When to use Vector Search

Vector search and Hybrid search are now available for Standard Tier customers. But what are they, and when should you use them? There are plenty of technical articles out there on vector search, but they show engineering details and mathematical concepts.

As a product owner or decision maker, it's your job to understand when or when not to add a feature to serve customers. But you have a problem: this vector search thing that all the engineers talk about is very technical. So technical, in fact, that you are likely struggling to grasp how it will actually deliver value.

This post starts from the top - no engineering details and no math. We describe it at a product level to give you the knowlege required to make a decision for your domain and business needs. This article will explain exactly why vector search helps, and also how it can hurt your search experience. We'll also discuss Hybrid Search as a solution to these potential gaps. We'll first give a brief explanation, and then show specific examples in various domains: e-commerce, research, enterprise, entertainment, and customer support. Even if your product domain isn't in the list, the examples will give you an excellent starting point for addressing your customers' information needs.

Two kinds of search

Before vector search, we were stuck with "lexical" aka "keyword" search. Lexical is very good at matching exactly what the customer is searching for. It's really helpful for finding names, identifiers, terms of art, brands, exact phrases, and advanced search expressions.

However, lexical search is not very good at finding things that mean the same thing, but are expressed with different words. That's where vector search comes in: it's very good when trying to find concepts, synonyms, misspellings, different languages, and the general meaning of passages.

Here's a simple way to think about it: lexical search matches words, vector search matches meaning.

Since both technologies offer strengths, they complement each other when used together. We call the blending of these two "hybrid search". It enables us to take the strengths of exact matching and combine them with similar meaning.

Hybrid Search

So which one should you use, lexical or vector? Both. They complement each other and should be used together. Each one covers the other's gaps.

Lexical search is precise, fast, and predictable. When a user types an exact term, you want exact results. No guessing, no interpretation, just a match. It's also cheaper to run and easier to debug. If your product deals heavily in structured data, codes, identifiers, or expert terminology, lexical search will be doing most of the heavy lifting.

Vector search is flexible, forgiving, and exploratory. When a user doesn't know the right words, or describes what they want in their own language, vector search fills the gap. It's particularly valuable for natural language queries, multilingual content, and discovery-oriented experiences. If your product serves a broad audience who may not know your domain vocabulary, vector search will earn its keep fast.

Here's a practical guide for deciding how to weight each side:

Lean heavier on lexical when:

• Your users are experts who know exact terminology
• Your content has strong structured metadata (IDs, codes, categories)
• Precision matters more than discovery
• You need Boolean logic (AND, OR, NOT)

Lean heavier on vector when:

• Your users are non-experts or consumers
• Your content is mostly unstructured text
• Discovery and exploration matter more than exact matching
• You serve multilingual audiences or deal with lots of synonyms

Use a balanced hybrid when:

• You serve both expert and casual users
• Your content is a mix of structured and unstructured
• You want the highest overall relevance across diverse query types

The exact blend depends on your data, your users, and your use case. Start with a 50/50 split, measure what your users click on and engage with, and adjust from there. There's no magic ratio, only the ratio that works for your customers.

Where vector search can hurt

There are real tradeoffs you should know about when using vector search:

Precision loss for exact queries. If someone searches for a specific product ID or error code, vector search can muddy the results by surfacing "similar" things that aren't what the user wanted. This is exactly why hybrid search exists: let lexical handle the exact stuff.

Embedding quality matters. Vector search is only as good as the model generating your vectors. A poorly chosen or poorly tuned embedding model will give you irrelevant results dressed up as intelligence.

Harder to debug. When lexical search returns bad results, you can look at the tokens and figure out why. When vector search returns bad results, you're staring at a bunch of numbers from a model trying to understand what went wrong. It's less transparent, and that opacity makes tuning harder.

We've helped many teams choose optimized search relevance, choose the right embedding model, and help debug and improve their search quality specific to their domain. Contact us if you're interested in Bonsai and would like a consultation.

Domain Examples

Let's look at how each technology shines in practice. In the tables below, the left column shows where lexical search excels, and the right column shows where vector search picks up the slack.

