Skip to content

CAR-T Intelligence — Advanced Learning Guide

For engineers extending or operating this subject.

Source: core/agents/cart · 62 Python files · 22,155 LOC · 11 test files

Registered capabilities

Capability Type Status Endpoint
cart-intelligence-agent agent live localhost:8521

UI :8521 · API :8522 (platform convention: registry endpoint is the UI, API is UI + 1)

Principal modules

src/knowledge.py

get_target_context, get_toxicity_context, get_manufacturing_context, get_all_context_for_query, get_knowledge_stats, get_pediatric_cart_context

  • get_target_context — Get formatted knowledge context for a target antigen.
  • get_toxicity_context — Get formatted knowledge context for a toxicity profile.
  • get_manufacturing_context — Get formatted knowledge context for a manufacturing process.
  • get_all_context_for_query — Extract all relevant knowledge context from a query string.

src/export.py

generate_filename, export_markdown, export_json, export_pdf, export_fhir_r4

  • generate_filename — Generate a timestamped filename for export.
  • export_markdown — Export a query result as a Markdown report.
  • export_json — Export a query result as structured JSON.
  • export_pdf — Export a query result as a professionally styled PDF report.

src/query_expansion.py

expand_query, expand_query_by_category, get_expansion_stats

  • expand_query — Extract expansion terms from a user query.
  • expand_query_by_category — Like expand_query but returns terms grouped by expansion category.
  • get_expansion_stats — Return the number of keywords and total terms per expansion map.

app/cart_ui.py

init_engine, init_agent, render_evidence_cards, build_conversation_context

  • init_engine — Initialize the CAR-T RAG engine (cached across reruns).
  • init_agent — Initialize the autonomous CAR-T Intelligence Agent.
  • render_evidence_cards — Render evidence cards with relevance indicators.
  • build_conversation_context — Build conversation context from recent exchanges for follow-up queries.

src/vector_collections.py

CARTCollectionManager

  • CARTCollectionManager — Manages 11 CAR-T Milvus collections (10 owned + 1 read-only genomic).

Dependencies

anthropic>=0.18.0, apscheduler>=3.10.0, biopython>=1.83, fastapi>=0.109.0, loguru>=0.7.0, lxml>=5.0.0, numpy>=1.24.0, opentelemetry-api>=1.29.0, opentelemetry-sdk>=1.29.0, prometheus-client>=0.20.0, pydantic-settings>=2.7, pydantic>=2.0, pymilvus>=2.4.0, python-dotenv>=1.0.0

Running the tests

.venv/bin/python scripts/run_all_tests.py cart

Two traps the shared harness handles, which a hand-rolled pytest invocation will not:

  1. Several subjects ship src/vector_collections.py, which shadows the Python standard library. Putting their src/ on PYTHONPATH kills the interpreter before collection.
  2. structural-biology/vendor_rfdiffusion/ is vendored third-party code needing gated GPU packages and is excluded.

Operational notes

On-target off-tumour toxicity and cytokine release are real. Any construct suggestion carries its monitoring requirement.

Before changing a port, read ../../build/PORT_MAP.md. The convention is enforced by scripts/validate_registry.py, which also cross-checks health-monitor.sh — a port change in one place and not the other fails the build.

Extending it

  1. Add or change code under core/agents/cart.
  2. Keep the capability entry in lib/hcls_common/capabilities.json truthful — a live status must answer a health probe. Two capabilities were found registered live with nothing bound to their ports; do not add a third.
  3. Run the gate: ruff, pytest lib/hcls_common, validate_registry.py, run_all_tests.py.