Insight · September 20, 2026

A database that
searches by meaning.

Every AI answer grounded in your own data, every agent that remembers, reaches through the same piece of infrastructure. Here is what it is.

01 · The shift

Meaning is the thing you cannot look up.

For fifty years a database found a row by matching it exactly. You asked for the customer named Ortega and it returned the row that said Ortega. That worked because the question and the answer used the same word.

It breaks the moment the match is about meaning. A support article about a refund and a customer asking to get their money back share almost no words. A keyword index sees two unrelated strings. A person sees one question. Closing the gap between those two is the job a vector database exists to do, and retrieval augmented generation made that job standard across 2026.

02 · The problem

A word is a string. Meaning is a number.

A model does not read text the way you do. It turns each piece of text into an embedding, a long list of numbers that places the meaning of that text at a point in space. Two pieces that mean the same thing land near each other, even when they share no words. Refund and get my money back sit close. Refund and quarterly earnings sit far apart.

Once your documents are numbers, the question of which ones are relevant becomes a question of which points sit nearest. That is not something a normal database is built to answer. It matches values. It does not measure distance. More than 80 percent of enterprise data is unstructured, the notes and tickets and transcripts that never fit the exact match model, and that is exactly the data a vector database was built to search.

The one line to keep

“A vector database does not look for the row that matches. It looks for the ones that are close.”

03 · What actually happens inside

Store the meaning, then search by distance.

Embed

Every document, image, or record goes through an embedding model and comes out as a vector, a list of numbers that fixes its meaning as a point in space.

Store

The database keeps each vector next to the original, so a match can always point back to the real thing it stands for.

Index

It builds a structure over the vectors so it can find the near ones without comparing the query to every record it holds.

Search

A query is embedded the same way, then the database returns the records whose vectors sit closest, measured by cosine similarity, Euclidean distance, or dot product.

Return

The nearest records come back ranked by closeness, and whatever asked the question, a model or a person, gets meaning instead of a keyword.

04 · Close enough, on purpose

The search is approximate, and that is the point.

Finding the true nearest point means comparing the query to every vector you hold. At a few thousand records that is instant. At a few hundred million it is hopeless. So a vector database does not look for the exact nearest neighbor. It looks for the approximate one. Approximate nearest neighbor search trades a sliver of accuracy for orders of magnitude of speed, and for search by meaning that trade is almost always right.

The dominant method is a graph called the hierarchical navigable small world, or HNSW, introduced by Yury Malkov and Dmitry Yashunin in 2016. It builds the vectors into a graph in layers. The top layer holds a few long range links that jump across the whole space. Each layer below adds shorter links and more detail. A search starts at the top, moves quickly toward the right region, then drops down layer by layer to pin the closest points. The result is search that stays fast as the data grows into the billions.

05 · The example that shows why it matters

The same question, grounded in your own data.

Here is where it earns its place. A language model knows only what it was trained on. Ask it about your own policies, your own tickets, your own catalog, and it guesses. Retrieval augmented generation fixes that by putting a lookup in front of the model. The vector database holds your documents as embeddings. A question comes in, gets embedded, and the database returns the handful of passages that sit closest in meaning.

Those passages go into the prompt, and the model answers from them instead of from memory. That is what grounds an answer in fact and keeps it from inventing one. It is also what gives an agent a memory: the same lookup, run against what the agent has already seen, so it can recall instead of repeat.

question   →  embed   →  vector

vector     →  nearest neighbor search   →  matching passages

passages   +  question   →  model   →  answer grounded in your data

A retrieval step · the model answers from what was found, not what it recalls

06 · Choosing one

You may already have one running.

The category looks crowded, and it is, but the choices sort into three shapes. There are purpose built engines, among them Pinecone, Weaviate, Milvus, Qdrant, and Chroma, each built from the ground up for vector search. There are extensions that add vector search to a database you already run, the most common being pgvector, which turns PostgreSQL into a vector store. And there are libraries like Faiss, the vector search toolkit from Meta, which give you the raw search with no database around it.

The honest default for most teams in 2026 is the one they already operate. For a retrieval workload under roughly ten million vectors, pgvector on existing Postgres carries most of the load with no new infrastructure. Cross into tens of millions of vectors, or need hybrid search that blends meaning with keywords and filters, and a purpose built engine like Qdrant or Weaviate starts to pay for itself.

The market reflects the pull. Analysts put the vector database category near 2.49 billion dollars in 2025 and around 3.09 billion in 2026, on a path toward 17.37 billion by 2034. The number that matters more is quieter: agents now issue far more queries than people do, and every one of them is a lookup by meaning.

