Knowledge Base Academy Community My Tickets
Fabric 8.5
Fabric 8.5

Installation and Setup

AI Fusion 3 is delivered as a set of K2exchange extensions that you install into your Fabric project from Studio. This article lists the extensions, the order in which to install them, and the configuration needed before you build your first agent.

Prerequisites

  • K2view Fabric and Web Studio. The extension manifests require Fabric 8.4.0 or later (AI Fusion Studio requires 8.4.7). Several AI Fusion 3 capabilities require Fabric 8.5.2 or later — in particular per-agent Observation configuration and the Analyze (Studio AI) actions in AI Fusion Studio. Fabric 8.5.2 or later is recommended.
  • An LLM provider account — Anthropic, AWS Bedrock, OpenAI (or an OpenAI-compatible endpoint), Google Vertex AI, or Ollama — and the credentials your organization approved for it.
  • A database for insights — SQLite (default, suitable for development) or PostgreSQL for shared and production environments.

Extensions

Extension Required What it provides
aifusion Yes The core platform: the aifusion Data Product (agent runtime, session tables, observation and evaluation flows, Observation web application), the AI Fusion web services in k2_ws, shared actors such as AgentInvoke, LLMInvoke and Embed, the Vectors LU and SQLite vector library, and the assurance, evaluation_storage and snapshot_storage interfaces.
aifusion-studio (AI Fusion - Studio) Yes The Studio development experience: AI Explorer, Agent Editor, Preview chat and Trace, Eval Editor, and the Studio AI skills for building agents and diagnosing answers.
pipeline Required for evaluations The pipeline framework (PipelineLu, Pipeline web app and services) used to run evaluations and observation analysis pipelines.
LLM connector: anthropic, aws-bedrock, openai, google-genai or ollama At least one The AI interface types for the provider (LLM and, where supported, embedding). Install the connectors your organization uses.
mcp (MCP Client) Optional The MCP interface type and actors, needed for agents that use tools from MCP servers.
aifusion-demo Optional (recommended for learning) A complete banking demo: the customer_bank Data Product with the banking agent, its sub-agents, reference packs and hook flows and the demo source databases.

Version notes (see AI Fusion Versions for the full compatibility table):

  • The xhigh and max thinking levels require recent connector versions (anthropic 3.0.3 or later; aws-bedrock, openai and google-genai 3.0.1 or later). Older connectors treat them as off.
  • When a connector is installed alongside aifusion, use aifusion 3.0.4 or later. From that version the core and the connectors no longer share files, so they can be upgraded independently.

Installation Steps

  1. In Web Studio, open K2exchange and install the aifusion, aifusion-studio and pipeline extensions. From aifusion 3.0.8 they are separate extensions and are not installed automatically. See Install an Extension.
  2. Install one or more LLM connector extensions.
  3. Optionally install mcp and aifusion-demo.
  4. Configure an LLM interface (below).
  5. Configure the insights database (below), if you are not using the SQLite default.
  6. Deploy the aifusion, k2_ws and PipelineLu Data Products, and every Data Product that contains agents. The first full (non-soft) deploy of aifusion creates the insights database tables.

After the deploy, open the AI Explorer in Studio to confirm that the installation works. If you installed the demo, the banking agent appears under the customer_bank Data Product.

Configure an LLM Interface

Agents call LLMs through Fabric AI interfaces of the connector's type (for example Anthropic, AWSBedrockLLM, OpenAI, GoogleLLM, OllamaLLM).

  1. In Studio, create a new interface of the connector type under Interfaces.
  2. Set the model, the credentials and other provider settings. Store credentials using Fabric's secret management.
  3. Set its Tag to default to make it the default LLM interface. Agents that do not name an interface use llm://default, which resolves to the interface tagged default.
  4. Make sure the interface is active in the environment you run. An interface that is not listed in the active environment is not available, and agents fail with Unsupported IO interface llm://default.

An agent can use a different interface through settings.llm_interface in its agent.yaml, and a different, typically cheaper, interface for context compaction and chat naming. See LLM Connectors and AI LLM Interface.

Configure the Insights Database

The insights database stores session metrics, Observation signals, evaluation results and batch jobs. It is set by shared globals:

Global Default Description
INSIGHTS_INTERFACE assurance The DB interface of the insights database. The default assurance interface is a SQLite file in the Fabric workspace.
INSIGHTS_INTERFACE_TYPE sqlite The database dialect: sqlite or postgres.
INSIGHTS_SCHEMA main The schema that holds the tables (for example public on PostgreSQL).
RETENTION_DAYS 30 Session retention used by the cleanup process (7–365 days).
EVALUATION_STORAGE_INTERFACE evaluation_storage Storage for evaluation snapshots.
EVALUATION_STORAGE_RETENTION_DAYS 90 Retention of evaluation snapshot files (30–365 days).
SNAPSHOT_STORAGE_INTERFACE snapshot_storage Storage for session snapshots taken during development.

To use PostgreSQL:

  1. Define a PostgreSQL DB interface.
  2. Override INSIGHTS_INTERFACE, INSIGHTS_INTERFACE_TYPE=postgres and INSIGHTS_SCHEMA (for example in the environment of the space).
  3. The pipeline extension has equivalent globals (PIPELINE_INTERFACE, PIPELINE_SCHEMA, PIPELINE_DIALECT) for its own database — point them to PostgreSQL as well.
  4. Deploy the aifusion LU (full deploy) to create the tables.

The deploy is idempotent: it creates missing tables and migrates tables created by earlier versions.

Observation

The Observation dashboard is a tab of the AI Data Fusion web application, enabled by default. See Opening Observation.

