fix readme

benchmarks/financebench/README.md

@@ -0,0 +1,115 @@
+# FinanceBench Benchmark for LEANN-RAG
+
+FinanceBench is a benchmark for evaluating retrieval-augmented generation (RAG) systems on financial document question-answering tasks.
+
+## Dataset
+
+- **Source**: [PatronusAI/financebench](https://huggingface.co/datasets/PatronusAI/financebench)
+- **Questions**: 150 financial Q&A examples
+- **Documents**: 368 PDF files (10-K, 10-Q, 8-K, earnings reports)
+- **Companies**: Major public companies (3M, Apple, Microsoft, Amazon, etc.)
+- **Paper**: [FinanceBench: A New Benchmark for Financial Question Answering](https://arxiv.org/abs/2311.11944)
+
+## Structure
+
+```
+benchmarks/financebench/
+├── setup_financebench.py        # Downloads PDFs and builds index
+├── evaluate_financebench.py     # Intelligent evaluation script
+├── data/
+│   ├── financebench_merged.jsonl     # Q&A dataset
+│   ├── pdfs/                         # Downloaded financial documents
+│   └── index/                        # LEANN indexes
+│       └── financebench_full_hnsw.leann
+└── README.md
+```
+
+## Usage
+
+### 1. Setup (Download & Build Index)
+
+```bash
+cd benchmarks/financebench
+python setup_financebench.py
+```
+
+This will:
+- Download the 150 Q&A examples
+- Download all 368 PDF documents (parallel processing)
+- Build a LEANN index from 53K+ text chunks
+- Verify setup with test query
+
+### 2. Evaluation
+
+```bash
+# Basic retrieval evaluation
+python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann
+
+
+# RAG generation evaluation with Qwen3-8B
+python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann --stage 4 --complexity 64 --llm-backend hf --model-name Qwen/Qwen3-8B --output results_qwen3.json
+```
+
+## Evaluation Methods
+
+### Retrieval Evaluation
+Uses intelligent matching with three strategies:
+1. **Exact text overlap** - Direct substring matches
+2. **Number matching** - Key financial figures ($1,577, 1.2B, etc.)
+3. **Semantic similarity** - Word overlap with 20% threshold
+
+### QA Evaluation
+LLM-based answer evaluation using GPT-4o:
+- Handles numerical rounding and equivalent representations
+- Considers fractions, percentages, and decimal equivalents
+- Evaluates semantic meaning rather than exact text match
+
+## Benchmark Results
+
+### LEANN-RAG Performance (sentence-transformers/all-mpnet-base-v2)
+
+**Retrieval Metrics:**
+- **Question Coverage**: 100.0% (all questions retrieve relevant docs)
+- **Exact Match Rate**: 0.7% (substring overlap with evidence)
+- **Number Match Rate**: 120.7% (key financial figures matched)*
+- **Semantic Match Rate**: 4.7% (word overlap ≥20%)
+- **Average Search Time**: 0.097s
+
+**QA Metrics:**
+- **Accuracy**: 42.7% (LLM-evaluated answer correctness)
+- **Average QA Time**: 4.71s (end-to-end response time)
+
+**System Performance:**
+- **Index Size**: 53,985 chunks from 368 PDFs
+- **Build Time**: ~5-10 minutes with sentence-transformers/all-mpnet-base-v2
+
+*Note: Number match rate >100% indicates multiple retrieved documents contain the same financial figures, which is expected behavior for financial data appearing across multiple document sections.
