Build hybrid search with a frozen query set

Hybrid search is easy to describe: combine lexical and vector retrieval. Proving that the combination is better takes a more careful sequence.

The 30-Day Retrieval Basics roadmap starts with lexical search. The learner builds a tokenizer, inverted index, and BM25 ranker before reaching embeddings. That order matters. Lexical retrieval gives exact terms, inspectable scores, and a strong baseline that is cheap to run.

Week two freezes the query set and builds an evaluation harness. Ranking metrics now have concrete inputs: queries, relevance judgments, result lists, and a cutoff. Changing the retriever without keeping this set stable would make the comparison harder to trust.

Week three adds vector retrieval. A brute-force index keeps the first implementation readable. The learner can inspect similarity, top-k behavior, and metadata filters before approximate nearest-neighbor structures introduce another quality and latency tradeoff.

The final week combines the two result lists and adds reranking. Reciprocal Rank Fusion is useful here because it can merge rankings without pretending the scores from BM25 and vector similarity share one scale. The capstone must show whether the hybrid path improves the frozen queries and where it makes them worse.

This workflow resists a common search mistake: adding embeddings before defining the failures they are meant to fix. Some queries need semantic recall. Others depend on an exact identifier, name, or phrase. A hybrid system earns its complexity when the evaluation shows that each retriever covers a real weakness in the other.

The complete roadmap is at https://learn.significanthobbies.com/curriculum/roadmaps/retrieval-30-day.