Kayak

Exact late-interaction retrieval for Python. Kayak gives you an explicit surface for exact late-interaction search backed by a real Mojo CPU path, and a practical way to keep your current vector database as the durable system of record.

Get started Quickstart Python API

Common workflows

Local text retrieval — input starts as raw text, one object owns encoding and search. → Text Encoders, Usage Patterns

Exact search from vectors — you already own token-level query and document vectors, search directly. → Quickstart, Python API

Database handoff — LanceDB, PgVector, Qdrant, Weaviate, or Chroma already owns persistence; Kayak owns the search step. → Storage + Search, Vector Databases

Repeated-query serving — the slice stays fixed, query traffic is the variable. → Usage Patterns, Hosted Engine Python

Where to start

If you want to...	Open...
choose by situation instead of API name	Choose Your Path
install Kayak and verify the exact backend	Installation
run the shortest possible exact search	Quickstart
pass a Hugging Face ColBERT checkpoint or your own model	Text Encoders
keep an existing vector database	Storage + Search
compare adapter-specific behavior	Vector Databases
open runnable notebooks and scripts	Examples
inspect the stable public signatures	Python API

Start with Installation → Quickstart → Usage Patterns.