Hybrid and Semantic Search

This is the retrieval stack that powers modern RAG applications. You will learn to combine two different ways of finding relevant documents: exact keyword matching and meaning-based vector similarity. By the end you will understand why production systems rarely use pure vector search, and how to build a search that handles both 'refund policy page 3' and 'what happens if I cancel my subscription?' with equal competence. This unlocks further study in reranking, chunking, agentic retrieval, and multi-document retrieval — all of which build on the hybrid foundation you will master here.

Imagine you are an archivist in a vast library. A researcher asks you: 'Find me everything about the 2016 election results.' You have two very different tools at your disposal. The first is an exact index: a card catalogue that lists every mention of the phrase '2016 election results' verbatim. This is fast and precise — you find the documents that literally contain those words. But if the researcher actually means 'the outcome of the presidential race in 2016' or 'who won in November 2016', the card catalogue is useless because it only matches exact strings. That is keyword search. The second tool is a semantic map: every document is a point on a map, and similar documents cluster together. You find the documents closest to your query's point, even if they don't share a single word with your question. That is vector search. The problem: the card catalogue misses paraphrases, and the semantic map can be confused by rare or specific terms like 'EIN-84-2093' (an ID code) which get blurred into an average position on the map. Hybrid search uses both tools at once, combines their scores, and returns a ranked list that is better than either alone. In practice, this means a search that understands meaning AND respects exact terms.

Think of a restaurant that has two ways of finding a dish on its menu. The first is the index at the back of the menu — it lists every page number where the word 'chicken' appears. If you look up 'chicken', you get page 2 and 5. That's keyword search: fast, exact, but if you ask for 'poultry' instead, the index finds nothing. The second way is the chef's mental map of ingredients — he knows that 'poultry', 'chicken', 'bird', and 'fowl' all point to the same section of the kitchen, because they are related in meaning. That's vector search: it understands synonyms and concepts, but if a customer asks for a specific dish code like 'TN-12' (a special number in the menu), the chef's mental map doesn't know what that means, because it's not a concept — it's just a string. Hybrid search is a waiter who does both: she looks up the exact code in the index, and also asks the chef for 'something like this dish' based on its meaning. She combines the two lists, gives higher rank to dishes that appear in both, and presents the best matches to the customer. The analogy breaks down when we think about the combination step: in the restaurant, the waiter just uses her judgment. In hybrid search, we need a mathematical way to merge two scores — a weighted sum. The waiter could, for instance, trust the exact index 70% and the chef's suggestion 30%. That's the core of score fusion: decide how much to trust each source, add them up, and sort by the total.

Hybrid search is a retrieval method that combines lexical keyword matching (which finds exact terms) with dense vector semantic similarity (which finds meaning-based matches) to produce a ranked list of documents that leverages both exactness and semantic understanding, typically by fusing the two score sets.

This concept builds on your existing knowledge of embeddings (which you have notes on) and vector search (which you have studied as a neighbouring topic). You know that embeddings turn text into lists of numbers, and vector search finds documents whose embeddings are closest to a query's embedding. Hybrid search adds a second retrieval signal on top of that: lexical matching. It also connects to reranking, because hybrid search is often the first stage of a two-stage retrieval pipeline: you retrieve a wide set using both methods, then rerank that set with a more expensive but more accurate model. Your notes on 'Admissible Evidence Spans' and 'Context Precision and Recall' are relevant here: hybrid search directly improves recall (the chance of retrieving all relevant passages) because two different methods cover each other's blind spots. Your note that 'hybrid search plus reranking beats pure vector similarity in most production settings' — from Large Language Models: From Foundations to Production AI — is the thesis of this page.