Course Wiki

Next Steps — LLM Zoomcamp Module 6

In this module, we covered hybrid search and document reranking, two of the most impactful techniques for improving RAG quality.

llmz-module-06

Notes

In this module, we covered hybrid search and document reranking, two of the most impactful techniques for improving RAG quality.

The key takeaways:

Hybrid search combines vector and keyword search for better retrieval. It works because vector search captures semantic meaning while keyword search catches exact term matches.
Reranking re-orders results with a better relevance score. Even a simple RRF implementation can improve metrics.
LangChain wraps Elasticsearch functionality in a cleaner interface, but the underlying queries are the same.

We only covered two of the five techniques from the intro lesson.

Here are pointers for the rest:

Small-to-big chunk retrieval: use small chunks for embedding but retrieve surrounding context for the LLM. Look at LangChain’s ParentDocumentRetriever.
Document metadata: add titles, topics, dates as filterable fields in your search index. Use them to narrow results before ranking.
Query rewriting: ask an LLM to reformulate the user’s question into a clearer search query before hitting the database.