+
+### LEANN-RAG Generation Performance (Qwen3-8B)
+
+- **Stage 4 (Index Comparison):**
+  - Compact Index: 5.0 MB
+  - Non-compact Index: 172.2 MB
+  - **Storage Saving**: 97.1%
+- **Search Performance**:
+  - Non-compact (no recompute): 0.009s avg per query
+  - Compact (with recompute): 2.203s avg per query
+  - Speed ratio: 0.004x
+
+**Generation Evaluation (20 queries, complexity=64):**
+- **Average Search Time**: 1.638s per query
+- **Average Generation Time**: 45.957s per query
+- **LLM Backend**: HuggingFace transformers
+- **Model**: Qwen/Qwen3-8B (thinking model with <think></think> processing)
+- **Total Questions Processed**: 20
+
+## Options
+
+```bash
+# Use different backends
+python setup_financebench.py --backend diskann
+python evaluate_financebench.py --index data/index/financebench_full_diskann.leann
+
+# Use different embedding models
+python setup_financebench.py --embedding-model facebook/contriever
+```

benchmarks/financebench/evaluate_financebench.py

@@ -0,0 +1,923 @@
+"""
+FinanceBench Evaluation Script - Modular Recall-based Evaluation
+"""
+
+import argparse
+import json
+import logging
+import os
+import pickle
+import time
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import openai
+from leann import LeannChat, LeannSearcher
+from leann_backend_hnsw import faiss
+
+from ..llm_utils import evaluate_rag, generate_hf, generate_vllm, load_hf_model, load_vllm_model
+
+# Setup logging to reduce verbose output
+logging.basicConfig(level=logging.WARNING)
+logging.getLogger("leann.api").setLevel(logging.WARNING)
+logging.getLogger("leann_backend_hnsw").setLevel(logging.WARNING)
+
+
+class RecallEvaluator:
+    """Stage 2: Evaluate Recall@3 (searcher vs baseline)"""
+
+    def __init__(self, index_path: str, baseline_dir: str):
+        self.index_path = index_path
+        self.baseline_dir = baseline_dir
+        self.searcher = LeannSearcher(index_path)
+
+        # Load FAISS flat baseline
+        baseline_index_path = os.path.join(baseline_dir, "faiss_flat.index")
+        metadata_path = os.path.join(baseline_dir, "metadata.pkl")
+
+        self.faiss_index = faiss.read_index(baseline_index_path)
+        with open(metadata_path, "rb") as f:
+            self.passage_ids = pickle.load(f)
+        print(f"📚 Loaded FAISS flat baseline with {self.faiss_index.ntotal} vectors")
+
+    def evaluate_recall_at_3(
+        self, queries: list[str], complexity: int = 64, recompute_embeddings: bool = True
+    ) -> float:
+        """Evaluate recall@3 for given queries at specified complexity"""
+        recompute_str = "with recompute" if recompute_embeddings else "no recompute"
+        print(f"🔍 Evaluating recall@3 with complexity={complexity} ({recompute_str})...")
+
+        total_recall = 0.0
+        num_queries = len(queries)
+
+        for i, query in enumerate(queries):
+            # Get ground truth: search with FAISS flat
+            from leann.api import compute_embeddings
+
+            query_embedding = compute_embeddings(
+                [query],
+                self.searcher.embedding_model,
+                mode=self.searcher.embedding_mode,
+                use_server=False,
+            ).astype(np.float32)
+
+            # Search FAISS flat for ground truth using LEANN's modified faiss API
+            n = query_embedding.shape[0]  # Number of queries
+            k = 3  # Number of nearest neighbors
+            distances = np.zeros((n, k), dtype=np.float32)
+            labels = np.zeros((n, k), dtype=np.int64)
+
+            self.faiss_index.search(
+                n,
+                faiss.swig_ptr(query_embedding),
+                k,
+                faiss.swig_ptr(distances),
+                faiss.swig_ptr(labels),
+            )
+
+            # Extract the results
+            baseline_ids = {self.passage_ids[idx] for idx in labels[0]}
+
+            # Search with LEANN at specified complexity
+            test_results = self.searcher.search(
+                query,
+                top_k=3,
+                complexity=complexity,
+                recompute_embeddings=recompute_embeddings,
+            )
+            test_ids = {result.id for result in test_results}
+
+            # Calculate recall@3 = |intersection| / |ground_truth|
+            intersection = test_ids.intersection(baseline_ids)
+            recall = len(intersection) / 3.0  # Ground truth size is 3
+            total_recall += recall
+
+            if i < 3:  # Show first few examples
+                print(f"  Query {i + 1}: '{query[:50]}...' -> Recall@3: {recall:.3f}")
+                print(f"    FAISS ground truth: {list(baseline_ids)}")
+                print(f"    LEANN results (C={complexity}, {recompute_str}): {list(test_ids)}")
+                print(f"    Intersection: {list(intersection)}")
+
+        avg_recall = total_recall / num_queries
+        print(f"📊 Average Recall@3: {avg_recall:.3f} ({avg_recall * 100:.1f}%)")
+        return avg_recall
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if hasattr(self, "searcher"):
+            self.searcher.cleanup()
+
+
+class FinanceBenchEvaluator:
+    def __init__(self, index_path: str, openai_api_key: Optional[str] = None):
+        self.index_path = index_path
+        self.openai_client = openai.OpenAI(api_key=openai_api_key) if openai_api_key else None
+
+        self.searcher = LeannSearcher(index_path)
+        self.chat = LeannChat(index_path) if openai_api_key else None
+
+    def load_dataset(self, dataset_path: str = "data/financebench_merged.jsonl"):
+        """Load FinanceBench dataset"""
+        data = []
+        with open(dataset_path, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data.append(json.loads(line))
+
+        print(f"📊 Loaded {len(data)} FinanceBench examples")
+        return data
+
+    def analyze_index_sizes(self) -> dict:
+        """Analyze index sizes with and without embeddings"""
+
+        print("📏 Analyzing index sizes...")
