How knowmind finds the right passage
knowmind searches your memories on three levels at once — for exact keywords, for meaning, and across the network of connections. That way it also finds passages you only vaguely remember, and returns them with a reference to their source.
Most AI retrieval systems work with a single method. Either they match words literally, and then they miss everything phrased differently. Or they search purely by meaning, and then they lose exact names, contract numbers and proper nouns. knowmind combines both and adds a third: the links between related memories. The quality we measure against single-method retrieval comes out of that interplay.
What you search for literally, you find right away
knowmind first matches the terms in your query word for word — proper names, contract numbers, technical designations, everything you remembered precisely is found reliably. This stage is fast and returns the obvious hits.
Even when you called it something else back then
In parallel, knowmind looks for passages that are related in substance. If you ask about “that thing with the delivery date”, the second stage also finds notes about “the deadline in the contract” or “shipping in May” — even when not a single word of your question appears in them.
You see what the matter is connected to
Once the first hits are in, knowmind looks at which memories are linked to them — same person, same project, same case. So you get the passage you were after plus the context that a complete answer needs.
The result is a ranking you can account for
The three levels end in a single, sorted result list. Memories that show up on several levels rank higher — they are relevant for more than one reason. Passages found by only one level appear further down and stay reachable. What you get is a ranking in which every finding has its place for a reason.
Technical detail
- Average response time under 300 milliseconds
- Multilingual (German and English can appear in one query)
- Reference to the original note with every hit
- Configurable number of hits per query (five by default, 25 maximum)
- Search is always strictly tenant-separated — no hits from other workspaces