LLM Connectors

An agent never talks to a model provider directly. It calls a Fabric interface — llm://<name or tag> — and the interface's connector translates the call into the provider's API. Switching an agent from one model or provider to another is therefore a configuration change, not a code change.

This article lists the connector extensions, their interface types and key properties, and explains how agents select an interface, how thinking levels and prompt caching map to each provider, and how MCP interfaces add external tools. For creating an interface in Fabric Studio, see the Fabric article AI LLM Interface. For installation order and prerequisites, see Installation and Setup.

Connector Extensions

Each provider is a separate K2exchange extension. Install the ones you need next to the aifusion extension, then create interfaces of the types they add.

The LLM connectors require Fabric 8.4.0 or later. They do not depend on the aifusion extension, but when both are installed, aifusion must be 3.0.4 or later.

Each connector caches one client per interface configuration. Changing any property of an interface creates a new client on the next request; deploying any Data Product (LU) clears the cache.

Key Interface Properties

All AI interface types have a Tag property, used to select the interface (see Selecting an Interface).

Anthropic

AWS Bedrock

The IAM identity needs bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream, plus bedrock:ApplyGuardrail when a guardrail is set. AWSBedrockEmbedding takes ModelType (Amazon Titan Embed or Cohere Embed) or a custom Model, and Region. AWSBedrockKnowledgeBase takes KbId and Region.

Google GenAI

GoogleEmbedding adds Type, the embedding task type (RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING).

OpenAI

OpenAIEmbedding takes AuthenticationType, Token, Model and Host.

Ollama (preview)

OllamaLLM takes HOST, PORT (default 11434), Path (default /api/chat), ProtocolType, optional HTTP authentication (Basic, Bearer or OAuth2) for an Ollama behind a gateway, Model (exactly as shown by ollama list), Temperature and Think. OllamaEmbedding uses the /api/embed path.

Selecting an Interface

The default Tag

Agents normally use the default interface. llm://default resolves to the interface whose Tag is default; more generally, llm://<id> resolves to the interface with that name, then to an interface with that tag, then to the default one (see AI LLM Interface).

• Keep exactly one LLM interface tagged default in each environment. New Anthropic and OllamaLLM interfaces are created with the Tag default, so change it when you add a second one.
• The default interface must be active in the environment the agent runs in. If none is, agent calls to llm://default fail with an unsupported-interface error.
• Because default is resolved at run time, you can point the same agents at a different model per environment (for example a cheaper model in QA) without editing any agent.

Per Agent, Per Task and Per Turn

settings:
  llm_interface: claude_sonnet        # an interface name or tag; "default" if omitted
  compaction_interface: claude_haiku
  chat_name_interface: claude_haiku

Thinking Levels

settings.thinking sets how hard the model reasons before answering: off (default), low, medium, high, xhigh or max. It is a provider-neutral intensity, not a token budget; each connector maps it to its own API and clamps when the provider has no exact equivalent.

• Ollama has no intensity dial: any level other than off turns thinking on.
• xhigh and max need Anthropic connector 3.0.3 or later, or Bedrock, OpenAI or Google GenAI connector 3.0.1 or later. Older connectors ignore or mis-map them.
• xhigh is the recommended level for agentic Claude agents that loop over tools or delegate. max costs considerably more for a small gain; high suits simple single-domain agents.
• Thinking shares the response budget. On Claude 4.6 and later, reasoning and answer both come out of the max-tokens budget; the connectors raise it to at least 16,000 when thinking is on. A tight budget at a high level can produce a truncated answer.
• off really is off. Claude 4.6 and later reason by default, so the connectors send an explicit "disabled" for off.
• An unrecognized level (a typo) is logged and treated as off.
• Reasoning is streamed to the client as thinking_delta events. The level applies to the main agent's LLM calls.

Prompt Caching

settings.use_cache (default true) asks the connector to cache the stable prompt prefix — the system prompt and tool definitions — so repeated calls in a session cost less and respond faster.

Anything that changes the prompt prefix — a different skill set, a context compaction — misses the cache for that call.

Embeddings

Embedding interfaces (OpenAIEmbedding, AWSBedrockEmbedding, BedrockLambdaEmbedding, GoogleEmbedding, OllamaEmbedding) are used through the Embed actor in Broadway flows — for example a flow tool that runs a semantic search. Embed takes the interface, an input that is either one string or an array of strings (batched in one request when the provider supports it), and an optional dimensionsOverride; it returns embeddings, total_tokens and dimensions. Anthropic offers no embedding models.

Batch Inference

Several connectors support the providers' asynchronous batch APIs, for offline workloads where no one waits for the answer. Each needs different configuration on the LLM interface:

MCP Interfaces

The MCP Client extension (preview) adds the Mcp interface type, which connects Fabric to a server implementing the Model Context Protocol. An agent can use every tool the server exposes:

mcp_interfaces:
  - crm_mcp            # the Mcp interface name

The server's tools are advertised to the model as MCP__<interface>__<tool> — for example MCP__crm_mcp__get_open_tickets — and called through the interface. Tool hooks and the tool-name guard apply to them like any other tool. See Tools.

Fabric 8.4 and later can expose Data Products as MCP servers, so an agent in one environment can use another Data Product's capabilities as MCP tools. See Data Product MCP.

The extension also provides Broadway actors (McpListTools, McpCallTool, McpListPrompts, McpGetPrompt, McpListResources, McpReadResource) for using MCP servers from your own flows.

Credentials and Secrets

Store provider keys and credentials through Fabric's secret manager or environment-level interface overrides, not in plain text in the project. Prefer identity-based authentication where the platform offers it — the instance IAM role for Bedrock, Application Default Credentials for Google. See Secrets Management and Security and Privacy.

Next article: Security and Privacy