JingsongLi opened a new pull request, #126:
URL: https://github.com/apache/paimon-vector-index/pull/126

   DiskANN currently cannot serve distance-band queries through the IVF range 
API. This adds dedicated single-query, batch and Roaring-filtered DiskANN range 
APIs in Rust, C, C++, Java/JNI and Python.
   
   The independent implementation follows Microsoft's [two-stage range-search 
approach](https://github.com/microsoft/DiskANN/blob/a3e8963c204c631ad09cac155bf6785ad4e5ba8e/diskann/src/graph/search/range_search.rs).
 It first converges a bounded candidate set, then expands a radius-based 
frontier. Final membership uses the requested half-open band and persisted 
F32/F16 vector distances; filters and lower bounds do not block graph 
navigation. Graph discovery remains approximate, with an explicit exhaustive 
mode and automatic scanning for unbounded upper cuts and zero cosine queries.
   
   Range queries use separate traversal scratch and bounded query-local window 
caches, batch known I/O windows within adapter limits, and return their own 
diagnostics. Existing Top-K traversal, caches and statistics are preserved, 
including the new unified search statistics. Existing IVF parameters, C ABI 
layouts, C++ result aggregates and Python result dataclass fields remain 
compatible. The index format is unchanged.
   
   Validation on latest main (`7b4b19b`):
   
   - Rust workspace: 751 passed, 2 existing ignored; build, Clippy with 
warnings denied, and formatting checks passed.
   - Python: 599 passed. C/C++ strict builds with UBSan and Java/JNI tests 
passed; each language passed the shared 168-case range oracle.
   - Regression coverage includes uncapped results, 
metric/layout/PQ/vector-encoding combinations, graph bridges through excluded 
nodes, numerical errors, working-memory limits, partial I/O failures, bounded 
remote read batching, legacy result compatibility, and preservation of Top-K 
results/statistics.
   - ASF source-header and binary legal-inventory checks passed.
   
   Performance evidence: before rebasing onto latest main, an 8-round 
alternating comparison against the original base (`f6a5558`) reused identical 
serialized index bytes and a fixed read plan. Recall was identical; median 
single/batch Top-K QPS changed by +3.7%/+3.1%, with no observed regression. 
These local measurements do not establish a remote-storage latency guarantee. 
Added a reproducible range recall/I/O and Top-K interleaving benchmark.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to