When Not to Use a Vector Database

Vector databases are great at fuzzy similarity search over unstructured text. They are a poor default for everything else. Before you spin one up, check the list below.

Skip the vector DB when

• The lookup is by exact key. Use a hash table.
• The data is small enough to fit in the prompt. Use a static file.
• The data has a stable schema. Use a SQL index.
• The query is structured (range, prefix, regex). Use a database that supports it.

Use the vector DB when

• You have thousands of documents and no stable schema.
• Your query is "find me paragraphs that look like this."
• You can tolerate approximate answers and a small recall loss.

The hidden cost

Vector databases need embeddings, embeddings need a model, the model has a version, the version has a behavior change, and now you have a dependency chain that is hard to test. Plain keyword search has none of that.

A small hybrid

For most personal knowledge bases, full-text search plus a small reference file in the prompt is faster to ship, easier to reason about, and good enough 90% of the time.