AI LLM Interface
The AI LLM interface type defines the connection to a Large Language Model (LLM) provider. Like any other Fabric interface, it is a Shared Object: the provider endpoint, the model and the API credentials are defined once, in the project, and are then used by every Fabric component that needs an LLM.
An AI LLM interface is consumed by:
- GenAI Data Fusion agents - see Use in GenAI Data Fusion.
- Broadway LLM utility Actors, addressed as
llm://<tag>. - External AI clients, through Fabric's OpenAI-compatible provider endpoint - including Studio AI.
Creating an AI LLM Interface
Step 1: Install the Provider Extension
LLM provider connectors are delivered as K2exchange extensions. Examples of supported providers are OpenAI, Anthropic, AWS Bedrock and Google Vertex AI.
- Open K2exchange in Fabric Studio.
- Choose the desired LLM connector extension and click Install.
- Deploy all changes.
Note: as with any installed extension, add its files to your project's Git repository.
Step 2: Create the Interface
Go to Project Tree > Shared Objects, right click Interfaces, select New Interface and then select AI LLM from the Interface Type dropdown menu. Populate the connection settings and click Save.
Go to Project Tree > Implementation > Shared Objects, right click Interfaces and select New Interface. Or: Go to the Top Menu, then select > Fabric > New Interface
Select an AI LLM from the AI section, enter a name for the interface and click Create. Populate the connection settings and click Save.
Connection Settings
Selecting an Interface
A project can hold several AI LLM interfaces - for example a lighter model for simple tasks and a stronger model for complex reasoning, or a secondary provider used as a fallback. Consumers select one by name or by tag, and Fabric resolves the reference in the following order:
- An interface whose name matches the given identifier.
- An interface carrying a matching tag.
- Otherwise, the interface tagged default.
In a Broadway flow, set the LLM Actor's interface parameter to llm://<tag>, for example llm://sql-generator.
Scenarios for Multiple Interfaces
When an LLM task is triggered, Fabric looks up the configured AI LLM interface and uses it to communicate with the underlying model, using the credentials defined in that interface. Tagging separate interfaces per task - for example a sql-generator interface alongside the default one - is what lets an implementation choose a different model for each of these steps.
Use in GenAI Data Fusion
The GenAI Data Fusion agent framework invokes LLMs at several points in the agent workflow:
When an LLM task is triggered, Fabric looks up the configured AI LLM interface and uses it to communicate with the underlying model, using the credentials defined in that interface. Tagging separate interfaces per task - for example a sql-generator interface alongside the default one - is what lets an implementation choose a different model for each of these steps.
Fabric as an LLM Provider
Starting from V8.5.1, Fabric exposes its configured AI LLM interfaces through an OpenAI-compatible REST endpoint. Any client that can talk to the OpenAI chat completions API - including the Studio AI assistant - can therefore point at Fabric instead of at a foundation model provider.
This means that:
- Users do not need a personal API key, and the organization does not need to distribute provider keys.
- The provider credentials stay in the Fabric project's AI LLM interface.
- Access is granted and audited with standard Fabric authorization.
- The model behind the endpoint can be swapped by editing the interface, without touching any client.
Endpoints
The base URL is the Fabric web services URL, followed by /api/v1, for example http://<fabric-host>:3213/api/v1.
The model field of the request selects the interface: use default for the default LLM engine, or the name or tag of a specific AI LLM interface.
Authorization
Requests are authenticated with a Fabric user or API token, passed in the standard Authorization: Bearer header, and require the LLM_INVOKE permission (or ALL):
GRANT LLM_INVOKE ON * TO <role>;
For more information about roles and permissions, see Fabric Authorization Commands.
Example
curl -s -X POST http://localhost:3213/api/v1/chat/completions \
-H "Authorization: Bearer <fabric-token>" \
-H "Content-Type: application/json" \
-d '{"model":"default","messages":[{"role":"user","content":"Say hi in three words."}],"stream":false}'
Connecting a Client
Any client that speaks the OpenAI chat completions API can be pointed at this endpoint by setting its base URL to the Fabric /api/v1 URL and leaving its API key empty.
For the Fabric Studio AI assistant, see Custom LLM Providers.
