[EXP] Update the benchmark code (#71)

* chore(hnsw): reorder imports to satisfy ruff I001

* chore: sync changes; fix Ruff import order; update examples, benchmarks, and dependencies

- Fix import order in packages/leann-backend-hnsw/leann_backend_hnsw/hnsw_backend.py (Ruff I001)

- Update benchmarks/run_evaluation.py

- Update apps/base_rag_example.py and leann-core API usage

- Add benchmarks/data/README.md

- Update uv.lock

- Misc cleanup

- Note: added paru-bin as an embedded git repo; consider making it a submodule (git rm --cached paru-bin) if unintended

* chore: remove unintended embedded repo paru-bin and ignore it

Fix CI: avoid missing .gitmodules entry by removing gitlink and adding to .gitignore.

* ci: retrigger after removing unintended gitlink (paru-bin)

* feat(benchmarks): add --batch-size option and plumb through to HNSW search (default 0)

* feat(hnsw): add batch_size to LeannSearcher.search and LeannChat.ask; forward only for HNSW backend

* chore(logging): surface recompute and batching params; enable INFO logging in benchmark

* feat(embeddings): add optional manual tokenization path (HF tokenizer+model) with mean pooling; default remains SentenceTransformer.encode

* fix micro bench and fix pre commit

* update readme

---------

Co-authored-by: yichuan-w <yichuan-w@users.noreply.github.com>

benchmarks/data/README.md

@@ -0,0 +1,44 @@
+---
+license: mit
+---
+
+# LEANN-RAG Evaluation Data
+
+This repository contains the necessary data to run the recall evaluation scripts for the [LEANN-RAG](https://huggingface.co/LEANN-RAG) project.
+
+## Dataset Components
+
+This dataset is structured into three main parts:
+
+1.  **Pre-built LEANN Indices**:
+    *   `dpr/`: A pre-built index for the DPR dataset.
+    *   `rpj_wiki/`: A pre-built index for the RPJ-Wiki dataset.
+    These indices were created using the `leann-core` library and are required by the `LeannSearcher`.
+
+2.  **Ground Truth Data**:
+    *   `ground_truth/`: Contains the ground truth files (`flat_results_nq_k3.json`) for both the DPR and RPJ-Wiki datasets. These files map queries to the original passage IDs from the Natural Questions benchmark, evaluated using the Contriever model.
+
+3.  **Queries**:
+    *   `queries/`: Contains the `nq_open.jsonl` file with the Natural Questions queries used for the evaluation.
+
+## Usage
+
+To use this data, you can download it locally using the `huggingface-hub` library. First, install the library:
+
+```bash
+pip install huggingface-hub
+```
+
+Then, you can download the entire dataset to a local directory (e.g., `data/`) with the following Python script:
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id="LEANN-RAG/leann-rag-evaluation-data",
+    repo_type="dataset",
+    local_dir="data"
+)
+```
+
+This will download all the necessary files into a local `data` folder, preserving the repository structure. The evaluation scripts in the main [LEANN-RAG Space](https://huggingface.co/LEANN-RAG) are configured to work with this data structure.

benchmarks/run_evaluation.py

@@ -12,7 +12,7 @@ import time
 from pathlib import Path
 
 import numpy as np
-from leann.api import LeannBuilder, LeannSearcher
+from leann.api import LeannBuilder, LeannChat, LeannSearcher
 
 
 def download_data_if_needed(data_root: Path, download_embeddings: bool = False):
@@ -197,6 +197,25 @@ def main():
     parser.add_argument(
         "--ef-search", type=int, default=120, help="The 'efSearch' parameter for HNSW."
     )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=0,
+        help="Batch size for HNSW batched search (0 disables batching)",
+    )
+    parser.add_argument(
+        "--llm-type",
+        type=str,
+        choices=["ollama", "hf", "openai", "gemini", "simulated"],
+        default="ollama",
+        help="LLM backend type to optionally query during evaluation (default: ollama)",
+    )
+    parser.add_argument(
+        "--llm-model",
+        type=str,
+        default="qwen3:1.7b",
+        help="LLM model identifier for the chosen backend (default: qwen3:1.7b)",
+    )
     args = parser.parse_args()
 
     # --- Path Configuration ---
@@ -318,9 +337,24 @@ def main():
 
         for i in range(num_eval_queries):
             start_time = time.time()
-            new_results = searcher.search(queries[i], top_k=args.top_k, ef=args.ef_search)
+            new_results = searcher.search(
+                queries[i],
+                top_k=args.top_k,
+                complexity=args.ef_search,
+                batch_size=args.batch_size,
+            )
             search_times.append(time.time() - start_time)
 
+            # Optional: also call the LLM with configurable backend/model (does not affect recall)
+            llm_config = {"type": args.llm_type, "model": args.llm_model}
+            chat = LeannChat(args.index_path, llm_config=llm_config, searcher=searcher)
+            answer = chat.ask(
+                queries[i],
+                top_k=args.top_k,
+                complexity=args.ef_search,
+                batch_size=args.batch_size,
+            )
+            print(f"Answer: {answer}")
             # Correct Recall Calculation: Based on TEXT content
             new_texts = {result.text for result in new_results}
 

benchmarks/simple_mac_tpt_test.py

@@ -20,7 +20,7 @@ except ImportError:
 
 @dataclass
 class BenchmarkConfig:
-    model_path: str = "facebook/contriever"
+    model_path: str = "facebook/contriever-msmarco"
     batch_sizes: list[int] = None
     seq_length: int = 256
     num_runs: int = 5
@@ -34,7 +34,7 @@ class BenchmarkConfig:
 
     def __post_init__(self):
         if self.batch_sizes is None:
-            self.batch_sizes = [1, 2, 4, 8, 16, 32, 64]
+            self.batch_sizes = [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
 
 
 class MLXBenchmark:
@@ -179,11 +179,14 @@ class Benchmark:
 
     def _run_inference(self, input_ids: torch.Tensor) -> float:
         attention_mask = torch.ones_like(input_ids)
-
+        # print shape of input_ids and attention_mask
+        print(f"input_ids shape: {input_ids.shape}")
+        print(f"attention_mask shape: {attention_mask.shape}")
         start_time = time.time()
         with torch.no_grad():
             self.model(input_ids=input_ids, attention_mask=attention_mask)
-        # mps sync
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
         if torch.backends.mps.is_available():
             torch.mps.synchronize()
         end_time = time.time()