learn-ai — Architecture & Data Flows
Generated 2026-06-24 13:33 UTC · c4gen dev
learn-ai is MIT Open Learning's conversational/agentic discovery sidecar: a Django + DRF service whose request path runs entirely on Channels/ASGI (Granian) so it can stream LLM responses over Server-Sent Events. Each chat turn drives a LangGraph ReAct agent that calls tools against the MIT Learn API (vector resource search, syllabus/content-file search, video transcript search, tutor problem sets) and calls an LLM through a LiteLLM proxy. APISIX fronts the service, enforcing Keycloak OIDC for learner-facing agents and static key-auth for Canvas integrations. Chat history (LangGraph checkpoints) and sessions persist in Postgres; Redis backs the Channels layer, Django cache, and the Celery broker.
This is a C4 view of learn-ai within the MIT Open Learning SOA, focused on how data is created and propagated — synchronous request paths and asynchronous (queued, scheduled, event-driven) flows alike. Use it for onboarding and as a holistic reference when realigning flows or hunting harmful cycles and fragile linkages.
How to read these diagrams
These are C4 model diagrams (C4-PlantUML). Read them top-down: System Context (the whole SOA) → Container (one system's runtime units) → Dynamic (a single data flow, step by step).
- People are rounded boxes; systems and containers are rectangles; databases and queues have distinct shapes.
- Each arrow is a data flow labelled with what moves.
- Solid arrows are synchronous (request/response, caller blocks).
- Amber dashed arrows are asynchronous (queued, scheduled, or event-driven — caller does not block).
- Drag to pan, scroll to zoom. Boxes with a link drill into the next level.
Contents
- System Context — learn-ai and the systems it exchanges data with.
- Containers — the runtime units inside learn-ai.
- Data Flows — key interactions, step by step (sync & async).
- Dependencies & Cycles — graph-derived coupling, cycles, fragile links.
Keeping this current
These pages are generated from a structured model by
architecture_maps/c4gen. The cross-service edges are extracted deterministically
from the witan-code graph; node prose and scenarios are curated. See
the generator README
to regenerate after the system changes.