07 · When not to reach for one

Not every search needs one.

A vector database is not a replacement for the database you have. If your queries are exact, give me the order with this id, show every invoice from March, a relational database is faster, cheaper, and exactly right. Vector search earns its cost only when the question is about meaning and the match is fuzzy.

It also adds moving parts. Embeddings have to be generated, kept in sync when the source changes, and regenerated when you swap the model that made them. An index has to be tuned. And approximate search means the occasional near miss, which is fine for finding related documents and wrong for anything that has to be precise.

The rule is simple. Reach for a vector database when meaning is the query. Keep your existing database for everything else. Most real systems run both, and the skill is knowing which question belongs to which.

Closing

Search stopped being about the words. It is about the meaning.

The unit of search is no longer the word you typed. It is the meaning behind it, stored as a number, found by distance. A vector database is the piece that makes that lookup possible. Once you see it, you see it everywhere: in the answer that cites your own documents, in the assistant that remembers, in the search that finally understands the question.

Written by · September 20, 2026

Donny Smith · ECD, Founder, Bttr.

Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.

Sources · Fortune Business Insights, Vector Database Market, 2026 · Zilliz and MongoDB on the hierarchical navigable small world, method introduced by Malkov and Yashunin, 2016 · DataCamp and Firecrawl, Best Vector Databases 2026 · IBM, Vector Databases for RAG · all retrieved September 2026

Share this perspective

More insights

Adjacent perspectives.

8 min read

How to Evaluate a RAG System

Evaluating a RAG system means scoring its two stages on their own, because a retrieval augmented generation answer is only as good as the passages behind it, and a wrong answer has two possible causes a single number cannot tell apart. Retrieval is measured before the model writes anything. The LlamaIndex evaluation guide scores a retriever with hit rate, the share of questions where the right passage lands in the top results, and mean reciprocal rank, which rewards ranking it high. Generation is measured given what was retrieved. Ragas, the open evaluation toolkit now maintained by VibrantLabs, splits the answer into claims and scores faithfulness as the share the retrieved context supports, then scores answer relevance by generating questions from the answer and comparing them to the real one. Read as two pairs, the four metrics locate the fault: high recall with low faithfulness is a generation problem, low recall with high faithfulness is a retrieval one. Anthropic showed the payoff in September 2024, cutting a retrieval failure rate measured by recall at 20 from 5.7 percent to 1.9 percent by adding context, keyword search, and reranking. And it is not a one time check. Models drift, documents go stale, so the teams that trust their systems run these scores on live traffic the way they run unit tests.

8 min read

What Is a Reasoning Model

A reasoning model is a large language model trained to work through a problem in steps before it answers, rather than reply in one pass. It spends extra computation at the moment you ask, called test-time compute, to write out a chain of thought and check its own work. Anthropic describes its extended thinking as letting the very same model give itself more time and effort rather than switching to a separate model, and reported that Claude 3.7 Sonnet reached 84.8% on the hard GPQA science set, with a 96.5% physics subscore, using parallel sampling at a 64K token thinking budget. The habit is trained, mostly by reward. DeepSeek showed in January 2025 that its R1 Zero model learned to reason purely through reinforcement learning with no supervised fine-tuning first, developing self checking and long chains of thought on its own, and shipped DeepSeek R1 under the MIT license at performance comparable to OpenAI o1. The cost is real, so the newest models add a dial. Claude Opus 4.5 hit its best score on SWE-bench Verified using 76% fewer output tokens at a medium effort. The skill is matching how hard the model thinks to how hard the task is.

8 min read

What Is Fine-Tuning

Fine-tuning is training a pre-trained model further on your own labeled examples so a behavior is set into its weights. It changes how a model acts, its tone, its format, its grip on a narrow task, rather than what facts it knows. Google frames the move as taking a general-purpose model and adapting it to your domain on a curated dataset of your own examples. The gains are measurable. When Anthropic opened fine-tuning for Claude 3 Haiku in Amazon Bedrock, a team fine-tuned it to moderate comments and lifted classification accuracy from 81.5% to 99.6% while cutting tokens per query by 85%. Lighter methods such as LoRA keep the base model frozen and train a small set of new weights beside it. The order that saves money is a plain prompt first, retrieval when the model lacks facts, fine-tuning only when it lacks the right behavior. A fact that changes is a retrieval problem. A behavior that drifts is a training problem.

View all insights →

Bttr. Field Brief

The brief Bttr. writes for senior buyers.

Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.

Build search that understands the question

A database that
searches by meaning.

Meaning is the thing you cannot look up.

A word is a string. Meaning is a number.