Vector Database.

The foundation is the embedding: a model translates every passage of text into a long sequence of numbers whose position in space represents its meaning. Related passages land close together, independent of the wording chosen. The database then measures the distance between the query vector and the stored vectors. Because comparing against every single entry would be too slow, vector databases use approximate methods such as HNSW: they do not guarantee the exactly most similar hits, but in practice they deliver the right ones very quickly — a deliberate trade of accuracy for speed.

In operation, three details decide quality. First, the split: sections that are too large blur the meaning, sections that are too small lose the context; what usually works is splitting along the editorial structure with a little overlap. Second, the metadata. Every section should carry its language, date, source and permissions, so that filtering can happen before the similarity search. Third, the combination with classic keyword search, because pure meaning-based search handles article numbers, proper names and case references poorly. This hybrid search with subsequent reranking of the best hits is the standard setup today.

A dedicated vector database is needed less often than the attention on the topic suggests. Up to a few hundred thousand sections, an existing relational database with a vector extension is frequently enough, which simplifies operations and data protection because no additional infrastructure appears. Specialised systems pay off with very large holdings, many concurrent queries or strict tenant separation. The real work lies elsewhere anyway: in keeping content current, splitting it cleanly, and answering the question of who is allowed to see which sections at all.

A case from a mid-sized business: an internal knowledge search should make three thousand documents from projects, contracts and manuals searchable. The order is decisive here — filter first by permission and currency, then search by similarity, otherwise the answer cheerfully quotes a contract the person asking should never have seen. Keyword search runs alongside so that project numbers match reliably. And there has to be a rule for outdated documents: what no longer applies must be removed from the holding, or it will be quoted convincingly.