GeezerVison

A low-latency, billion-scale, and updatable graph-based vector store on SSD.

✨ Key Features

📊 Performance Comparison

🔌 Already on Milvus? Switch in one line

# In-process, drop-in for pymilvus.MilvusClient (URI is a directory):
from pipeann import MilvusClient
client = MilvusClient(uri="./pipeann-data")

# Or talk to the PipeANN gRPC server with the stock pymilvus client:
from pymilvus import MilvusClient
client = MilvusClient(uri="http://localhost:19530")   # PipeANN server, same wire protocol

client.create_collection("demo", dimension=128, metric_type="L2")
client.insert("demo", [{"id": 1, "vector": [0.1] * 128, "color": "red"}])
client.create_index("demo")   # builds the SSD graph; lazy/no-op for the in-process client
client.search("demo", data=[[0.1] * 128], filter="color == 'red'", limit=5)

🚀 Quick Start

🏗️ Build

# Ubuntu >= 22.04
# libmkl could be replaced by other BLAS libraries (e.g., openblas).
sudo apt install make cmake g++ libaio-dev libgoogle-perftools-dev \
                 clang-format libmkl-full-dev libeigen3-dev

# gRPC, Protobuf, and RocksDB — required by the Milvus-compatible layer
# (the gRPC server *and* the in-process MilvusClient, which share the same
# RocksDB-backed collection engine). The Python module links RocksDB, so
# these are needed even if you only use the Python interface.
sudo apt install libgrpc++-dev protobuf-compiler-grpc libprotobuf-dev \
                 protobuf-compiler librocksdb-dev

# For Python interface
pip install "pybind11[global]"

# Build liburing
cd third_party/liburing
./configure && make -j
cd ../..

# For C++ users: build C++ binaries under build/
bash ./build.sh

# For Python users: build and install the Python interface.
# This also builds and bundles the Milvus-compatible gRPC server binary,
# so `pipeann-server` is available straight after install.
pip install -e .

Milvus-compatible gRPC server

pipeann-server --data_dir ./data --port 19530 --threads 8

cmake -B build -DBUILD_MILVUS_SERVER=ON . && cmake --build build --target pipeann_milvus_server

⚡ C++

# Search an existing on-disk index with PipeANN pipelined search
build/tests/search_disk_index uint8 index_prefix 1 32 query.bin gt.bin 10 l2 pq 2 10 10 20 30 40

🐍 Python

from pipeann import IndexPipeANN, Metric

idx = IndexPipeANN(data_dim=128, data_type='float32', metric=Metric.L2)
idx.omp_set_num_threads(32)
idx.set_index_prefix(index_prefix)
idx.add(vectors, tags)                          # insert (auto disk-convert at 100K)
ids, dists = idx.search(queries, topk=10, L=50) # search
idx.save(index_prefix)                          # persist

📰 Updates

📖 Citation

@misc{arxiv26pipeannfilter,
      title={PipeANN-Filter: An Efficient Filtered Vector Search System on SSD}, 
      author={Hao Guo and Jiwu Shu and Youyou Lu},
      year={2026},
      eprint={2605.17992},
      archivePrefix={arXiv},
      primaryClass={cs.OS},
      url={https://arxiv.org/abs/2605.17992}, 
}

@inproceedings{fast26odinann,
  author    = {Hao Guo and Youyou Lu},
  title     = {OdinANN: Direct Insert for Consistently Stable Performance 
               in Billion-Scale Graph-Based Vector Search},
  booktitle = {24th USENIX Conference on File and Storage Technologies (FAST 26)},
  year      = {2026},
  address   = {Santa Clara, CA},
  pages     = {133--147},
  publisher = {USENIX Association}
}

@inproceedings{osdi25pipeann,
  author    = {Hao Guo and Youyou Lu},
  title     = {Achieving Low-Latency Graph-Based Vector Search via 
               Aligning Best-First Search Algorithm with SSD},
  booktitle = {19th USENIX Symposium on Operating Systems Design and Implementation (OSDI 25)},
  year      = {2025},
  address   = {Boston, MA},
  pages     = {171--186},
  publisher = {USENIX Association}
}