Add cuVS performance dashboard docs

fern/build_docs.sh

@@ -68,6 +68,10 @@ generate_api_reference() {
   python3 "${SCRIPT_DIR}/scripts/generate_api_reference.py"
 }
 
+export_performance_dashboard_data() {
+  python3 "${SCRIPT_DIR}/scripts/export_performance_dashboard_data.py"
+}
+
 run_checks() {
   pushd "${REPO_DIR}" >/dev/null
   run_fern check --warnings
@@ -78,24 +82,28 @@ run_checks() {
 case "${MODE}" in
   check)
     generate_api_reference
+    export_performance_dashboard_data
     run_checks
     ;;
   preview)
     generate_api_reference
+    export_performance_dashboard_data
     run_checks
     pushd "${REPO_DIR}" >/dev/null
     run_fern generate --docs --preview "$@"
     popd >/dev/null
     ;;
   publish)
     generate_api_reference
+    export_performance_dashboard_data
     run_checks
     pushd "${REPO_DIR}" >/dev/null
     run_fern generate --docs "$@"
     popd >/dev/null
     ;;
   dev)
     generate_api_reference
+    export_performance_dashboard_data
     pushd "${REPO_DIR}" >/dev/null
     run_fern docs dev "$@"
     popd >/dev/null

fern/docs.yml

@@ -47,6 +47,7 @@ css:
   - "./styles/metrics-table.css"
   - "./styles/notebook-viewer.css"
   - "./styles/trajectory-viewer.css"
+  - "./styles/performance-dashboard.css"
 experimental:
   mdx-components:
     - "./theme/nvidia/components"
@@ -96,6 +97,8 @@ navigation:
             path: "./pages/tuning_guide.md"
       - page: "Integrations"
         path: "./pages/integrations.md"
+      - page: "Performance"
+        path: "./pages/performance.md"
       - page: "Use-cases"
         path: "./pages/use_cases.md"
       - page: "Using cuVS APIs"

fern/pages/performance.md

@@ -0,0 +1,13 @@
+import PerformanceDashboard from "@/theme/nvidia/components/PerformanceDashboard";
+
+# Performance
+
+Compare cuVS-Bench performance across hardware and algorithms. Adjust the filters to see comparative performance for index build time, search throughput (QPS), and search latency.
+
+<PerformanceDashboard />
+
+## About this data
+
+These results come from [cuVS-Bench](cuvs_bench/index.md) runs on cuVS 26.04. The source dataset is [MIRACL](https://huggingface.co/datasets/miracl/miracl-corpus), embedded with [Llama Nemotron Embed 1B](https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2). Each row represents a tuned configuration bucketed by recall range (`90%`, `95%`, `99%`). Green bars denote GPU SKUs; blue bars denote CPU SKUs.
+
+To reproduce or extend these benchmarks, see the [cuVS Bench guide](cuvs_bench/index.md).

fern/scripts/export_performance_dashboard_data.py

@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Export cuVS bench CSV results to JSON and TypeScript for the Performance dashboard."""
+
+from __future__ import annotations
+
+import csv
+import json
+from pathlib import Path
+
+REPO_DIR = Path(__file__).resolve().parents[2]
+DEFAULT_CSV = REPO_DIR / "data" / "benchmarks" / "results_cuvs_26_04.csv"
+OUTPUT_JSON = REPO_DIR / "fern" / "assets" / "data" / "benchmark_results.json"
+BENCHMARK_DATA_TS = (
+    REPO_DIR / "fern" / "theme" / "nvidia" / "components" / "benchmarkData.ts"
+)
+
+NUMERIC_COLUMNS = {
+    "Index Build Time (s)",
+    "Search Batch Size",
+    "TopK",
+    "Mean Search Throughput (QPS)",
+    "Mean Search Latency (ms)",
+    "Mean Recall",
+    "N Points in Bucket",
+    "Total Vectors",
+    "Dimensions",
+}
+
+
+def parse_row(row: dict[str, str]) -> dict:
+    parsed: dict = {}
+    for key, value in row.items():
+        if key in NUMERIC_COLUMNS and value not in ("", "NA", None):
+            try:
+                parsed[key] = float(value) if "." in value else int(value)
+            except ValueError:
+                parsed[key] = value
+        else:
+            parsed[key] = value
+    return parsed
+
+
+def write_benchmark_data_ts(rows: list[dict], output_path: Path) -> None:
+    output_path.write_text(
+        """/**
+ * SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+export type BenchmarkRow = Record<string, string | number | null>;
+
+export const BENCHMARK_ROWS: BenchmarkRow[] = """
+        + json.dumps(rows, indent=2)
+        + ";\n",
+        encoding="utf-8",
+    )
+
+
+def export(csv_path: Path = DEFAULT_CSV, output_path: Path = OUTPUT_JSON) -> int:
+    if not csv_path.is_file():
+        raise FileNotFoundError(f"Benchmark CSV not found: {csv_path}")
+
+    with csv_path.open(newline="", encoding="utf-8") as handle:
+        rows = [parse_row(row) for row in csv.DictReader(handle, delimiter="\t")]
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(json.dumps(rows, indent=2), encoding="utf-8")
+    print(f"Wrote {len(rows)} rows to {output_path}")
+
+    write_benchmark_data_ts(rows, BENCHMARK_DATA_TS)
+    print(f"Wrote {len(rows)} rows to {BENCHMARK_DATA_TS}")
+
+    return len(rows)
+
+
+def main() -> int:
+    export()
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())