+
+        # Get all index-related files
+        index_path = Path(self.index_path)
+        index_dir = index_path.parent
+        index_name = index_path.stem  # Remove .leann extension
+
+        sizes = {}
+        total_with_embeddings = 0
+
+        # Core index files
+        index_file = index_dir / f"{index_name}.index"
+        meta_file = index_dir / f"{index_path.name}.meta.json"  # Keep .leann for meta file
+        passages_file = index_dir / f"{index_path.name}.passages.jsonl"  # Keep .leann for passages
+        passages_idx_file = index_dir / f"{index_path.name}.passages.idx"  # Keep .leann for idx
+
+        for file_path, name in [
+            (index_file, "index"),
+            (meta_file, "metadata"),
+            (passages_file, "passages_text"),
+            (passages_idx_file, "passages_index"),
+        ]:
+            if file_path.exists():
+                size_mb = file_path.stat().st_size / (1024 * 1024)
+                sizes[name] = size_mb
+                total_with_embeddings += size_mb
+
+            else:
+                sizes[name] = 0
+
+        sizes["total_with_embeddings"] = total_with_embeddings
+        sizes["index_only_mb"] = sizes["index"]  # Just the .index file for fair comparison
+
+        print(f"  📁 Total index size: {total_with_embeddings:.1f} MB")
+        print(f"  📁 Index file only: {sizes['index']:.1f} MB")
+
+        return sizes
+
+    def create_compact_index_for_comparison(self, compact_index_path: str) -> dict:
+        """Create a compact index for comparison purposes"""
+        print("🏗️ Building compact index from existing passages...")
+
+        # Load existing passages from current index
+
+        from leann import LeannBuilder
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        print(f"📄 Loading passages from {passage_file}...")
+
+        # Build compact index with same passages
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            embedding_model=meta["embedding_model"],
+            embedding_mode=meta.get("embedding_mode", "sentence-transformers"),
+            is_recompute=True,  # Enable recompute (no stored embeddings)
+            is_compact=True,  # Enable compact storage
+            **meta.get("backend_kwargs", {}),
+        )
+
+        # Load all passages
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    builder.add_text(data["text"], metadata=data.get("metadata", {}))
+
+        print(f"🔨 Building compact index at {compact_index_path}...")
+        builder.build_index(compact_index_path)
+
+        # Analyze the compact index size
+        temp_evaluator = FinanceBenchEvaluator(compact_index_path)
+        compact_sizes = temp_evaluator.analyze_index_sizes()
+        compact_sizes["index_type"] = "compact"
+
+        return compact_sizes
+
+    def create_non_compact_index_for_comparison(self, non_compact_index_path: str) -> dict:
+        """Create a non-compact index for comparison purposes"""
+        print("🏗️ Building non-compact index from existing passages...")
+
+        # Load existing passages from current index
+
+        from leann import LeannBuilder
+
+        current_index_path = Path(self.index_path)
+        current_index_dir = current_index_path.parent
+        current_index_name = current_index_path.name
+
+        # Read metadata to get passage source
+        meta_path = current_index_dir / f"{current_index_name}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.load(f)
+
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not Path(passage_file).is_absolute():
+            passage_file = current_index_dir / Path(passage_file).name
+
+        print(f"📄 Loading passages from {passage_file}...")
+
+        # Build non-compact index with same passages
+        builder = LeannBuilder(
+            backend_name="hnsw",
+            embedding_model=meta["embedding_model"],
+            embedding_mode=meta.get("embedding_mode", "sentence-transformers"),
+            is_recompute=False,  # Disable recompute (store embeddings)
+            is_compact=False,  # Disable compact storage
+            **{
+                k: v
+                for k, v in meta.get("backend_kwargs", {}).items()
+                if k not in ["is_recompute", "is_compact"]
+            },
+        )
+
+        # Load all passages
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    builder.add_text(data["text"], metadata=data.get("metadata", {}))
+
+        print(f"🔨 Building non-compact index at {non_compact_index_path}...")
