One binary handles dense, sparse, tensor, and full-text search

What it does

Infinity is a C++20 database that stores and queries structured data alongside dense vectors, sparse vectors, tensors, and full-text. It targets retrieval-augmented generation and other LLM workloads by offering a unified search backend instead of forcing developers to wire up separate vector and text engines. The project ships as one self-contained binary with no external dependencies and exposes a Python API.

The interesting bit

The “hybrid search” pitch is common, but Infinity also treats tensors—multi-vector representations like ColBERT embeddings—as a first-class data type with native reranking support. That suggests it is trying to absorb newer retrieval models, not just classic k-NN and BM25.

Key highlights

• Self-reported latency of 0.1 ms and 15K+ QPS on million-scale vector datasets, and 1 ms with 12K+ QPS on 33M-document full-text corpora.
• Supports hybrid queries across all four search types—dense, sparse, tensor, and full-text—with built-in filtering and rerankers including RRF, weighted sum, and ColBERT.
• Runs as a standalone server or embedded directly in Python; the server is a single dependency-free binary.
• Requires x86_64 with AVX2; compatible with Linux, macOS, or Windows via WSL2.

Verdict

Teams building RAG pipelines who are tired of gluing together a vector database, a sparse retrieval engine, and a text search index should evaluate this. If you only need simple vector similarity or already run a heavily customized search stack, it may be overkill.