InfinityFlow
InfinityFlow: Infinity is an AI-Native database for LLM applications, offering top-notch performance and flexibility to meet future AI application challenges. It is known for...
Last verified:
What is InfinityFlow?
InfinityFlow is an AI-native database built specifically for LLM applications, providing incredibly fast hybrid search capabilities across dense vectors, sparse vectors, tensors, and full-text data. It was officially open-sourced on December 21, 2023, and is primarily designed for Retrieval-Augmented Generation (RAG) applications. The database supports a wide range of data types including strings, numerics, vectors, and structured data, making it versatile for diverse AI workloads.
InfinityFlow pricing
Pricing model: Free
Infinity is fully open-source and free for any purpose under the Apache-2.0 license. The Community Edition is released with Apache2 license, meaning FREE for any purpose including commercial and non-commercial use without fees. Users can self-host the entire database on their own infrastructure with no vendor lock-in or per-query charges. The software can be deployed via Docker using the nightly image (infiniflow/infinity:nightly) or installed as a binary. The client SDK is available via pip: infinity-sdk==0.6.0.
InfinityFlow pros
- 0.1ms query latency on million-scale vector datasets
- 15K+ QPS on million-scale vector datasets
- 1ms latency and 12K+ QPS for full-text search on 33M documents
- Hybrid search of dense embedding, sparse embedding, tensor, and full text
- Supports multiple rerankers: RRF, weighted sum, and ColBERT
- Rich data types including strings, numerics, vectors, and structured data
- Intuitive Python API for easy development
- Single-binary architecture with no dependencies
- Easy deployment via Docker
- Can be embedded in Python as a module
- Open-source under Apache-2.0 license
- Free for any purpose with no licensing fees
- Designed specifically for RAG applications
- Supports filtering in hybrid search
- No vendor lock-in when self-hosted
- 4.1k GitHub stars showing strong community adoption
- Available on Linux, Windows, and MacOS
- HTTP API support for flexibility
InfinityFlow cons
- Requires x86_64 CPU with AVX2 support
- Linux requires glibc 2.17+
- Windows requires WSL/WSL2 for Docker deployment
- Python 3.10+ required
- Limited to 31 releases so still maturing
- Only 48 contributors compared to larger databases
- C++ 87% of codebase makes it harder for Python developers to contribute
- No official cloud-hosted managed service mentioned
Frequently asked questions about InfinityFlow
What is Infinity database?
Infinity is an AI-native database built specifically for LLM applications, providing incredibly fast hybrid search of dense vector, sparse embedding, tensor (multi-vector), and full-text data. It was officially open-sourced on December 21, 2023, and is primarily used for Retrieval-Augmented Generation (RAG) applications including search, recommenders, question-answering, conversational AI, copilot, and content generation.
Is Infinity free to use?
Yes, Infinity is fully open-source and free for any purpose. It is released under the Apache-2.0 license, which means it is FREE for commercial and non-commercial use without any licensing fees. Users can self-host the database on their own infrastructure with complete control over their data and costs.
How fast is Infinity database?
Infinity achieves incredibly fast performance with 0.1 milliseconds query latency and 15K+ QPS on million-scale vector datasets. For full-text search, it achieves 1 millisecond latency and 12K+ QPS on 33M documents. This performance is comparable to single retrieval libraries while providing much richer search capabilities.
What search capabilities does Infinity support?
Infinity supports hybrid search combining dense embedding, sparse embedding, tensor, and full-text search, plus filtering. It includes multiple reranking methods: RRF (Reciprocal Rank Fusion), weighted sum, and ColBERT reranker. This makes it the most comprehensive multi-way retrieval solution for RAG applications.
What data types does Infinity support?
Infinity supports a wide range of data types including strings (varchar), numerics (integer, float), dense vectors, sparse vectors, tensors (multi-vector), and structured data. This richness allows users to store and search diverse data types in a single database.
How do I install Infinity?
Infinity can be deployed using Docker by pulling the nightly image: docker pull infiniflow/infinity:nightly, then running it with docker run -d --name infinity -v /var/infinity/:/var/infinity --ulimit nofile=500000:500000 --network=host infiniflow/infinity:nightly. The Python client is installed via pip: pip install infinity-sdk==0.6.0. Prerequisites include x86_64 CPU with AVX2, Python 3.10+, and supported OS (Linux with glibc 2.17+, Windows 10+ with WSL/WSL2, or MacOS).
What is Infinity's Python API like?
Infinity provides an intuitive Python API that is easy to use. You can connect using infinity.connect(), get or create databases, create tables with various column types, insert data, and perform vector search, full-text search, or hybrid search. The database can also be embedded directly in Python as a module, making it friendly for AI developers.
Is Infinity just another vector database?
No, Infinity is not just another vector database. While it supports dense vector search, it also provides sparse vector search (which can replace BM25/full-text search), tensor search, and full-text search. This multi-path recall capability makes it the most powerful database for RAG, combining multiple retrieval methods that traditional vector databases don't offer together.
What is the difference between dense and sparse vectors in Infinity?
Dense Embedding refers to vectors where dimensions may not be very high but each dimension has a numerical weight value. Sparse Embedding refers to vectors where most dimensions are zero with only a few dimensions having values, allowing very high overall dimensions. Sparse vectors provide precise queries necessary for RAG and can completely replace BM25/full-text search, enabling multi-path recall when combined with dense vectors.
How does Infinity's architecture differ from data warehouses?
Infinity has sophisticated designs at both storage engine and execution engine levels. Its execution engine compiles execution plans into pipeline execution plans after binding API queries. Unlike modern data warehouses where pipelines are designed mainly for parallel query execution, Infinity's pipeline serves both parallel querying and concurrent query execution to optimize scheduling strategies and CPU affinity for query operators during high-concurrency execution, avoiding overhead from invalid context switches.