CoreAI API

CoreAI API is the main backend service layer for CoreAI. It hosts application logic and integrates with identity, retrieval, model access, storage, and observability services while presenting a stable product-facing API to developers and clients. It also includes the model installer functionality that prepares and installs models for use in the platform, exposed through the existing /v1/model-installer API routes.

Why It Is Used

It gives the platform a central application service for orchestrating AI features, retrieval flows, and backend operations in a controlled way. It also protects clients from direct dependence on lower-level internal components, which means the platform can evolve its internal AI stack without changing the main developer integration surface.

Learn More

Anthropic Messages API

The CoreAI API also serves an Anthropic-compatible endpoint at /anthropic. This allows tools that speak the Anthropic wire format, such as Claude Code and the Anthropic SDK, to use platform-hosted models with platform authentication. No Anthropic account is required.

See Claude Code for configuration and usage.

Developer Guidance

Developers should generally integrate with the CoreAI API rather than depending directly on LiteLLM, Docling, or other internal services. This keeps application integrations aligned with the platform contract and gives BullSequana AI freedom to evolve the internal implementation over time.

Model Installation

Model installation is part of the CoreAI API backend. It connects model sources, model artifacts, model registry records, and the Runtime inference layer, handling the operational work required to move models from import and registration into usable inference deployments. It standardizes how models move from tracked artifacts into deployable inference assets, which is especially useful when models need to be pulled from Hugging Face or S3-compatible storage, registered in MLflow, and then exposed to Runtime inference services.

Typical flow:

Import a model from Hugging Face or from S3-compatible storage.
Upload or register the model artifacts in MLflow.
Optionally connect those artifacts to an inference registration flow.
Register the model into the Runtime inference environment, especially KubeAI.
Make the model available to higher-level CoreAI services.

Practical notes:

It supports both REST and gRPC APIs.
It can process long-running imports asynchronously and track task progress.
It can register models already stored in S3 directly in MLflow without re-uploading them.
It is the operational bridge between model lifecycle management and actual inference availability.

Interacts With

Keycloak for authentication and protected API access
LiteLLM behind the scenes for LLM calls and model gateway behavior
Milvus for vector storage and semantic retrieval
Docling for document parsing workflows
MLflow for model lifecycle metadata and GenAI observability traces
Rook Ceph for object storage and document artifacts through RGW
PostgreSQL for application state and transactional data
OpenTelemetry for trace instrumentation and telemetry export
ToolHive for MCP-oriented integrations
KubeAI, vLLM, and other Runtime inference services where installed models are consumed
Runtime ingress services that expose the API