+        builder.build_index(non_compact_index_path)
+
+        # Analyze the non-compact index size
+        temp_evaluator = FinanceBenchEvaluator(non_compact_index_path)
+        non_compact_sizes = temp_evaluator.analyze_index_sizes()
+        non_compact_sizes["index_type"] = "non_compact"
+
+        return non_compact_sizes
+
+    def compare_index_performance(
+        self, non_compact_path: str, compact_path: str, test_data: list, complexity: int
+    ) -> dict:
+        """Compare performance between non-compact and compact indexes"""
+        print("⚡ Comparing search performance between indexes...")
+
+        import time
+
+        from leann import LeannSearcher
+
+        # Test queries
+        test_queries = [item["question"] for item in test_data[:5]]
+
+        results = {
+            "non_compact": {"search_times": []},
+            "compact": {"search_times": []},
+            "avg_search_times": {},
+            "speed_ratio": 0.0,
+        }
+
+        # Test non-compact index (no recompute)
+        print("  🔍 Testing non-compact index (no recompute)...")
+        non_compact_searcher = LeannSearcher(non_compact_path)
+
+        for query in test_queries:
+            start_time = time.time()
+            _ = non_compact_searcher.search(
+                query, top_k=3, complexity=complexity, recompute_embeddings=False
+            )
+            search_time = time.time() - start_time
+            results["non_compact"]["search_times"].append(search_time)
+
+        # Test compact index (with recompute)
+        print("  🔍 Testing compact index (with recompute)...")
+        compact_searcher = LeannSearcher(compact_path)
+
+        for query in test_queries:
+            start_time = time.time()
+            _ = compact_searcher.search(
+                query, top_k=3, complexity=complexity, recompute_embeddings=True
+            )
+            search_time = time.time() - start_time
+            results["compact"]["search_times"].append(search_time)
+
+        # Calculate averages
+        results["avg_search_times"]["non_compact"] = sum(
+            results["non_compact"]["search_times"]
+        ) / len(results["non_compact"]["search_times"])
+        results["avg_search_times"]["compact"] = sum(results["compact"]["search_times"]) / len(
+            results["compact"]["search_times"]
+        )
+
+        # Performance ratio
+        if results["avg_search_times"]["compact"] > 0:
+            results["speed_ratio"] = (
+                results["avg_search_times"]["non_compact"] / results["avg_search_times"]["compact"]
+            )
+        else:
+            results["speed_ratio"] = float("inf")
+
+        print(
+            f"    Non-compact (no recompute): {results['avg_search_times']['non_compact']:.3f}s avg"
+        )
+        print(f"    Compact (with recompute): {results['avg_search_times']['compact']:.3f}s avg")
+        print(f"    Speed ratio: {results['speed_ratio']:.2f}x")
+
+        # Cleanup
+        non_compact_searcher.cleanup()
+        compact_searcher.cleanup()
+
+        return results
+
+    def evaluate_timing_breakdown(
+        self, data: list[dict], max_samples: Optional[int] = None
+    ) -> dict:
+        """Evaluate timing breakdown and accuracy by hacking LeannChat.ask() for separated timing"""
+        if not self.chat or not self.openai_client:
+            print("⚠️  Skipping timing evaluation (no OpenAI API key provided)")
+            return {
+                "total_questions": 0,
+                "avg_search_time": 0.0,
+                "avg_generation_time": 0.0,
+                "avg_total_time": 0.0,
+                "accuracy": 0.0,
+            }
+
+        print("🔍🤖 Evaluating timing breakdown and accuracy (search + generation)...")
