So far, your retrieval pipeline has treated every document as if it were the same size. Today you learn why that breaks everything. Long documents dominate the context window, hide relevant information, and force the LLM to sift through noise it can’t meaningfully process.
Day 2 is all about fixing that with chunking, the process of splitting large documents into smaller, useful pieces before embedding. You'll explore fixed-size chunking with LangChain’s token splitter, experiment with different chunk sizes and overlaps, and immediately see how those choices shape what your system can actually retrieve. Then you move into structural chunking, using headings, code blocks, and chapter markers to create chunks that respect a document’s natural boundaries.
Once your chunks look solid, you test how retrieval behaves. You compare BM25 for keywords, embeddings for meaning, and RRF for a combined ranking. Different queries surface different weaknesses, and you get a feel for how chunk size, overlap, and structure impact the quality of your results.
Finally, you'll implement reranking which is a lightweight LLM pass that filters the top search results down to only the truly relevant chunks. Instead of overwhelming the model with 30 candidates, the reranker selects the few that actually answer the query.
By the end of Day 3, you know how to structure documents, understand how retrieval methods interact with your chunks, and have a reranker that sharpens your results. You’re no longer hoping retrieval will work, you’re engineering the context your model depends on.