Add support for Agentic AI to Sonador: RAG in Orthanc, MCP server, inference runtime, diagnosis of CXR conditions and report generation

NANDA: blueprint/architecture for agentic systems.

Architecture and changes to Sonador:

• Implement RAG/embeddings endpoints within Sonador
  • Add python-pgvector to Orthanc Cloud plugin
  • Create embeddings models (similar to comments) for patient, study, and series
  • Create REST endpoints (three total) for embeddings: patient, study, and series
• Update the Sonador environment
  • Change the Orthanc database from PostgreSQL image to PgVector image
  • Update Kafka and drop etcd
• Sonador maintenance
  • Update Django to 5.2 from 4.2 (beyond EoL)
• MCP server PoC: expose embeddings capabilities to agents

PoC: demonstrate the feasibility of end-to-end case planning for total knees. Needed actions/tasks: receive medical images, segment structures of interest (femur and tibia), approve segmentations (or refine), determine anatomic landmarks, determine plan parameters (cut-plan, implant, sizing, ... )

• Inference server
  • NVIDIA Triton
    • Able to serve any type of self-hosted model
    • Supports VLLM
  • VLLM
    • Limited to LLMs only
  • Cloud-native / provider tooling
• Runtime environment (DevOps): cloud versus edge (bare-metal K8s)
  • MLflow: supports NVIDIA triton as a backend