+
+        if max_samples:
+            data = data[:max_samples]
+            print(f"📝 Using first {max_samples} samples for timing evaluation")
+
+        search_times = []
+        generation_times = []
+        total_times = []
+        correct_answers = 0
+
+        for i, item in enumerate(data):
+            question = item["question"]
+            ground_truth = item["answer"]
+
+            try:
+                # Hack: Monkey-patch the ask method to capture internal timing
+                original_ask = self.chat.ask
+                captured_search_time = None
+                captured_generation_time = None
+
+                def patched_ask(*args, **kwargs):
+                    nonlocal captured_search_time, captured_generation_time
+
+                    # Time the search part
+                    search_start = time.time()
+                    results = self.chat.searcher.search(args[0], top_k=3, complexity=64)
+                    captured_search_time = time.time() - search_start
+
+                    # Time the generation part
+                    context = "\n\n".join([r.text for r in results])
+                    prompt = (
+                        "Here is some retrieved context that might help answer your question:\n\n"
+                        f"{context}\n\n"
+                        f"Question: {args[0]}\n\n"
+                        "Please provide the best answer you can based on this context and your knowledge."
+                    )
+
+                    generation_start = time.time()
+                    answer = self.chat.llm.ask(prompt)
+                    captured_generation_time = time.time() - generation_start
+
+                    return answer
+
+                # Apply the patch
+                self.chat.ask = patched_ask
+
+                # Time the total QA
+                total_start = time.time()
+                generated_answer = self.chat.ask(question)
+                total_time = time.time() - total_start
+
+                # Restore original method
+                self.chat.ask = original_ask
+
+                # Store the timings
+                search_times.append(captured_search_time)
+                generation_times.append(captured_generation_time)
+                total_times.append(total_time)
+
+                # Check accuracy using LLM as judge
+                is_correct = self._check_answer_accuracy(generated_answer, ground_truth, question)
+                if is_correct:
+                    correct_answers += 1
+
+                status = "✅" if is_correct else "❌"
+                print(
+                    f"Question {i + 1}/{len(data)}: {status} Search={captured_search_time:.3f}s, Gen={captured_generation_time:.3f}s, Total={total_time:.3f}s"
+                )
+                print(f"  GT: {ground_truth}")
+                print(f"  Gen: {generated_answer[:100]}...")
+
+            except Exception as e:
+                print(f"  ❌ Error: {e}")
+                search_times.append(0.0)
+                generation_times.append(0.0)
+                total_times.append(0.0)
+
+        accuracy = correct_answers / len(data) if data else 0.0
+
+        metrics = {
+            "total_questions": len(data),
+            "avg_search_time": sum(search_times) / len(search_times) if search_times else 0.0,
+            "avg_generation_time": sum(generation_times) / len(generation_times)
+            if generation_times
+            else 0.0,
+            "avg_total_time": sum(total_times) / len(total_times) if total_times else 0.0,
+            "accuracy": accuracy,
+            "correct_answers": correct_answers,
+            "search_times": search_times,
+            "generation_times": generation_times,
+            "total_times": total_times,
+        }
+
+        return metrics
+
+    def _check_answer_accuracy(
+        self, generated_answer: str, ground_truth: str, question: str
+    ) -> bool:
+        """Check if generated answer matches ground truth using LLM as judge"""
+        judge_prompt = f"""You are an expert judge evaluating financial question answering.
+
+Question: {question}
+
+Ground Truth Answer: {ground_truth}
+
+Generated Answer: {generated_answer}
+
+Task: Determine if the generated answer is factually correct compared to the ground truth. Focus on:
+1. Numerical accuracy (exact values, units, currency)
+2. Key financial concepts and terminology
+3. Overall factual correctness
+
+For financial data, small formatting differences are OK (e.g., "$1,577" vs "1577 million" vs "$1.577 billion"), but the core numerical value must match.
+
+Respond with exactly one word: "CORRECT" if the generated answer is factually accurate, or "INCORRECT" if it's wrong or significantly different."""
