Semantic search
Semantic search ranks results by similarity of meaning rather than by shared words, usually by converting both the query and the documents into embeddings — numeric vectors — and returning the passages whose vectors lie closest to the query’s.
- Also called
- Vector search, embedding search
- Good at
- Same idea, different words
- Weak at
- Exact terms, names, numbers
What it does well
Keyword search needs the words. If your notes say “stormwater retention” and you search for “how cities cope with heavy rain”, keyword search finds nothing. Semantic search can, because it compares what the passages are about rather than which characters they contain. It also crosses languages: a question in English can surface a passage written in German.
That is a real advantage, and it is why semantic search became the default retrieval step in most AI tools that answer from documents.
What it misses on your own files
Public web search and personal files are different problems. In your own material you usually remember something specific — a term, a name, a number, a phrase — and that is exactly where similarity is weakest:
- Exact strings blur. “Section 4.2(b)”, a statute number, a drug name, a variable — an embedding places them near things that look similar, not things that are identical. The passage with the exact clause can rank below three passages that merely discuss the topic.
- Misses are silent. Results come back ranked by closeness, with no sign of what fell just below the cut. As with RAG, a good result list and an incomplete one look the same.
- It needs an index. Every document has to be embedded before it can be found, and re-embedded when it changes — a build step, a wait, and something to keep current.
- Counting is not its job. “Every place this term appears” is a question about words, and a similarity ranking cannot answer it.
- The same idea in different words
- A question asked in another language
- Exact strings — a clause number, a name, a variable
- Any signal of what it ranked out
- Anything not yet indexed
- Exact terms, names and numbers
- Every occurrence, countable
- The same idea in different words
- A question asked in another language
Each approach has a blind spot. Which one matters depends on whether you remember the idea or the words.
How MindMax handles it
By not putting a vector index in the middle. There is no vector database behind the search. It finds material the way you would — by structure, by the words on the page, and by trying other wordings — so each answer traces to a passage you can open. The trade-off: a question phrased very differently from your files, or asked in another language, can miss.
In the search panel, one search returns two lists side by side: literal matches, counted across the folder, and relevant passages ranked by how much of your query they cover — without needing your exact phrase or word order. When you want judgement by meaning, the AI re-picks: it reads the candidates and keeps the ones that fit what you meant, and it does this on its own when a question finds nothing that clearly matches. Because it chooses among what the wording turned up, a passage that shares no words at all with your question can still be missed. Every result opens on the sentence it found.
Relevant passages and literal matches in one panel, each counted across files — with the AI re-pick one click away.
Common questions
What is the difference between semantic search and keyword search?
Keyword search matches the words you typed; semantic search matches meaning, usually by comparing embeddings. Keyword search is better for exact terms, names and numbers and can count every occurrence; semantic search is better when you remember the idea but not the wording.
Do I need a vector database to search my documents by meaning?
Not necessarily. MindMax has no vector database: it ranks passages by the words they share with your query, then lets the AI re-pick candidates by meaning. The trade-off is that a question phrased very differently from your files, or in another language, can miss.
Is semantic search more accurate?
It is better at paraphrase and worse at exact strings, and neither tells you what it ranked out. For your own files, where you often remember a specific term, having literal matches alongside is usually what makes a search trustworthy.