E-Commerce | Research and Expert Search | Enterprise and Internal Knowledge | Entertainment and Media | Customer Support

E-Commerce

If you run a product catalog, you've got a mix of structured data (SKUs, brands, sizes) and unstructured data (descriptions, reviews, style notes). Lexical search is your workhorse for exact lookups. A customer searching for "Nike Air Max Size 12" wants exactly that, and keyword matching nails it.

But what about the customer who searches for "comfortable running shoes for flat feet"? There's no SKU for that. No brand field matches. Vector search understands the intent behind those words and surfaces products whose descriptions discuss arch support, cushioning, and comfort for flat-footed runners.

Lexical	Vector
Product IDs / SKUs / ISBNs	Categories and themes
Brand names	Styles and aesthetics
Sizes and measurements	Descriptions and reviews
Exact colors (e.g. "Midnight Blue #003366")	Similar colors (e.g. "dark blue")
Model numbers	Use cases and benefits
Exact price filters	"Affordable" or "premium" intent

Research and Expert Search

Researchers and analysts live in a world of precision. They know exactly what they're looking for: a specific author, a citation, a legal statute, a chemical compound. Lexical search is built for this. Boolean queries with AND, OR, and NOT are the bread and butter of expert search, and have been for decades.

But research also involves exploration. A scientist studying "CRISPR gene editing side effects" might benefit from results about "off-target mutations in genome modification" even though those words barely overlap. Vector search bridges that vocabulary gap, surfacing relevant work that uses different terminology to describe the same concepts.

Lexical	Vector
Named entities (people, places, orgs)	Conceptual meaning
Citations and references	Related ideas and themes
Terms of art and jargon	Plain-language equivalents
Exact phrases and quotes	Paraphrased content
Logical / Boolean expressions	Exploratory queries
Statute or regulation numbers	Interpretations and commentary

Enterprise and Internal Knowledge

Every company has a pile of internal docs: wikis, runbooks, HR policies, onboarding guides, and Confluence pages that haven't been updated since 2019. When an employee searches for "PTO policy" they want the exact document with that title. Lexical search handles that.

But when a new hire searches "how do I request time off," they're asking a question, not citing a document title. Vector search understands that "request time off" and "PTO policy" are about the same thing. This is especially powerful for companies where documentation is scattered across tools and written by different people using different words for the same processes.

Lexical	Vector
Document titles and IDs	Natural language questions
Employee names and departments	Role descriptions and responsibilities
Policy numbers and codes	Situational queries ("what do I do when...")
Product and project names	Conceptual documentation
Error codes and log messages	Troubleshooting descriptions

Entertainment and Media

Streaming platforms, music services, news sites, and content libraries all depend on search to help users discover things. A user who knows the title of a movie or song will find it instantly with lexical search. "Bohemian Rhapsody" returns Bohemian Rhapsody. Simple.

But entertainment is driven by discovery. Someone searching "feel-good movies about underdogs" isn't looking for a specific title. They're describing a mood, a theme, a vibe. Vector search connects that description to movies whose plots, reviews, and metadata match the sentiment, even when they don't share a single keyword.

Lexical	Vector
Titles and artist names	Moods and themes
Genre tags	"Movies like..." or "songs similar to..."
Release years and ratings	Narrative descriptions
Actor / director names	Plot summaries and reviews
Episode or track numbers	Tonal and emotional similarity

Customer Support

Support teams deal with a constant stream of tickets, FAQs, and knowledge base articles. When a support agent searches "error code 5012," they need lexical precision. That's an exact match problem.

But customers don't usually know error codes. They write things like "my screen goes blank when I try to upload a photo." Vector search maps that description to known issues about upload failures, display rendering bugs, or file size limits, even if the knowledge base article never uses the word "blank."

Lexical	Vector
Error codes and ticket IDs	Symptom descriptions
Product names and versions	"How do I..." questions
Feature names	Workaround and troubleshooting intent
Status codes	Customer frustration signals
Exact log messages	Natural language problem descriptions

Go build something

You now have a framework for thinking about when vector search helps and when it doesn't. The key takeaway: it's not lexical or vector. It's about understanding where each technology shines and letting them work together.

If you want to go deeper on the search fundamentals that underpin all of this, check out What AI Engineers Should Know About Search and Surviving the AI Mandate. And if you're ready to start building, we'd love to help.

See you next time!