example(benchmark): add VIBE integration and performance evaluation suite

- Implement reproducible ANN benchmark pipeline comparing DEG configurations
- Measure Pareto-optimal QPS vs Recall trade-offs across standardized datasets
- Include plotting scripts and result serialization

Author: Neiko2002

examples/README.md

@@ -43,5 +43,7 @@ uv sync --reinstall-package deglib
 - [`knng`](./knng/): k-Nearest Neighbor Graph (k-NNG) construction benchmark using EVP quantization and FP16 reranking (SISAP 2026 Challenge Task 1).
 - [`mips`](./mips/): Maximum Inner Product Search (MIPS) benchmark using $(d+1)$-dimensional $L_2$ transformation, FLAS pre-sorting, and SIMD FP16 inner products (SISAP 2026 Challenge Task 2).
 - [`static_data`](./static_data/): DEG paper search benchmark reproduction (Recall vs. QPS) on static datasets (`sift1m`, `deep1m`, `glove-100`, `audio`, `enron`).
+- [`vibe`](./vibe/): Vector Index Benchmark for Embeddings (VIBE) ANNS top-100 benchmark on modern embedding datasets (`agnews-mxbai`, `arxiv-nomic`, `landmark-dino`, `msmarco-qwen`, `gooaq-distilroberta`, `laion-clip`, `imagenet-align`, `imagenet-clip`, `yandex`, `yahoo-minilm`).
 - [`dynamic_data`](./dynamic_data/): DEG dynamic data streaming benchmark (`AddHalf`, `AddHalfRemoveAndAddOneAtATime`, `AddAllRemoveHalf`).
 - [`sliding_window`](./sliding_window/): DEG sliding window benchmark reproducing the dynamic continuous update experiment against DEG from the CleANN paper.
+

examples/vibe/README.md

@@ -0,0 +1,60 @@
+# VIBE ANNS Benchmark Reproduction with DEG
+
+This example reproduces the Approximate Nearest Neighbor Search (ANNS) experiments from the **[VIBE (Vector Index Benchmark for Embeddings)](https://vector-index-bench.github.io/)** project using the Dynamic Exploration Graph (DEG).
+
+VIBE provides realistic embedding datasets covering in-distribution and out-of-distribution (OOD) search across text, vision, and multi-modal models with exact pre-computed Top-100 ground truth neighbors.
+
+## Supported VIBE Datasets
+
+| Dataset Key | VIBE Name | Type | Size ($N$) | Dimension ($D$) | Metric |
+| :--- | :--- | :--- | :--- | :--- | :--- |
+| `agnews-mxbai` | AGNews-mxbai | In-Distribution | 120,000 | 1,024 | L2 (Euclidean) |
+| `arxiv-nomic` | ArXiv-nomic | In-Distribution | 2,000,000 | 768 | Inner Product |
+| `landmark-dino` | Landmark-dino | In-Distribution | 760,757 | 768 | Cosine |
+| `msmarco-qwen` | MSMARCO-qwen | In-Distribution | 8,841,823 | 1,024 | Inner Product |
+| `gooaq-distilroberta`| GooAQ-distilroberta | In-Distribution | 1,471,375 | 768 | Inner Product |
+| `laion-clip` | LAION-clip | Out-of-Distribution | 1,000,000 | 512 | Inner Product |
+| `imagenet-align` | ImageNet-align | Out-of-Distribution | 1,281,167 | 640 | Inner Product |
+| `imagenet-clip` | ImageNet-clip | In-Distribution | 1,281,167 | 512 | Inner Product |
+| `yandex` | Yandex-200 | Out-of-Distribution | 1,000,000 | 200 | Cosine |
+| `yahoo-minilm` | Yahoo-minilm | In-Distribution | 677,305 | 384 | Inner Product |
+
+## Quick Start
+
+### 1. Environment Setup
+
+```bash
+cd examples/vibe
+uv sync
+```
+
+### 2. Run Benchmark
+
+Datasets are automatically downloaded on demand directly from the official [VIBE Hugging Face repository](https://huggingface.co/datasets/vector-index-bench/vibe) and stored in `D:/Data/VIBE` (or `~/.cache/vibe` / `VIBE_CACHE_DIR`).
+
+```bash
+# Run on AGNews-mxbai (compact dataset ~120k vectors)
+uv run main.py --dataset agnews-mxbai
+
+# Run on Yahoo-MiniLM (677k vectors)
+uv run main.py --dataset yahoo-minilm
+
+# Run without GUI plot popups
+uv run main.py --dataset arxiv-nomic --no-show
+```
+
+### 3. Command-Line Options
+
+```bash
+uv run main.py --help
+```
+
+- `--dataset`, `-d`: Dataset name to benchmark (e.g. `agnews-mxbai`, `arxiv-nomic`, `landmark-dino`, `msmarco-qwen`, `gooaq-distilroberta`, `laion-clip`, `imagenet-align`, `imagenet-clip`, `yandex`, `yahoo-minilm`).
+- `--cache-dir`, `-c`: Custom directory for dataset files and saved `.deg` graphs.
+- `--build-threads`, `-t`: Number of CPU threads for graph building (default: half of CPU cores, `threads // 2`).
+- `--k`: Graph degree $k$ (out-degree per vertex).
+- `--extend-k`: Exploration width during graph building.
+- `--eps`: Build $\varepsilon$ parameter.
+- `--anns-k`: Number of nearest neighbors to evaluate (default: 100).
+- `--rebuild-graph`: Force graph re-construction even if a cached graph file exists.
+- `--no-show`: Do not open interactive matplotlib plot window.

examples/vibe/config.yml

@@ -0,0 +1,45 @@
+float:
+  any:
+  - base_args: ['@metric']
+    constructor: DegANN
+    disabled: false
+    gpu: false
+    ood: false
+    singularity_image: deg
+    module: vibe.algorithms.deg
+    name: deg
+    run_groups:
+      base:
+        args:
+          k: [16, 24, 30, 40, 48]
+          opt_target: ['LowLID', 'HighLID', 'StreamingData']
+          extend_k: [60]
+          build_eps: [0.1]          
+          improve_k: [0]
+          improve_eps: [0.0]
+          threads: [1]
+        query_args:
+          search_eps: [0.0, 0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 0.7, 1.0, 1.5, 2.0]
+
+uint8:
+  euclidean:
+  - base_args: ['@metric']
+    constructor: DegANN
+    disabled: false
+    gpu: false
+    ood: false
+    singularity_image: deg
+    module: vibe.algorithms.deg
+    name: deg
+    run_groups:
+      base:
+        args:
+          k: [16, 24, 30, 40, 48]
+          opt_target: ['LowLID', 'HighLID', 'StreamingData']
+          extend_k: [60]
+          build_eps: [0.1]          
+          improve_k: [0]
+          improve_eps: [0.0]
+          threads: [1]
+        query_args:
+          search_eps: [0.0, 0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 0.7, 1.0, 1.5, 2.0]