Infinity
infiniflow.org
- Free plan
- Windows
- macOS
- Linux
- API
- Self-hosted
What Infinity is
Infinity is an open-source, AI-native database for LLM applications and retrieval-augmented generation. It provides hybrid search over dense embeddings, sparse embeddings, tensors, and full text, with metadata filtering. Supported data includes strings, numerics, and vectors, while listed rerankers include RRF, weighted sum, and ColBERT. The homepage claims 0.1 millisecond query latency and up to 15K QPS on million-scale vector datasets. Developers can use Python and HTTP APIs, with a Python SDK installed through pip. Deployment documentation covers Docker and Linux packages in deb, rpm, or tgz formats. Supported environments include Linux, macOS, and Windows 10 or later through WSL or WSL2, on x86_64 systems with AVX2 support; Python 3.10 or later is required. HNSW targets faster dense-vector search, while IVF uses less memory. Streaming vectors are searched by brute force alongside indexed results. Tensor search may create significant storage overhead, and full-text indexing requires a tokenizer or analyzer. The project uses the Apache-2.0 license.
Who it suits
It suits developers building retrieval-augmented generation or other LLM applications needing hybrid vector and text search. Teams comfortable with Docker, APIs, supported operating systems, and index configuration may find its deployment model appropriate.
Good to know, and what to watch
- Open-source Apache-2.0 project
- Combines dense, sparse, tensor, and full-text search
- Provides Python and HTTP APIs
- Supports Docker and Linux package deployment
- Includes several index and reranking options
- Python client and server versions must match exactly
- Tensor search can require significant storage
- Full-text indexing needs a tokenizer or analyzer
- Conditional filters must operate through an index
The verdict
Infinity brings hybrid retrieval, filtering, reranking, and multiple deployment paths to an open-source database for LLM workloads. Its documentation also highlights version, indexing, and storage considerations that teams should plan for.
What the maker publishes
- API options
- The docs provide Python and HTTP API references, and the homepage says Infinity has a Python API. source
- Cluster storage
- The cluster guide says supported shared storage is MinIO and describes a cluster with one leader, up to four followers, and learner nodes. source
- Community
- The homepage links to Infinity's GitHub repository and Discord community. source
- Community support
- The homepage links to the project’s X account, GitHub, and Discord community. source
- Data types
- Infinity supports data types including strings, numerics, and vectors. source
- Deployment
- The quickstart documents deploying Infinity with Docker, while the binary guide offers Linux packages in deb, rpm, or tgz formats. source
- Developer interface
- The homepage describes an intuitive Python API, and the docs provide a Python SDK and HTTP API reference. source
- Full-text search
- The search guide says full-text indexing requires specifying a tokenizer or analyzer, and most built-in analyzers require resource files. source
- Intended use
- The maker describes Infinity as designed for retrieval-augmented generation applications. source
- License
- The project’s GitHub repository lists the Apache-2.0 license. source
- Performance
- The homepage states 0.1 milliseconds query latency and up to 15K QPS on million-scale vector datasets. source
- Performance claims
- The homepage claims 0.1 millisecond query latency and up to 15K QPS on million-scale vector datasets. source
- Purpose
- Infinity describes itself as an AI-native database for LLM applications that provides hybrid search. source
- Python API
- The homepage describes an intuitive Python API, and the docs show installing the client with pip. source
- Ranking
- Its listed rerankers include RRF, weighted sum, and ColBERT. source
- Rerankers
- The homepage names RRF, weighted sum, and ColBERT as supported rerankers. source
- Search
- It supports hybrid search across dense embeddings, sparse embeddings, tensors, and full text, with filtering. source
- Search constraints
- The search guide says conditional filters must work through an index; secondary-index filters support logical comparisons and membership expressions. source
- Search methods
- It supports hybrid search across dense embeddings, sparse embeddings, tensors, and full text, as well as filtering. source
- Supported environments
- The quickstart lists Linux, macOS, and Windows 10 or later with WSL or WSL2; it specifies x86_64 with AVX2 support and Python 3.10 or later. source
- Tensor guidance
- The search guide cautions that tensor search can have significant storage overhead and recommends tensors for reranking in hybrid search. source
- Vector search
- The search guide says HNSW is aimed at faster dense-vector search and IVF uses less memory; incoming streaming vectors are searched by brute force alongside indexed results. source
- Version requirement
- The Python API documentation says client and server versions must match exactly in client-server mode. source
Checked against the maker's pages on 8 Oct 2026.