+
+        try:
+            judge_response = self.openai_client.chat.completions.create(
+                model="gpt-4o-mini",
+                messages=[{"role": "user", "content": judge_prompt}],
+                max_tokens=10,
+                temperature=0,
+            )
+            judgment = judge_response.choices[0].message.content.strip().upper()
+            return judgment == "CORRECT"
+        except Exception as e:
+            print(f"  ⚠️  Judge error: {e}, falling back to string matching")
+            # Fallback to simple string matching
+            gen_clean = generated_answer.strip().lower().replace("$", "").replace(",", "")
+            gt_clean = ground_truth.strip().lower().replace("$", "").replace(",", "")
+            return gt_clean in gen_clean
+
+    def _print_results(self, timing_metrics: dict):
+        """Print evaluation results"""
+        print("\n🎯 EVALUATION RESULTS")
+        print("=" * 50)
+
+        # Index comparison analysis
+        if "current_index" in timing_metrics and "non_compact_index" in timing_metrics:
+            print("\n📏 Index Comparison Analysis:")
+            current = timing_metrics["current_index"]
+            non_compact = timing_metrics["non_compact_index"]
+
+            print(f"  Compact index (current): {current.get('total_with_embeddings', 0):.1f} MB")
+            print(
+                f"  Non-compact index (with embeddings): {non_compact.get('total_with_embeddings', 0):.1f} MB"
+            )
+            print(
+                f"  Storage saving by compact: {timing_metrics.get('storage_saving_percent', 0):.1f}%"
+            )
+
+            print("  Component breakdown (non-compact):")
+            print(f"    - Main index: {non_compact.get('index', 0):.1f} MB")
+            print(f"    - Passages text: {non_compact.get('passages_text', 0):.1f} MB")
+            print(f"    - Passages index: {non_compact.get('passages_index', 0):.1f} MB")
+            print(f"    - Metadata: {non_compact.get('metadata', 0):.1f} MB")
+
+        # Performance comparison
+        if "performance_comparison" in timing_metrics:
+            perf = timing_metrics["performance_comparison"]
+            print("\n⚡ Performance Comparison:")
+            print(
+                f"  Non-compact (no recompute): {perf.get('avg_search_times', {}).get('non_compact', 0):.3f}s avg"
+            )
+            print(
+                f"  Compact (with recompute): {perf.get('avg_search_times', {}).get('compact', 0):.3f}s avg"
+            )
+            print(f"  Speed ratio: {perf.get('speed_ratio', 0):.2f}x")
+
+        # Legacy single index analysis (fallback)
+        if "total_with_embeddings" in timing_metrics and "current_index" not in timing_metrics:
+            print("\n📏 Index Size Analysis:")
+            print(f"  Total index size: {timing_metrics.get('total_with_embeddings', 0):.1f} MB")
+
+        print("\n📊 Accuracy:")
+        print(f"  Accuracy: {timing_metrics.get('accuracy', 0) * 100:.1f}%")
+        print(
+            f"  Correct Answers: {timing_metrics.get('correct_answers', 0)}/{timing_metrics.get('total_questions', 0)}"
+        )
+
+        print("\n📊 Timing Breakdown:")
+        print(f"  Total Questions: {timing_metrics.get('total_questions', 0)}")
+        print(f"  Avg Search Time: {timing_metrics.get('avg_search_time', 0):.3f}s")
+        print(f"  Avg Generation Time: {timing_metrics.get('avg_generation_time', 0):.3f}s")
+        print(f"  Avg Total Time: {timing_metrics.get('avg_total_time', 0):.3f}s")
+
+        if timing_metrics.get("avg_total_time", 0) > 0:
+            search_pct = (
+                timing_metrics.get("avg_search_time", 0)
+                / timing_metrics.get("avg_total_time", 1)
+                * 100
+            )
+            gen_pct = (
+                timing_metrics.get("avg_generation_time", 0)
+                / timing_metrics.get("avg_total_time", 1)
+                * 100
+            )
+            print("\n📈 Time Distribution:")
+            print(f"  Search: {search_pct:.1f}%")
+            print(f"  Generation: {gen_pct:.1f}%")
+
+    def cleanup(self):
+        """Cleanup resources"""
+        if self.searcher:
+            self.searcher.cleanup()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Modular FinanceBench Evaluation")
+    parser.add_argument("--index", required=True, help="Path to LEANN index")
+    parser.add_argument("--dataset", default="data/financebench_merged.jsonl", help="Dataset path")
+    parser.add_argument(
+        "--stage",
+        choices=["2", "3", "4", "all"],
+        default="all",
+        help="Which stage to run (2=recall, 3=complexity, 4=generation)",
+    )
+    parser.add_argument("--complexity", type=int, default=None, help="Complexity for search")
+    parser.add_argument("--baseline-dir", default="baseline", help="Baseline output directory")
+    parser.add_argument("--openai-api-key", help="OpenAI API key for generation evaluation")
+    parser.add_argument("--output", help="Save results to JSON file")
+    parser.add_argument(
+        "--llm-backend", choices=["openai", "hf", "vllm"], default="openai", help="LLM backend"
+    )
+    parser.add_argument("--model-name", default="Qwen3-8B", help="Model name for HF/vLLM")
+
+    args = parser.parse_args()
+
+    try:
+        # Check if baseline exists
+        baseline_index_path = os.path.join(args.baseline_dir, "faiss_flat.index")
+        if not os.path.exists(baseline_index_path):
+            print(f"❌ FAISS baseline not found at {baseline_index_path}")
+            print("💡 Please run setup_financebench.py first to build the baseline")
+            exit(1)
+
+        if args.stage == "2" or args.stage == "all":
+            # Stage 2: Recall@3 evaluation
+            print("🚀 Starting Stage 2: Recall@3 evaluation")
+
+            evaluator = RecallEvaluator(args.index, args.baseline_dir)
+
+            # Load FinanceBench queries for testing
+            print("📖 Loading FinanceBench dataset...")
