feat(zvec): add DiskANN benchmark support

README.md

@@ -626,6 +626,16 @@ Options:
 
 ### Run Zvec from command line
 
+HNSW remains the default. For both HNSW and DiskANN, the distance metric is determined by the selected dataset during task
+assembly, not by a Zvec CLI option. For example, `Performance768D10M` uses COSINE.
+
+In the run-test UI, select Zvec and choose HNSW or DISKANN in the case settings. Only parameters for the selected algorithm
+are displayed. Load cases expose build parameters; performance and streaming cases also expose search parameters. Defaults
+match the CLI. Use separate collection paths for HNSW and DiskANN, and keep DiskANN build parameters unchanged for search-only
+runs.
+
+Existing HNSW commands are unchanged:
+
 ```bash
 vectordbbench zvec --path Performance768D10M --db-label 16c64g-v0.1 \
     --case-type Performance768D10M --num-concurrency 12,14,16,18,20 \
@@ -645,6 +655,68 @@ To list the options for zvec, execute vectordbbench zvec --help
                                   results
 ```
 
+#### Zvec DiskANN
+
+DiskANN requires `zvec>=0.7.0`, which provides `Collection.close()`. The SDK must also expose
+`DiskAnnIndexParam` and `DiskAnnQueryParam`.
+This adapter uses zvec's native C++ DiskANN, not the Microsoft Rust implementation.
+From this updated checkout on Linux, install the adapter with its Zvec and test dependencies, then run the checks:
+
+```bash
+python -m pip install -e ".[zvec,test]"
+vectordbbench zvec --help
+python -m pytest tests/test_zvec.py -q
+ZVEC_NATIVE_TESTS=1 python -m pytest tests/test_zvec.py -k native_diskann -v
+```
+
+The `zvec` extra enforces the SDK minimum. `--no-deps` bypasses dependency resolution; use it only after
+independently verifying the installed SDK version and required APIs.
+
+The native test builds 1,200 32-dimensional vectors in a temporary directory, checks insertion and index completion,
+closes/reopens the collection, searches it, and exercises rebuilding. It is a functionality test, not a recall or
+performance benchmark. It is skipped unless explicitly enabled on Linux. Run it successfully before a large benchmark.
+
+| Option | Default | Meaning |
+| --- | --- | --- |
+| `--index-type` | hnsw | Select hnsw or diskann |
+| `--max-degree` | 64 | DiskANN graph degree, 1-100 |
+| `--build-list-size` | 100 | Construction candidate list, 10-100 |
+| `--search-list-size` | 300 | Query candidate list, at least 1; the engine uses at least top-k |
+| `--pq-chunk-num` | 0 | PQ bytes per vector; 0 is automatic, otherwise at most min(dimension, 1024) |
+
+DiskANN stores FP32 vectors and uses internal PQ for graph traversal. HNSW options (`--m`, `--ef-construction`,
+`--ef-search`, `--quantize-type`, `--is-using-refiner`) are rejected in DiskANN mode, and DiskANN options are rejected
+in HNSW mode. PQ chunks are not equivalent to HNSW INT8 scalar quantization.
+
+After the checks pass, build a separate DiskANN collection. The load stage includes `optimize()`, which builds the
+target index; flushing alone does not establish that a DiskANN index was built. `--drop-old` replaces any existing
+DiskANN collection at this path. The adapter refuses to use or delete an existing non-DiskANN collection.
+
+```bash
+export DATASET_LOCAL_DIR="$PWD/dataset"
+vectordbbench zvec --index-type diskann --path Performance768D10M_DiskANN \
+  --case-type Performance768D10M \
+  --max-degree 64 --build-list-size 100 --pq-chunk-num 96 \
+  --drop-old --load --skip-search-serial --skip-search-concurrent
+```
+
+Then search the same index without reloading or rebuilding it:
+
+```bash
+vectordbbench zvec --index-type diskann --path Performance768D10M_DiskANN \
+  --case-type Performance768D10M \
+  --max-degree 64 --build-list-size 100 --pq-chunk-num 96 \
+  --search-list-size 300 --num-concurrency 12,14,16,18,20 \
+  --skip-drop-old --skip-load --search-serial --search-concurrent
+```
+
+Search-only runs validate the stored dimensions, metric, build parameters, positive document count and complete
+indexing. Keep build parameters unchanged when reopening; sweep `--search-list-size` to compare QPS at matched recall.
+These examples use the Cohere dataset's COSINE metric for both building and searching. For an HNSW/DiskANN
+comparison, use the same dataset and separate collection paths; do not relabel an existing IP index as COSINE.
+Multiple search processes can duplicate in-memory PQ codes and metadata, so measure total process memory as well
+as QPS.
+
 ### Run Doris from command line
 
 Doris supports ann index with type hnsw from version 4.0.x

pyproject.toml

@@ -88,7 +88,7 @@ alisql          = [ "mysqlclient" ]
 polardb         = [ "PyMySQL" ]
 doris           = [ "doris-vector-search" ]
 turbopuffer     = [ "turbopuffer" ]
-zvec            = [ "zvec" ]
+zvec            = [ "zvec>=0.7.0" ]
 endee           = [ "endee==0.1.10" ]
 lindorm         = [ "opensearch-py" ]
 seekdb          = [ "mysql-connector-python" ]