Next Steps

Next article: Agent Platform

Installation and Setup

AI Fusion 3 is delivered as a set of K2exchange extensions that you install into your Fabric project from Studio. This article lists the extensions, the order in which to install them, and the configuration needed before you build your first agent.

Prerequisites

  • K2view Fabric and Web Studio. The extension manifests require Fabric 8.4.0 or later (AI Fusion Studio requires 8.4.7). Several AI Fusion 3 capabilities require Fabric 8.5.2 or later — in particular per-agent Observation configuration and the Analyze (Studio AI) actions in AI Fusion Studio. Fabric 8.5.2 or later is recommended.
  • An LLM provider account — Anthropic, AWS Bedrock, OpenAI (or an OpenAI-compatible endpoint), Google Vertex AI, or Ollama — and the credentials your organization approved for it.
  • A database for insights — SQLite (default, suitable for development) or PostgreSQL for shared and production environments.

Extensions

Extension Required What it provides
aifusion Yes The core platform: the aifusion Data Product (agent runtime, session tables, observation and evaluation flows, Observation web application), the AI Fusion web services in k2_ws, shared actors such as AgentInvoke, LLMInvoke and Embed, the Vectors LU and SQLite vector library, and the assurance, evaluation_storage and snapshot_storage interfaces.
aifusion-studio (AI Fusion - Studio) Yes The Studio development experience: AI Explorer, Agent Editor, Preview chat and Trace, Eval Editor, and the Studio AI skills for building agents and diagnosing answers.
pipeline Required for evaluations The pipeline framework (PipelineLu, Pipeline web app and services) used to run evaluations and observation analysis pipelines.
LLM connector: anthropic, aws-bedrock, openai, google-genai or ollama At least one The AI interface types for the provider (LLM and, where supported, embedding). Install the connectors your organization uses.
mcp (MCP Client) Optional The MCP interface type and actors, needed for agents that use tools from MCP servers.
aifusion-demo Optional (recommended for learning) A complete banking demo: the customer_bank Data Product with the banking agent, its sub-agents, reference packs and hook flows and the demo source databases.

Version notes (see AI Fusion Versions for the full compatibility table):

  • The xhigh and max thinking levels require recent connector versions (anthropic 3.0.3 or later; aws-bedrock, openai and google-genai 3.0.1 or later). Older connectors treat them as off.
  • When a connector is installed alongside aifusion, use aifusion 3.0.4 or later. From that version the core and the connectors no longer share files, so they can be upgraded independently.

Installation Steps

  1. In Web Studio, open K2exchange and install the aifusion, aifusion-studio and pipeline extensions. From aifusion 3.0.8 they are separate extensions and are not installed automatically. See Install an Extension.
  2. Install one or more LLM connector extensions.
  3. Optionally install mcp and aifusion-demo.
  4. Configure an LLM interface (below).
  5. Configure the insights database (below), if you are not using the SQLite default.
  6. Deploy the aifusion, k2_ws and PipelineLu Data Products, and every Data Product that contains agents. The first full (non-soft) deploy of aifusion creates the insights database tables.

After the deploy, open the AI Explorer in Studio to confirm that the installation works. If you installed the demo, the banking agent appears under the customer_bank Data Product.

Configure an LLM Interface

Agents call LLMs through Fabric AI interfaces of the connector's type (for example Anthropic, AWSBedrockLLM, OpenAI, GoogleLLM, OllamaLLM).

  1. In Studio, create a new interface of the connector type under Interfaces.
  2. Set the model, the credentials and other provider settings. Store credentials using Fabric's secret management.
  3. Set its Tag to default to make it the default LLM interface. Agents that do not name an interface use llm://default, which resolves to the interface tagged default.
  4. Make sure the interface is active in the environment you run. An interface that is not listed in the active environment is not available, and agents fail with Unsupported IO interface llm://default.

An agent can use a different interface through settings.llm_interface in its agent.yaml, and a different, typically cheaper, interface for context compaction and chat naming. See LLM Connectors and AI LLM Interface.

Configure the Insights Database

The insights database stores session metrics, Observation signals, evaluation results and batch jobs. It is set by shared globals:

Global Default Description
INSIGHTS_INTERFACE assurance The DB interface of the insights database. The default assurance interface is a SQLite file in the Fabric workspace.
INSIGHTS_INTERFACE_TYPE sqlite The database dialect: sqlite or postgres.
INSIGHTS_SCHEMA main The schema that holds the tables (for example public on PostgreSQL).
RETENTION_DAYS 30 Session retention used by the cleanup process (7–365 days).
EVALUATION_STORAGE_INTERFACE evaluation_storage Storage for evaluation snapshots.
EVALUATION_STORAGE_RETENTION_DAYS 90 Retention of evaluation snapshot files (30–365 days).
SNAPSHOT_STORAGE_INTERFACE snapshot_storage Storage for session snapshots taken during development.

To use PostgreSQL:

  1. Define a PostgreSQL DB interface.
  2. Override INSIGHTS_INTERFACE, INSIGHTS_INTERFACE_TYPE=postgres and INSIGHTS_SCHEMA (for example in the environment of the space).
  3. The pipeline extension has equivalent globals (PIPELINE_INTERFACE, PIPELINE_SCHEMA, PIPELINE_DIALECT) for its own database — point them to PostgreSQL as well.
  4. Deploy the aifusion LU (full deploy) to create the tables.

The deploy is idempotent: it creates missing tables and migrates tables created by earlier versions.

Observation

The Observation dashboard is a tab of the AI Data Fusion web application, enabled by default. See Opening Observation.

Next Steps

Next article: Agent Platform

© Copyright 2026 K2VIEW