Providers (EN)

Why you're reading this page: This page explains embedding providers (Mock, OpenAI, Gemini, Mistral, Azure, Claude): when to use which, env vars, and minimal code. It is the right place if you are choosing a provider for your first project or configuring for production.

Intentum uses embedding providers to turn behavior (e.g. "user:login") into vectors so the intent model can infer confidence. You can use a mock provider (no API key, for local runs and tests) or a real provider (OpenAI, Gemini, Mistral, Azure OpenAI, Claude) for production.

This page explains each provider: what it does, env vars, minimal code, and DI setup. For easy / medium / hard usage examples for every AI provider (Mock, OpenAI, Gemini, Mistral, Azure, Claude), see How to use AI providers. For the overall flow (Observe → Infer → Decide), see API Reference and Setup.

When to use which provider

Provider	Use when
MockEmbeddingProvider (Intentum.AI)	Local runs, tests, demos — no API key.
OpenAI	You already use OpenAI; good model choice and docs.
Gemini	You prefer Google; often good latency and pricing.
Mistral	You want a European option or Mistral models.
Azure OpenAI	You run on Azure or need enterprise SLAs.
Claude	You use Anthropic; supports message-based intent scoring.

You only need one embedding provider per app; pick the one that matches your stack and region. All use the same flow: LlmIntentModel(embeddingProvider, similarityEngine) then Infer(space).

OpenAI

What it does: Calls OpenAI’s embedding API (e.g. text-embedding-3-large) to turn behavior keys into vectors.

Env vars: OPENAI_API_KEY, OPENAI_EMBEDDING_MODEL, OPENAI_BASE_URL (optional, e.g. for proxy).

Minimal code (no DI):

using Intentum.AI.OpenAI;
using Intentum.AI.Models;
using Intentum.AI.Similarity;

var options = OpenAIOptions.FromEnvironment();
options.Validate();

var httpClient = new HttpClient
{
    BaseAddress = new Uri(options.BaseUrl ?? "https://api.openai.com/v1/")
};
httpClient.DefaultRequestHeaders.Authorization =
    new System.Net.Http.Headers.AuthenticationHeaderValue("Bearer", options.ApiKey);

var provider = new OpenAIEmbeddingProvider(options, httpClient);
var model = new LlmIntentModel(provider, new SimpleAverageSimilarityEngine());

var space = new BehaviorSpace().Observe("user", "login").Observe("user", "submit");
var intent = model.Infer(space);

DI (e.g. ASP.NET Core):

using Intentum.AI.OpenAI;
using Microsoft.Extensions.DependencyInjection;

var options = OpenAIOptions.FromEnvironment();
services.AddIntentumOpenAI(options);

// Then inject IIntentEmbeddingProvider (or OpenAIEmbeddingProvider) and build LlmIntentModel.

Gemini

What it does: Calls Google’s Gemini embedding API (e.g. text-embedding-004) to turn behavior keys into vectors.

Env vars: GEMINI_API_KEY, GEMINI_EMBEDDING_MODEL, GEMINI_BASE_URL (optional).

Minimal code (no DI):

using Intentum.AI.Gemini;
using Intentum.AI.Models;
using Intentum.AI.Similarity;

var options = GeminiOptions.FromEnvironment();
var httpClient = new HttpClient { BaseAddress = new Uri(options.BaseUrl ?? "https://generativelanguage.googleapis.com/v1beta/") };
var provider = new GeminiEmbeddingProvider(options, httpClient);
var model = new LlmIntentModel(provider, new SimpleAverageSimilarityEngine());

DI:

using Intentum.AI.Gemini;
using Microsoft.Extensions.DependencyInjection;

var options = GeminiOptions.FromEnvironment();
services.AddIntentumGemini(options);

Mistral

What it does: Calls Mistral’s embedding API (e.g. mistral-embed) to turn behavior keys into vectors.

Env vars: MISTRAL_API_KEY, MISTRAL_EMBEDDING_MODEL, MISTRAL_BASE_URL (optional).

Minimal code (no DI):

using Intentum.AI.Mistral;
using Intentum.AI.Models;
using Intentum.AI.Similarity;

var options = MistralOptions.FromEnvironment();
var httpClient = new HttpClient { BaseAddress = new Uri(options.BaseUrl ?? "https://api.mistral.ai/v1/") };
httpClient.DefaultRequestHeaders.Authorization =
    new System.Net.Http.Headers.AuthenticationHeaderValue("Bearer", options.ApiKey);
var provider = new MistralEmbeddingProvider(options, httpClient);
var model = new LlmIntentModel(provider, new SimpleAverageSimilarityEngine());

DI:

using Intentum.AI.Mistral;
using Microsoft.Extensions.DependencyInjection;

var options = MistralOptions.FromEnvironment();
services.AddIntentumMistral(options);

Azure OpenAI

What it does: Calls your Azure OpenAI embedding deployment to turn behavior keys into vectors. Uses endpoint + API key + deployment name.

Env vars: AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY, AZURE_OPENAI_EMBEDDING_DEPLOYMENT, AZURE_OPENAI_API_VERSION (optional).

Minimal code (no DI):

using Intentum.AI.AzureOpenAI;
using Intentum.AI.Models;
using Intentum.AI.Similarity;

var options = AzureOpenAIOptions.FromEnvironment();
var httpClient = new HttpClient();
// Azure uses ApiKey in header; extension usually sets it.
var provider = new AzureOpenAIEmbeddingProvider(options, httpClient);
var model = new LlmIntentModel(provider, new SimpleAverageSimilarityEngine());

DI:

using Intentum.AI.AzureOpenAI;
using Microsoft.Extensions.DependencyInjection;

var options = AzureOpenAIOptions.FromEnvironment();
services.AddIntentumAzureOpenAI(options);

Claude

What it does: Anthropic’s Claude can be used for message-based intent scoring (ClaudeMessageIntentModel), not only embeddings. By default the Claude package may use a stub for embeddings; use message scoring for full intent inference.

Env vars: CLAUDE_API_KEY, CLAUDE_MODEL, CLAUDE_BASE_URL, CLAUDE_API_VERSION, CLAUDE_USE_MESSAGES_SCORING (optional).

DI (recommended):

using Intentum.AI.Claude;
using Microsoft.Extensions.DependencyInjection;

var options = ClaudeOptions.FromEnvironment();
services.AddIntentumClaude(options);

Then inject the intent model (e.g. ClaudeMessageIntentModel) as needed. See package docs for message-based scoring usage.

Security and configuration

Never commit API keys. Use environment variables or a secret manager.
Avoid logging raw request/response bodies in production.
Region and latency: Choose a provider and endpoint close to your users if latency matters.
Rate limits: Respect each provider’s limits; consider retries and backoff (or a dedicated middleware) for production.

For full env var names and optional fields, see each provider’s *Options class and FromEnvironment() in the repo.

Next step: When you're done with this page → How to use AI providers or API Reference.

Table of Contents

Providers (EN)

When to use which provider

OpenAI

Gemini

Mistral

Azure OpenAI

Claude

Security and configuration