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 specific limits this course ran into along the way, each paired with where in this series (or in a production system beyond it) that limit gets addressed. Nothing here is new information, every limit was already demonstrated hands-on in an earlier lesson; this just names them together, in one place, before this course ends.
Why this matters
This course committed to one implementation choice in Lesson 1, captioning-then-embed, on purpose, and spent every lesson since building out its consequences, good (reuses the existing pipeline, inspectable captions) and bad (lossy summaries, a per-image vision cost at generation time). A student who only remembers "multimodal RAG means captioning images" missed half of Lesson 1; a student who can also name what captioning costs, and what joint embedding spaces trade instead, actually understands the choice this course made rather than having memorized its outcome.
Where to go from here
Multimodal RAG extended naive_rag's exact pipeline with one new idea, captioning-then-embed, and every lesson since Lesson 1 built out that idea's real consequences, not just its happy path. Adaptive RAG (this series' next course) picks up the "fixed retrieval depth" thread directly: this course's ask(query, state, k) always used a k chosen in advance, Adaptive RAG decides k, or whether to retrieve at all, per question. This course's ingest()/ask() (Lesson 22) is built to slot into that course's routing layer without modification, exactly the point of the shared Strategy protocol.
Congratulations on completing the course.