LangChain¶
Seven weeks on LLM apps as backend systems: prompts as templates, tools as functions, memory as a store, RAG as search + a prompt, eval as tests.
This track assumes backend engineering fluency and the course prerequisites. It is not a first introduction to Python, HTTP, JSON, testing, or LLMs.
Scope: get the idea, build a small working chain, then use the official docs to go further. Read the framework track guide before starting.
The model is a dependency. LangChain is middleware. If you cannot redraw a chain as a sequence of function calls, the abstraction is hiding a bug.
| Week | Idea | You are done when… |
|---|---|---|
| 1 — Chains | A prompt template is f-string + a schema. |
Valid input returns the schema; malformed output is rejected. |
| 2 — Memory | Session store. What to keep, what to drop. | Two sessions remain isolated and old context is deliberately bounded. |
| 3 — Agents | A loop that picks tools. ReAct is not autonomy. | The trace shows tool selection, an observation, and a bounded failure. |
| 4 — RAG | Embed, retrieve, then prompt. Search quality first. | A small query set measures retrieval separately from answer quality. |
| 5 — Eval | You cannot ship what you cannot score. | A golden set produces repeatable pass/fail results. |
| 6 — Production | Timeouts, fallbacks, tracing—the operational shape. | One request exposes a trace, timeout, fallback, and cost record. |
| 7 — Ticket bot | Golden file, allowlist, “I don’t know”, cost line. | The supplied golden-file checks pass. |
Think of it like… Express / FastAPI middleware
The LLM is the slow, flaky downstream service. Your job is the same as always: validate input, bound the work, parse output, log enough to debug the next incident.