, vLLM, LM Studio ollama:, vllm:, lmstudio: self-hosted: what the operator declares declared

GET /v1/admin/providers returns the catalog (without prices) with its coverage: how many facts it states and how many it leaves unknown. The console's Providers page shows it with every source.

A deployment's own entries

EVOKAI_PROVIDERS_FILE names a YAML file of the same shape, merged over the built-in catalog by id: a provider's fields replace the built-in ones, an endpoint with a known id replaces it, a new endpoint is tried before the built-in ones. An endpoint can copy the facts of a built-in one with like and state only what differs:

providers:
  - id: azure
    title: Azure OpenAI
    endpoints:
      - id: ours-sweden
        like: azure-datazone-eu         # the EU Data Zone's facts and sources
        models: ["azure:*"]             # azure:<deployment name>
        api: azure
        base_url: https://ours.openai.azure.com
        api_version: "2024-10-21"
        api_key_env: AZURE_OPENAI_API_KEY
        place: {geography: eu, countries: [SE], regions: [swedencentral]}
        declared_by: operator
  - id: google-vertex
    title: Google Vertex AI
    endpoints:
      - id: ours-eu
        like: vertex-eu-multiregion
        models: ["google-vertex:*"]
        api: google-vertex              # needs pydantic-ai-slim[google]
        location: eu
        project: my-project
  - id: our-vllm                        # a new provider: also a model prefix (our-vllm:<model>)
    title: Our vLLM in Milan
    self_hosted: true                   # no per-token charge
    entity: {name: ACME S.p.A., country: IT, controlled_from: IT}
    terms: {training: false, retention: none, zero_retention: default}
    endpoints:
      - id: milan
        models: ["our-vllm:*"]
        api: openai-compatible
        base_url: http://vllm.internal:8000/v1
        api_key_env: OUR_VLLM_KEY
        place: {geography: eu, countries: [IT]}
        declared_by: operator

A contract that changes a fact (an approved zero data retention, an Azure modified abuse monitoring) is declared the same way, on the endpoint (terms: {retention: none}, declared_by: operator). The model still has to be in EVOKAI_MODEL_ALLOW.

How models are built

evokai.llm.catalog.ModelCatalog builds a model from its endpoint's connection when it has one: openai-compatible (Scaleway, IONOS, STACKIT, OpenAI's EU residency, self-hosted servers), azure, google-vertex, mistral (Mistral's EU and US endpoints), anthropic, bedrock (with the endpoint's region); otherwise Bedrock in the deployment's region, or pydantic-ai's infer_model (OVHcloud, Ollama, Groq...). The API key comes from the environment variable the endpoint names (SCALEWAY_API_KEY, IONOS_API_KEY, STACKIT_API_KEY, MISTRAL_API_KEY, OPENAI_API_KEY); a missing key or base URL is a configuration error, never a silent fallback.

The data protection profile

An application's data_protection (code: evokai.protection) says which external processors its data may reach. Three independent dimensions, each optional; the ones that are set combine, and all must hold:

Dimension Values Holds when
residency geographies (eu, us, apac, global...) and/or countries (ISO codes) the endpoint's geography is listed, or every country it may process in is
jurisdiction eu and/or ISO country codes the provider is controlled from the EU or EEA (eu) or from a listed country
retention any, no_training, zero_retention the API terms say inputs are not used for training (and, for zero_retention, nothing is kept past the request)

An unknown fact never satisfies a rule. A model whose provider's jurisdiction the catalog cannot show is excluded by a jurisdiction rule with the reason unknown; a self-hosted server satisfies the rules its operator declares in EVOKAI_PROVIDERS_FILE. Something that runs inside Evokai (a test model, the local Prompt Guard) sends nothing out and satisfies every rule.

Example: EU residency, EU jurisdiction and zero data retention.

application:
  data_protection:
    residency: {geographies: [eu]}
    jurisdiction: [eu]
    retention: zero_retention

With the built-in catalog this allows Scaleway, IONOS, STACKIT, OVHcloud and Mistral's EU endpoint for residency and jurisdiction, excludes Bedrock, Azure, Vertex, OpenAI and Anthropic (US-controlled), and then keeps only the providers whose terms state that nothing is kept and nothing is trained on (Scaleway, for one; Mistral keeps 30 days for abuse monitoring). A deployment whose contract changes a fact declares it on its endpoint (terms: {retention: none}, declared_by: operator).

A profile that sets sovereignty is refused (at bootstrap and by PATCH /v1/admin/application): that dimension was removed before release. A stored profile that still holds it is read without it.

Where it applies

Processor Checked
Chat models at agent save (400 residency), token mint for a requested model (403 model_residency), every call and fallback (a model outside is never tried)
Embeddings when a knowledge base chooses its model, and for memories
Guard classifiers before a text is sent: an excluded classifier is skipped and the turn records classifier_unavailable with the reason
Code sandbox (AgentCore) code_run and skill_run are not offered when AgentCore in its region is excluded
Server speech (STT, TTS) at the voice offer (403 data_protection)
Connections (Microsoft 365, Google Workspace, GitHub, Linear, Notion, Slack) when a person starts one (403 data_protection)
Observability exporter deployment-wide: reported, not enforced per application
Browser voices and recognition the application's choice: declared in browser (online_voices, online_recognition, vendors) and reported

The older residency declaration (model_regions, models) keeps working beside the profile; both must hold. Its model_regions now reads geographies from the provider catalog, so an EU provider passes model_regions: [eu] too.

GET /v1/admin/data-protection returns the profile and every processor with its facts, whether it is allowed and why not (reasons: dimension and outside, unknown, training or retained). The console shows it on the Privacy page; the GDPR pack (section 2) states the profile and each processor's residency, jurisdiction and terms as known.

Regulatory spaces

The profile is written with the European regulatory space in mind (the GDPR's transfers and processors, a provider controlled from the EU or EEA, an EU contracting entity), but its dimensions are not EU-specific: residency and jurisdiction take any geography or country, and retention is about API terms anywhere. Another regulatory space adds what it needs the same way (a fact on catalog entries, stated only where an official source says so, and a rule in the profile), without touching the other dimensions.