Survey of the (Hybrid) Search Landscape

Erik Hatcher (MongoDB) opens the Hybrid Search series with a survey of the findability landscape: what techniques exist, how they differ in rankability, and how search-centric applications have historically combined them. The technical mechanics are deliberately deferred — what this article establishes is the discipline that makes combining anything worthwhile.

First article in the series:

1. This article — survey of the landscape
2. Reciprocal Rank Fusion and Relative Score Fusion — RRF and RSF in detail
3. Hybrid Search Blueprint Series Semantic Boosting — Semantic Boosting

---

Hybrid search, a definition

Hatcher offers a working definition drawn from experience rather than the literature:

combining two or more search techniques to produce results better than any single technique alone.

The load-bearing word is better. Combining techniques is only justified if the combination outperforms the parts, and the only way to know is to measure — which yields the mantra repeated throughout the series.

He frames the whole exercise pragmatically: there is a lot of low-hanging fruit, and one can go a long way just by turning on search and doing basic queries, or by vectorizing content and queries for instant semantic findability. The recurring check is “Is the juice worth the squeeze?” — the ability to over-engineer is not a reason to.

Measure, tune, repeat

There is no single best metric for what makes one result better than another. Hatcher contrasts two poles: e-commerce, highly dynamic and subjective, where the ultimate business metric is profit; and library search, more static and objective, where subject matter experts would mostly agree on the best results for a given query.

His starting practice, before any tuning:

• Log the queries — with context: timestamp, number of results presented, and user details to whatever level is available and permitted. Sometimes the document IDs returned are logged too, since the result set changes over time and can be useful for auditing; the top ~20 shown at a time are tractable to log.
• Log what users do with the results — clicks, including the position the result was at when clicked, plus downstream actions like add-to-cart or purchase. What to capture is entirely application-dependent: what would be useful implicit feedback in your environment?
• Analyze the query logs — which queries found nothing, which are most popular, what a user tried in a session before landing on a chosen result. Can you learn from the query trail before the goal?

Only once basic search analytics are in hand does parameter tuning begin: adjust, measure against the previous results, keep improvements, revert regressions.

Search techniques, ordered by rankability

Three basic technologies, presented in order of rankability — the ability of the query-document relationship to be given a numerical rank or score.

Key/value matching

The most basic query: exact field values, served by the database’s built-in B-Tree index. The same structure serves range queries (e.g. a published_date within the last three months). A document either matches or it doesn’t — no explicit scoring or ranking. Proximity can still be computed (how close published_date is to now), or documents can carry their own ranking value.

Vector search

Vectors from an embedding model give semantic similarity: documents about the same topics sit close together in a hyper-dimensional space, the query is embedded into the same space, and a geometric distance measure indicates closeness.

The embedding model choice is the primary relevancy factor. Models vary in representational capacity and technique — Matryoshka Embeddings and quantization awareness buy lower-dimensional and faster calculations plus reduced storage while retaining meaning (see Vector Quantization).

This is precisely why the “measure” half of the mantra matters: you are entrusting your users’ search success to what Hatcher calls the embedding black box, so evaluating that results are as expected is non-optional.

Lexical search

Ligatures, diacritics, synonyms, elisions, Unicode, stems, acronyms — words matter, yet typos happen. How do you find “the most popular snorkling gear” when the most important word in the query is misspelled? Finding needles in haystacks needs a dictionary, or three.

Documents are ranked on how well they match, as precisely or as loosely as is sensible: blending matches across fields and their particular dictionaries, attaching weights or formulas to query clauses, and factoring in document statistics (term frequency) and corpus-level statistics (document frequency). See Full-Text Search and BM25.

Sweet spots Hatcher highlights:

• As-you-type suggest — with a few characters and some context, the user finds what they want and has already acted on it. See Autocomplete.
• Part numbers and similar codes — SKUs, years, makes, models. Users enter them in a form slightly different from the canonical value; a spurious dash shouldn’t block them.
• Voice-to-text queries — the transcription carries the important terms but may be garbled, with spurious rambling attached. Fuzzy lexical matching plus typical or domain synonyms catches many mangled queries.

