What we're building
No code today, this is the last lesson, and it's a bridge, not a capstone. lesson.py prints a summary of four specific limits this course ran into along the way, each paired with the lesson that demonstrated it and what addresses it next. Nothing here is new information, every limit was already shown hands-on in an earlier lesson; this just names them together, in one place, before this course ends.
Why this matters
naive_rag Lesson 26 named Corrective RAG as the fix for naive retrieval's confidently-wrong top-k chunk, and this course built that fix, by hand, and actually watched it work (Lesson 17's precision@k jump from 0.80 to 1.00) and actually watched it have limits (Lesson 16's circularity demo, Lesson 21's bounded loop). That's worth more than an abstract description of "Corrective RAG has limitations," because you've now built the mechanism, tested its assumption, and found the real edge.
Where to go from here
Corrective RAG is a real improvement over naive RAG, not a discarded attempt, this course's Lesson 17 numbers prove that directly. Agentic RAG (this series' next course) is what addresses this course's last two limits: instead of a fixed correction ladder every question runs through the same way, an agent decides, per question, what to do next, including whether grading and correction are even the right tools to reach for. This course's grading, filtering, rewriting, and external search don't get thrown away there, they become capabilities an agent can choose among, rather than a sequence every question runs through identically.
Congratulations on completing the course.