What Is a Vector Database? A Plain-English Guide

· 6 min read

A vector database is a store for embeddings, the lists of numbers that stand for what a passage means, built to find the entries nearest to a question's numbers quickly. It is one way to keep search by meaning, "more like this" suggestions and the document lookup behind some assistants fast when there are many passages.

You may not need one, and a short test further down shows how to tell.

What does a vector database do?

It does three jobs, and writing the answer is left to a separate model.

  • Store. Each entry holds the embedding, usually the passage or a pointer to it, and details such as its file, date and who may read it. What an embedding is covers how text becomes numbers.
  • Find the nearest. Given the numbers for a question, it returns the closest entries, best first. To stay quick, many use shortcuts that skip most entries, so a result may not be the single closest.
  • Filter. It can limit the search by those details: only this folder, only this year, only what this reader may see.

An ordinary database answers exact questions about stored values, such as every order from one customer after March. A row matches or it does not. A vector database answers "which entries are most like this?", ranked by closeness.

A keyword search matches the words you typed. Suppose a handbook says "Staff may work remotely when their manager agrees" and you ask "Can I do my job from home?" The two share no meaningful word, so a word search can miss the sentence, while a search by meaning is likely to find it. Search for "invoice INV-2047" and the reverse holds: a word search finds that exact string, while a search by meaning may return similar invoices.

So they suit different questions, and good setups often combine them. Many ordinary databases can now search embeddings too, so the term names a job as much as a product.

What does searching a vector database look like?

Stay with the handbook. Suppose, for this example, three of its passages are stored, each tagged with its file and date: the remote-work sentence, "Annual leave is requested from your manager in advance" and "Expenses are repaid once a receipt is submitted".

You ask "Can I do my job from home?" The remote-work passage comes back first, leave second and expenses last, usually with a closeness score each. Add the filter "files dated this year" and older passages drop out first.

Now ask "Is there a dress code?" The handbook says nothing on it, yet a ranked list still comes back. Only a lower score hints that nothing fits, and a tool acts on that only if set up to.

What is a vector database used for?

  • Looking things up by idea. A shared drive of meeting notes, where you remember what was decided but not how anyone phrased it.
  • "More like this." Show the tickets that resemble the one open now, or flag a near-duplicate before it is filed.
  • Answering from your files. Many tools that answer questions about your documents search first and pass only the best few passages to the model that writes the reply. RAG vs. long context compares that with handing the model everything.
  • Remembering between chats. An assistant that carries notes between chats may keep them in a store like this and fetch the relevant ones later. AI memory vs. context separates that from the current conversation.

What are the limits of a vector database?

  • Nearest is not correct. The closest passage is the most similar to your question, which is not the same as right. "Staff may work remotely" and "Staff may not work remotely" are about the same subject and can rank side by side.
  • Splitting costs context. Documents are cut into passages before storing, often called chunking. An exception left in the paragraph after its rule, or a table row parted from its headings, can sit in an entry that never matches the question that needs it.
  • Only what went in can come out. A password-protected file or an image of a table may never have been stored, so "no result" can mean "never stored".
  • Old entries linger. Edit a travel-expense rule and the earlier passages stay until replaced. Old and new wording read almost alike, and copies of one passage can crowd the top places.
  • Stored numbers can leak. Research has shown that some of the original wording can sometimes be partly recovered from embeddings, so guard the store as you guard the documents.
  • Permissions must be enforced by the system. If one store holds both the public handbook and a salary sheet, who may see each passage must be checked before anything reaches the assistant. Telling the assistant "do not mention it" is a request, not a control: it has already seen the passage.
  • Deleting a document is not deleting its entries. The passages, their numbers and any copies in backups or caches all have to go. How to delete your data from an AI service covers what you can and cannot undo.

Do I need a vector database?

Often not. Stop at the first step that is enough.

  • Does it all fit in front of the assistant? If your documents fit in its context window, giving them whole is the simplest start: nothing is built and no search step can miss the page. The assistant can still overlook a detail in a long input. The context window explains the limit.
  • Does word search find what you need? If you know the terms, your viewer's find function or your computer's file search gives exact matches with nothing to maintain.
  • Do readers ask in their own words? When questions and documents use different words, search by meaning earns its keep. So does a collection too large to hand over.

Before building anything, have someone who has not read the documents write ten real questions, since people who have read them echo the wording, and note the passage that should answer each. Try plain word search on them. If it finds most, you probably do not need one; if the misses are paraphrases, search by meaning is worth building. After building, run the same ten, check the right passage lands in the top few, and rerun them whenever the documents change.

You may be using one already: tools that "chat with your files" can rely on one or something like it. Ask what is stored, for how long and how to delete it; the AI assistant privacy checklist lists the questions.

Where NAMELESS stands

NAMELESS is an AI model, still in training and not yet available, with no launch date, so nothing in this guide can be tried on it today. The guide describes vector databases in general, not how our model works. Our goal is a model on our own architecture, with unlimited context and no rereading of the conversation, extremely fast, with stronger logic and reasoning. None of that removes the need for the checks above: stored passages still have to be right, current and deleted when they should be, and an answer that matters should be checked against its source.

If that interests you, join the waitlist.

Keep reading

All posts