+            queries = []
+            with open(args.dataset, encoding="utf-8") as f:
+                for line in f:
+                    if line.strip():
+                        data = json.loads(line)
+                        queries.append(data["question"])
+
+            # Test with more queries for robust measurement
+            test_queries = queries[:2000]
+            print(f"🧪 Testing with {len(test_queries)} queries")
+
+            # Test with complexity 64
+            complexity = 64
+            recall = evaluator.evaluate_recall_at_3(test_queries, complexity)
+            print(f"📈 Recall@3 at complexity {complexity}: {recall * 100:.1f}%")
+
+            evaluator.cleanup()
+            print("✅ Stage 2 completed!\n")
+
+        # Shared non-compact index path for Stage 3 and 4
+        non_compact_index_path = args.index.replace(".leann", "_noncompact.leann")
+        complexity = args.complexity
+
+        if args.stage == "3" or args.stage == "all":
+            # Stage 3: Binary search for 90% recall complexity (using non-compact index for speed)
+            print("🚀 Starting Stage 3: Binary search for 90% recall complexity")
+            print(
+                "💡 Creating non-compact index for fast binary search with recompute_embeddings=False"
+            )
+
+            # Create non-compact index for binary search (will be reused in Stage 4)
+            print("🏗️ Creating non-compact index for binary search...")
+            evaluator = FinanceBenchEvaluator(args.index)
+            evaluator.create_non_compact_index_for_comparison(non_compact_index_path)
+
+            # Use non-compact index for binary search
+            binary_search_evaluator = RecallEvaluator(non_compact_index_path, args.baseline_dir)
+
+            # Load queries for testing
+            print("📖 Loading FinanceBench dataset...")
+            queries = []
+            with open(args.dataset, encoding="utf-8") as f:
+                for line in f:
+                    if line.strip():
+                        data = json.loads(line)
+                        queries.append(data["question"])
+
+            # Use more queries for robust measurement
+            test_queries = queries[:200]
+            print(f"🧪 Testing with {len(test_queries)} queries")
+
+            # Binary search for 90% recall complexity (without recompute for speed)
+            target_recall = 0.9
+            min_complexity, max_complexity = 1, 32
+
+            print(f"🔍 Binary search for {target_recall * 100}% recall complexity...")
+            print(f"Search range: {min_complexity} to {max_complexity}")
+
+            best_complexity = None
+            best_recall = 0.0
+
+            while min_complexity <= max_complexity:
+                mid_complexity = (min_complexity + max_complexity) // 2
+
+                print(
+                    f"\n🧪 Testing complexity {mid_complexity} (no recompute, non-compact index)..."
+                )
+                # Use recompute_embeddings=False on non-compact index for fast binary search
+                recall = binary_search_evaluator.evaluate_recall_at_3(
+                    test_queries, mid_complexity, recompute_embeddings=False
+                )
+
+                print(
+                    f"  Complexity {mid_complexity}: Recall@3 = {recall:.3f} ({recall * 100:.1f}%)"
+                )
+
+                if recall >= target_recall:
+                    best_complexity = mid_complexity
+                    best_recall = recall
+                    max_complexity = mid_complexity - 1
+                    print("  ✅ Target reached! Searching for lower complexity...")
+                else:
+                    min_complexity = mid_complexity + 1
+                    print("  ❌ Below target. Searching for higher complexity...")
+
+            if best_complexity is not None:
+                print("\n🎯 Optimal complexity found!")
