# AI (/docs/ai)



AI is the product domain for building and using governed AI experiences. It provides stable application interfaces while keeping model runtimes, storage engines, and workflow components behind platform boundaries.

## Choose an outcome [#choose-an-outcome]

<Cards>
  <Card title="Access models" href="/docs/ai/model-as-a-service" description="Discover and use platform-hosted or connected models." />

  <Card title="Build with the BSQAI API" href="/docs/get-started/build-ai-application" description="Follow an application journey from authentication to operation." />

  <Card title="Ground responses with RAG" href="/docs/ai/rag" description="Ingest organizational knowledge and retrieve relevant context." />

  <Card title="Connect tools with MCP" href="/docs/ai/mcp" description="Make Model Context Protocol servers available to an organization." />

  <Card title="Process documents" href="/docs/ai/components/docling" description="Convert documents for extraction and retrieval workflows." />
</Cards>

## Capability map [#capability-map]

### Model access and routing [#model-access-and-routing]

[Model as a Service](/docs/ai/model-as-a-service) covers model discovery, registration, cloud-model connections, and the hand-off into inference. Applications integrate through the [BSQAI API](/docs/ai/components/bsqai-api) instead of depending directly on internal gateways or serving engines.

Routing policies can select providers and models, apply fallback behavior, and retain telemetry without changing the application-facing contract.

### Conversational AI and knowledge [#conversational-ai-and-knowledge]

Conversational experiences combine model access, conversation state, organizational knowledge, and permission-aware retrieval. The Portal provides the user-facing experience; [RAG](/docs/ai/rag) explains ingestion, chunking, embeddings, vector search, and grounded responses.

Use the [AI and knowledge guided path](/docs/get-started/use-ai-and-knowledge) when the goal is to help users work with approved models and enterprise content.

### Documents, voice, and multimodal input [#documents-voice-and-multimodal-input]

Document processing converts files into structured content for extraction, classification, and retrieval workflows. Speech services add real-time or batch transcription and supported text-to-speech paths. These capabilities share authentication, workflow, storage, and observability services with other AI experiences.

Use [Docling](/docs/ai/components/docling) for document conversion details and [RAG](/docs/ai/rag) for the complete ingestion and retrieval flow.

### Agents and tool use [#agents-and-tool-use]

Agentic applications add orchestration, governed tools, connectors, and secure execution to model interactions. BullSequana AI exposes approved tool interfaces and carries the calling identity's tenant and team boundary into those integrations.

[MCP Servers](/docs/ai/mcp) is the tool-integration surface. A Platform Admin makes a Model Context Protocol server available to an organization and chooses which of its tools may be called. Each member links their own credential for the system behind it, and every call passes through the MCP Gateway.

Start with the [application-building path](/docs/get-started/build-ai-application), then use [How MCP works on the platform](/docs/ai/mcp/how-it-works) for the request flow and the [MCP Lifecycle Operator reference](/docs/ai/components/mcp-lifecycle-operator) for managed workload reconciliation.

### Model lifecycle [#model-lifecycle]

Training, fine-tuning, evaluation, promotion, registration, and serving form one lifecycle. Data workspaces and MLflow retain development and experiment context; AI exposes approved models through stable interfaces; Foundation runs the serving infrastructure.

The [Data and ML guided path](/docs/get-started/build-data-workflow) connects those responsibilities.

## Typical application flow [#typical-application-flow]

1. Identify whether the application needs direct generation, organizational knowledge, document processing, or tools.
2. Choose a model and confirm the tenant and team access boundary.
3. Authenticate to the BSQAI API and implement the smallest end-to-end request.
4. Add retrieval, documents, voice, or tools only when the use case requires them.
5. Deploy the application with health checks, resource limits, telemetry, and a rollback path.

## Architecture and implementation [#architecture-and-implementation]

<Cards>
  <Card title="Reference architecture" href="/docs/ai/reference-architecture" />

  <Card title="How AI builds on Foundation" href="/docs/ai/how-ai-builds-on-foundation" />

  <Card title="AI component catalog" href="/docs/ai/components" />

  <Card title="Develop applications" href="/docs/develop" />
</Cards>