| name | hipvs-ann |
| description | Selects, builds, benchmarks, and validates hipVS ANN indexes and query paths for ROCm-DS workloads. |
| tools | ["bash","python","git"] |
| inputs | ["dataset_shape","vector_dim","metric","recall_target","latency_target","gpu_facts"] |
| outputs | ["index_recommendation","benchmark_summary","validation_notes"] |
| tags | ["rocm","rocm-ds","hipvs","ann","vector-search"] |
Purpose
Choose the correct hipVS path for vector search workloads and validate that the selected ANN method is appropriate for the data, constraints, and hardware.
When to use
- Build or tune ANN/vector-search workflows on ROCm-DS
- Compare brute force, HNSW, IVF-Flat, IVF-PQ, CAGRA, or NN-Descent
- Validate query latency, throughput, and recall tradeoffs
- Design a GPU-native retrieval subsystem
When not to use
- The workload is ordinary tabular ETL with no vector retrieval requirement
- The target GPU is outside supported or realistically testable configurations and no experimental mode is explicitly allowed
- There is no benchmark dataset or success criterion
Required inputs
- Repository path
- Dataset shape and vector dimensionality
- Distance metric
- Recall / latency / throughput targets
- GPU and ROCm environment facts
- Whether source builds are allowed
Procedure
- Check compatibility assumptions first.
- Classify the workload:
- exact k-NN
- high-recall ANN
- memory-constrained ANN
- high-throughput batch query
- Pick candidate index families.
- Build the minimal reproducible benchmark.
- Compare build time, memory use, query latency, and recall.
- Recommend the least-complex path that satisfies the target.
- Document unsupported assumptions and fallback options.
Deliverables
- Recommended index type and rationale
- Build and runtime notes
- Benchmark summary
- Validation evidence
- Risks and fallback path