Week 3 — LLM Core
5 lessons · project on day 18
- Open
Day 13
Prompt Engineering
Prompting as engineering, not folklore — structured output, schema-constrained generation, and evals that survive a model upgrade.
- Open
Day 14
Embeddings
What a vector actually is, cosine similarity, and why chunking strategy — not the embedding model — is where most RAG systems fail.
Day 15
Vector Database
Day 16
RAG
Day 17
MCP
Day 18 project
Docs RAG Assistant
Point it at a documentation site, chunk and embed the pages, and answer questions with inline citations back to the source section. The interesting work is not the retrieval call — it is the chunking strategy, and discovering how badly naive fixed-size chunks perform against headings-aware ones.
Stack
- TypeScript
- pgvector
- OpenAI-compatible embeddings
- Next.js
- Zod
The hard parts
- Chunk boundaries that split a code sample or a table in half, which quietly destroys answer quality without ever throwing an error.
- Deciding when retrieval found nothing useful — a cosine similarity of 0.62 looks fine numerically and is often garbage.
- Citations that survive re-indexing, which means chunk IDs must be stable across runs rather than array positions.
Why it belongs in your portfolio
Everyone claims RAG experience. A repo with a chunking-strategy comparison and a retrieval evaluation set shows you have actually measured it instead of shipping the first thing that returned text.