Semantic Search

The pipeline

Semantic search in primer is a small pipeline of independently-configured pieces:

1. An embedding provider turns text into vectors.
2. A semantic search provider is the vector store that holds those vectors and returns the nearest neighbours for a query.
3. A collection ties an embedder and a vector store together and holds the documents you ingest. Documents are chunked, embedded, and stored on ingest.
4. At query time the embedder vectorises the query, the store returns top-k nearest chunks, and an optional cross-encoder reranks those candidates for precision.

Internal collections apply the same machinery to primer's own catalogue (agents, providers, and the like) so agents can search the platform itself.

Register embedding providers that turn text into vectors.

Register the vector store that holds embeddings and serves nearest-neighbour queries.

Create collections, ingest documents, and run searches over them.

Add a cross-encoder reranker to refine the top-k results.

Make primer's own catalogue searchable as internal collections.