More Than One Way to Ground
Vector search is the famous option, but it isn’t the only one. The right representation of your knowledge depends on your data and the questions people will ask of it.
Three Representations
Each stores knowledge in a different shape, and each answers a different kind of question well.
- Vector index — search by meaning and similarity
- Knowledge graph — entities and the relationships between them
- Semantic layer — a structured view over your database or records
Reach for a Vector Index
Use it when your data is unstructured text and the query is essentially “find me passages that are similar to this.”
Reach for a Graph or Layer
Use these when the answer needs to be exact and relational, not a fuzzy similarity match.
- Knowledge graph — connected facts: who reports to whom, which task links to which note
- Semantic layer — structured records you query precisely, the way you would query a database
Real Systems Mix Them
You rarely pick one and stop. A serious system often runs several representations at once — vector search over the prose, a structured query over the records. Pick per query type, not per fashion.
Build It
How to implement: classify your data. Is a typical question “find similar text,” “traverse relationships,” or “query structured records”? That answer picks the representation for you.
- Weekly AI Tasks tracker — tasks and notes are structured records with links, so a semantic layer over the database fits better than pure vectors, plus light vector search on note text.
- Personal brand site — mostly a small set of documents, so a simple vector index — or even including them directly — is enough.
What you learned
Three representations — vector index, knowledge graph, semantic layer. Vectors for similar text, graphs and layers for exact relational answers, and real systems mix them by query type.