Semantic Search API

A reference implementation of semantic search the way we build it for clients: not a thin wrapper over a vector database, but a service with retrieval quality treated as a measurable, defended metric.

What it demonstrates

• Hybrid retrieval — dense embeddings combined with keyword scoring, so exact matches never get lost.
• Re-ranking — a second-stage cross-encoder that reorders candidates for precision.
• Evaluation endpoint — recall@k and MRR computed against a labelled set, so quality is observable, not assumed.
• Clean API surface — a typed FastAPI service ready to drop behind any product.

Status

This is an active build and will ship with a live demo you can query directly. Want it applied to your corpus? Start a conversation.