Skip to main content
Version: v2026.0.0

LLM providers

Uxopian AI supports nine LLM providers. Provider configurations are loaded from llm-clients-config.yml at startup into OpenSearch, then managed at runtime via the Admin API.

Provider overview​

ProviderBean nameModels
OpenAIopenaigpt-5.1, gpt-4.1, gpt-4o
Anthropicanthropicclaude-sonnet-4, claude-opus-4
Azure OpenAIazure-openaigpt-4o via deployment
AWS Bedrockbedrockclaude-3-sonnet, cohere
Google Geminigeminigemini-2.5-pro, flash
Mistral AImistral-aimistral-large, mistral-small
HuggingFacehuggingfaceMistral-7B, Llama-3-8B
Ollamaollamallama3 (local)
NuExtractnu-extractspecialized extraction

Configuration structure​

Each provider is configured in llm-clients-config.yml under llm.provider.globals:

llm:
default:
provider: ${LLM_DEFAULT_PROVIDER:openai}
model: ${LLM_DEFAULT_MODEL:gpt-5.1}
base-prompt: ${LLM_DEFAULT_PROMPT:basePrompt}
context: ${LLM_CONTEXT_SIZE:10}
provider:
globals:
- provider: openai
defaultLlmModelConfName: gpt5
globalConf:
apiSecret: ${OPENAI_API_KEY:}
temperature: 1
timeout: 60
maxRetries: 3
llModelConfs:
- llmModelConfName: gpt5
modelName: gpt-5.1
multiModalSupported: true
functionCallSupported: true

Global configuration fields​

FieldDescription
providerProvider identifier (see table below)
defaultLlmModelConfNameDefault model configuration name for this provider
globalConf.apiSecretAPI key or secret credential
globalConf.endpointUrlBase URL for the provider API
globalConf.temperatureSampling temperature (0.0 to 1.0+)
globalConf.timeoutRequest timeout (e.g., 60s)
globalConf.maxRetriesNumber of retry attempts on failure
globalConf.extrasProvider-specific additional parameters

Model configuration fields​

FieldDescription
llmModelConfNameInternal name used to reference this model
modelNameActual model name sent to the provider API
multiModalSupportedWhether this model accepts image inputs
functionCallSupportedWhether this model supports function calling (required for tools)

Supported providers​

Each provider is a Spring @Service bean. The bean name (shown in the provider column) is the identifier used in llm-clients-config.yml and the admin UI.

Providerprovider valueAuth fieldsExtra parametersStreaming
OpenAIopenaiapiSecret, endpointUrl (optional)Yes
AnthropicanthropicapiSecret, endpointUrlYes
Azure OpenAIazure-openaiapiSecret, endpointUrlYes
AWS BedrockbedrockapiSecretAwsRegion, AwsAccessKey, AwsSessionTokenYes
Google GeminigeminiapiSecretYes
Mistral AImistral-aiapiSecret, endpointUrlYes
OllamaollamaendpointUrlYes
HuggingFacehuggingfaceapiSecretNo
NuExtractnu-extractapiSecret, endpointUrlproject, output, rasterizationDPI, enableThinking, thinkingTokenBudget, loopDetectionNo (streaming is refused)

All providers are implemented as LangChain4J wrappers. See Write a custom LLM client for the full source code of each provider and instructions to create your own.

Credential encryption​

API keys are encrypted with AES/GCM before being stored in OpenSearch. The encryption key is configured via app.security.secret-key in application.yml. If not set, a default development key is used. Set a unique key in production.

Default provider and model​

The provider and model of a turn are resolved in this order: the prompt of the turn (its defaultLlmProvider / defaultLlmModel), then the usage profile of the calling surface. A request cannot override them: the former provider, model, llm_provider and llm_model query parameters are no longer read.

The system-wide properties only seed usage profiles:

  • LLM_DEFAULT_PROVIDER: provider identifier (default: openai)
  • LLM_DEFAULT_MODEL: model name (default: gpt-5.1)
  • LLM_DEFAULT_PROMPT: base prompt ID (default: basePrompt)

llm.default.provider and llm.default.model are not a fallback at call time any more. They only fill the usage profile that is auto-created the first time a connection provider is used; when they are unset, that profile takes the tenant's first LLM provider configuration and that provider's default model. A call with no provider on its prompt nor on its usage profile fails with an explicit error.

Documents​

A document attached to a conversation (a PDF, or another file rendered to PDF) is only sent to a model flagged Reads documents (documentSupported, off by default) in the provider's model configuration; otherwise the request is refused with 422 LLM_CAPABILITY_ERROR. Native PDF input is supported for OpenAI, Anthropic, Gemini, Bedrock and Mistral; the other providers refuse documents. A per-model maxDocumentSizeMb and the global llm.document.max-size (default 20 MB) bound the size of a document.

Context size​

llm.context.size (or LLM_CONTEXT_SIZE) controls how many previous requests from the conversation are included in each LLM call. Default: 10.

Tenant overrides​

Per-tenant LLM provider configurations can be defined under llm.provider.tenants with a mergeStrategy of MERGE, OVERWRITE, or CREATE_IF_MISSING. This allows different tenants to use different API keys or models. See Multi-tenancy.