AI LLM Interface
The AI LLM interface type defines the connection to a Large Language Model (LLM) provider. Like any other Fabric interface, it is a Shared Object: the provider endpoint, the model and the API credentials are defined once, in the project, and are then used by every Fabric component that needs an LLM.
An AI LLM interface is consumed by:
- GenAI Data Fusion agents - see Use in GenAI Data Fusion.
- Broadway LLM utility Actors, addressed as
llm://<tag>. - External AI clients, through Fabric's OpenAI-compatible provider endpoint - including Studio AI.
Creating an AI LLM Interface
Step 1: Install the Provider Extension
LLM provider connectors are delivered as K2exchange extensions. Examples of supported providers are OpenAI, Anthropic, AWS Bedrock and Google Vertex AI.
- Open K2exchange in Fabric Studio.
- Choose the desired LLM connector extension and click Install.
- Deploy all changes.
Note: as with any installed extension, add its files to your project's Git repository.
Step 2: Create the Interface
Go to Project Tree > Shared Objects, right click Interfaces, select New Interface and then select AI LLM from the Interface Type dropdown menu. Populate the connection settings and click Save.
Go to Project Tree > Implementation > Shared Objects, right click Interfaces and select New Interface. Or: Go to the Top Menu, then select > Fabric > New Interface
Select an AI LLM from the AI section, enter a name for the interface and click Create. Populate the connection settings and click Save.
Connection Settings
Selecting an Interface
A project can hold several AI LLM interfaces - for example a lighter model for simple tasks and a stronger model for complex reasoning, or a secondary provider used as a fallback. Consumers select one by name or by tag, and Fabric resolves the reference in the following order:
- An interface whose name matches the given identifier.
- An interface carrying a matching tag.
- Otherwise, the interface tagged default.
In a Broadway flow, set the LLM Actor's interface parameter to llm://<tag>, for example llm://sql-generator.
Scenarios for Multiple Interfaces
When an LLM task is triggered, Fabric looks up the configured AI LLM interface and uses it to communicate with the underlying model, using the credentials defined in that interface. Tagging separate interfaces per task - for example a sql-generator interface alongside the default one - is what lets an implementation choose a different model for each of these steps.
Use in GenAI Data Fusion
The GenAI Data Fusion agent framework invokes LLMs at several points in the agent workflow:
When an LLM task is triggered, Fabric looks up the configured AI LLM interface and uses it to communicate with the underlying model, using the credentials defined in that interface. Tagging separate interfaces per task - for example a sql-generator interface alongside the default one - is what lets an implementation choose a different model for each of these steps.
Fabric as an LLM Provider
Starting from V8.5.1, Fabric exposes its configured AI LLM interfaces through an OpenAI-compatible REST endpoint. Any client that can talk to the OpenAI chat completions API - including the Studio AI assistant - can therefore point at Fabric instead of at a foundation model provider.
This means that:
- Users do not need a personal API key, and the organization does not need to distribute provider keys.
- The provider credentials stay in the Fabric project's AI LLM interface.
- Access is granted and audited with standard Fabric authorization.
- The model behind the endpoint can be swapped by editing the interface, without touching any client.
Endpoints
The base URL is the Fabric web services URL, followed by /api/v1, for example http://<fabric-host>:3213/api/v1.
The model field of the request selects the interface: use default for the default LLM engine, or the name or tag of a specific AI LLM interface.
Authorization
Requests are authenticated with a Fabric user or API token, passed in the standard Authorization: Bearer header, and require the LLM_INVOKE permission (or ALL):
GRANT LLM_INVOKE ON * TO <role>;
For more information about roles and permissions, see Fabric Authorization Commands.
Example
curl -s -X POST http://localhost:3213/api/v1/chat/completions \
-H "Authorization: Bearer <fabric-token>" \
-H "Content-Type: application/json" \
-d '{"model":"default","messages":[{"role":"user","content":"Say hi in three words."}],"stream":false}'
Connecting a Client
Any client that speaks the OpenAI chat completions API can be pointed at this endpoint by setting its base URL to the Fabric /api/v1 URL and leaving its API key empty.
For the Fabric Studio AI assistant, see Custom LLM Providers.
