0.3.0

It adds new search capabilities first, then the Python-facing stack built on top of them, plus the refactors and tests needed to support updates and filtering consistently.

Feature highlights

Speculative filtering

• Added speculative filtered ANNS for arbitrary attribute constraints.
• Uses lightweight in-memory probabilistic filters to explore a superset of valid vectors, then verifies final candidates exactly against SSD-resident attributes.
• A cost model chooses speculative pre-filtering, speculative in-filtering, or post-filtering per query.
• Supports label filters, range filters [l, r), and Boolean combinations through native selectors.
• Added typed attributes (Attributes, AttrsVec), on-disk attribute indexes, dense-neighbor support (range_dense), and JSON filter config loading.

OOD refinement

• Added NGFix-style out-of-distribution graph refinement.
• Exposed train_query_path, R_ood, and L_ood in C++ build tools and IndexPipeANN.build().
• Persisted OOD metadata in SSD index metadata.

Range search

• Added finite-threshold range search in C++ and Python.
• Results outside the threshold are filtered out and padded with UINT32_MAX / inf.
• Reuses the common pipelined traversal and result-copy path.

SPDK backend

• Added an optional SPDK I/O backend through -DIO_ENGINE=spdk for raw NVMe vector reads.
• Supports RAID-0-style striping across PCIe NVMe devices listed in spdk_bdevs.json, with one poller thread per device.
• Copies {index_prefix}_disk.index to the SPDK target on first open and reuses a marker to skip repeated copies.
• Keeps filtered-search attribute reads on io_uring while vector I/O uses SPDK.

Python API and integrations

• Added the current IndexPipeANN(data_dim, data_type, metric) API for build, load, search, insert, delete, save, filters, range search, and attribute-aware inserts.
• Added Collection and Client for SQLite-backed documents/payloads, vector CRUD, persistence, and collection auto-discovery.
• Added LangChain integration through pipeann.langchain.PipeANNVectorStore.
• Added a Qdrant-compatible FastAPI server with collection management, point upsert/query/scroll, payload indexes, filter delete, count, and save.
• Added schema.json persistence for collection config and attribute-index metadata.

Code refactoring and implementation changes

• Merged dynamic search, insert, delete, merge, and save behavior into header-only DynamicIndex<T>.
• Unified update save/merge with same-prefix double-version replacement.
• Merged PiPNN and Vamana-style build paths behind build_disk_index and the shared SSD file format.
• Refactored pipelined top-k, range, and filtered search around pipe_search_common.h and spec_filter_search.cpp.
• Simplified metric/distance handling and updated PiPNN file layout.
• Added IO_ENGINE selection for uring, aio, and spdk, separate dense-node I/O sizing, CMake/CI cleanup, and package version 0.3.0.

Tests and examples

• Switched the project license from MIT to Apache License 2.0 and updated NOTICE attribution, including NGFix graph-refinement logic under MIT.
• Added Python insert/delete search tests, including assertions that results include inserted data and exclude deleted data.
• Added filtered insert regression tests comparing full build vs. build-plus-insert with attributes.
• Added native selector, range-search, collection, LangChain, and Qdrant server tests/examples.
• Added C++ filtered build/search and update test coverage.

Breaking changes

• DynamicSSDIndex was removed; use DynamicIndex<T> in C++ or IndexPipeANN in Python.
• dynamic_index.cpp was removed; DynamicIndex<T> is now header-only.
• filter/label.h was removed; use filter/attribute.h and filter/selector.h.
• build_pipnn_index was removed; use build_disk_index with PiPNN-compatible parameters.
• The old schema-centered Python API was removed; construct IndexPipeANN(data_dim, data_type, metric) directly and let Collection / Client own collection persistence.
• PyIndexInterface construction changed from a Python dict to explicit (dim, dtype, metric) arguments.
• IndexPipeANN.search() now takes selector, query_attrs, and finite range keyword arguments.
• IndexPipeANN.build() now takes attrs, range_dense, train_query_path, R_ood, and L_ood; major build knobs auto-configure when left at 0.
• pipnn.h / pipnn.cpp moved under the utils layout.

0.2.0

We heavily refactored the codebase and added some features.

Updates

• Python interface for index, search, and update.
• Modularized neighbor abstraction & RaBitQ support.
• Faster PQ generation in graph building.
• Filtered ANNS support (post-filtering).
• Reduced memory usage.
• More metrics supported (L2, cosine, and inner-product).
• Vector size >= 4KB supported (experimental)

0.1.1

OdinANN (FAST'26) AE version, a graph-based vector store supporting direct inserts as well as pipe search.

0.1.0

This is the artifact version of our OSDI '25 paper, PipeANN, a low-latency and large-scale vector search system.