But he is explicit about the limits: it is just lexical matching. It isn’t the tool for true semantic search, and it’s overkill for exact-value-only matching.

Historical “hybrid search”

Before vector fusion, hybrid meant something broader — Hatcher treats “hybrid” as a philosophy or mindset of blending rather than a prescriptive recipe.

A search system shouldn’t be an island. Implicit contexts are factors like time of day, device type, location where available, and language settings. Explicit feedback includes the items a user interacts with, the activity type performed on them (add-to-cart, purchase, read-detail, save to favorites), the session behavior trail, and — notably — what a user was presented but did not interact with.

The loop:

1. Collect signals — any bits of information users provide via queries, clicks, and interactions.
2. Learn from the signals — ML jobs aggregate raw signals into actionable insights.
3. At query time, look up insights available for this user, search context, or query, and augment the search engine request with the relevant actions.
4. Log the query and its contextual information.

Concrete insights he gives:

• A user searches blue shoez; learned spell correction internally rewrites it to blue shoes from there onward. See Spelling Correction in Search.
• Previous blue shoes queries led to purchases of Product #3748, so that product ID is boosted in the main query. See Results Boosting.
• Desktop app users see several additional facets versus mobile users — a manual rule set up by online experience managers.

The synergy is the whole being greater than the sum of the parts: feedback loops that learn from system activity and feed useful insights back in.

Hybrid fusion techniques

With vector-based semantic matching on the rise, and lexical search bringing in otherwise-missing but potentially highly relevant content, it is now common practice to embed the query, issue both a vector search and a lexical search, and fuse the two result sets.

Conceptually it’s zipping two result sets together. Reciprocal Rank Fusion (RRF) and Relative Score Fusion (RSF) are the popular mechanisms — blending by relevancy order (ignoring the actual computed distance or relevancy scores), or by sorting on sensible score range normalizations. Detailed in Reciprocal Rank Fusion and Relative Score Fusion.

The 747 cockpit

Hatcher’s analogy for the tuning surface: knobs, dials, levers, meters, displays, graphs, buttons, pedals, status indicators, buzzers, and additional computers (agents). There is no one right setting for every condition — settings change in response to the current environment. Even in autopilot, the computer makes minute adjustments constantly to keep things smooth.

Beyond the ranked list

The hybrid search challenge is more than producing a single good result set. What matters is what the user does with it, and that depends on how they arrive at it: presentation, format, facets and navigation affect the overall experience immensely. Can the user find what they want in a few taps? Can they navigate sensibly despite their typos? Is the system understanding their information-seeking intent and making good use of all the implicit and explicit signals? Is information ambiently findable?

Related Concepts

• Hybrid Search — the family this article defines and surveys
• Reciprocal Rank Fusion — fusion by relevancy order
• Relative Score Fusion — fusion by normalized score range
• Semantic Boosting — the series’ third technique
• Full-Text Search — the lexical leg
• Dense Vector Retrieval — the semantic leg
• Matryoshka Embeddings — cited for lower-dimensional, faster embeddings
• Vector Quantization — quantization awareness for reduced storage
• Click Signals — the raw material of the signals loop
• Implicit Judgments — implicit feedback derived from behavior
• Autocomplete — a named lexical sweet spot
• Results Boosting — how learned insights re-enter the query
• Search Architecture — the surrounding system

Related Articles

• Reciprocal Rank Fusion and Relative Score Fusion — part 2 of this series
• Hybrid Search Blueprint Series Semantic Boosting — part 3 of this series
• RRF is Not Enough — a sharper critique of rank-only fusion
• Hybrid search > sum of its parts? Berlin Buzzwords 2022 — tests the “better than the parts” claim

Related Topics

• Search Observability — the logging and analytics practice this article prescribes
• E-commerce Search — his example of a dynamic, subjective relevance environment
• Search UX — presentation, facets and navigation as part of findability

People

• Erik Hatcher

Companies

• MongoDB