+                print(f"  Complexity: {best_complexity}")
+                print(f"  Recall@3: {best_recall:.3f} ({best_recall * 100:.1f}%)")
+
+                # Test a few complexities around the optimal one for verification
+                print("\n🔬 Verification test around optimal complexity:")
+                verification_complexities = [
+                    max(1, best_complexity - 2),
+                    max(1, best_complexity - 1),
+                    best_complexity,
+                    best_complexity + 1,
+                    best_complexity + 2,
+                ]
+
+                for complexity in verification_complexities:
+                    if complexity <= 512:  # reasonable upper bound
+                        recall = binary_search_evaluator.evaluate_recall_at_3(
+                            test_queries, complexity, recompute_embeddings=False
+                        )
+                        status = "✅" if recall >= target_recall else "❌"
+                        print(f"  {status} Complexity {complexity:3d}: {recall * 100:5.1f}%")
+
+                # Now test the optimal complexity with compact index and recompute for comparison
+                print(
+                    f"\n🔄 Testing optimal complexity {best_complexity} on compact index WITH recompute..."
+                )
+                compact_evaluator = RecallEvaluator(args.index, args.baseline_dir)
+                recall_with_recompute = compact_evaluator.evaluate_recall_at_3(
+                    test_queries[:10], best_complexity, recompute_embeddings=True
+                )
+                print(
+                    f"  ✅ Complexity {best_complexity} (compact index with recompute): {recall_with_recompute * 100:.1f}%"
+                )
+                complexity = best_complexity
+                print(
+                    f"  📊 Recall difference: {abs(best_recall - recall_with_recompute) * 100:.2f}%"
+                )
+                compact_evaluator.cleanup()
+            else:
+                print(f"\n❌ Could not find complexity achieving {target_recall * 100}% recall")
+                print("All tested complexities were below target.")
+
+            # Cleanup evaluators (keep non-compact index for Stage 4)
+            binary_search_evaluator.cleanup()
+            evaluator.cleanup()
+
+            print("✅ Stage 3 completed! Non-compact index saved for Stage 4.\n")
+
+        if args.stage == "4" or args.stage == "all":
+            # Stage 4: Comprehensive evaluation with dual index comparison
+            print("🚀 Starting Stage 4: Comprehensive evaluation with dual index comparison")
+
+            # Use FinanceBench evaluator for QA evaluation
+            evaluator = FinanceBenchEvaluator(
+                args.index, args.openai_api_key if args.llm_backend == "openai" else None
+            )
+
+            print("📖 Loading FinanceBench dataset...")
+            data = evaluator.load_dataset(args.dataset)
+
+            # Step 1: Analyze current (compact) index
+            print("\n📏 Analyzing current index (compact, pruned)...")
+            compact_size_metrics = evaluator.analyze_index_sizes()
+            compact_size_metrics["index_type"] = "compact"
+
+            # Step 2: Use existing non-compact index or create if needed
+            from pathlib import Path
+
+            if Path(non_compact_index_path).exists():
+                print(
+                    f"\n📁 Using existing non-compact index from Stage 3: {non_compact_index_path}"
+                )
+                temp_evaluator = FinanceBenchEvaluator(non_compact_index_path)
+                non_compact_size_metrics = temp_evaluator.analyze_index_sizes()
+                non_compact_size_metrics["index_type"] = "non_compact"
+            else:
+                print("\n🏗️ Creating non-compact index (with embeddings) for comparison...")
+                non_compact_size_metrics = evaluator.create_non_compact_index_for_comparison(
+                    non_compact_index_path
+                )
+
+            # Step 3: Compare index sizes
+            print("\n📊 Index size comparison:")
+            print(
+                f"  Compact index (current): {compact_size_metrics['total_with_embeddings']:.1f} MB"
+            )
+            print(
+                f"  Non-compact index: {non_compact_size_metrics['total_with_embeddings']:.1f} MB"
+            )
+            print("\n📊 Index-only size comparison (.index file only):")
+            print(f"  Compact index: {compact_size_metrics['index_only_mb']:.1f} MB")
+            print(f"  Non-compact index: {non_compact_size_metrics['index_only_mb']:.1f} MB")
+            # Use index-only size for fair comparison (same as Enron emails)
+            storage_saving = (
+                (non_compact_size_metrics["index_only_mb"] - compact_size_metrics["index_only_mb"])
+                / non_compact_size_metrics["index_only_mb"]
+                * 100
+            )
+            print(f"  Storage saving by compact: {storage_saving:.1f}%")
+
+            # Step 4: Performance comparison between the two indexes
+            if complexity is None:
+                raise ValueError("Complexity is required for performance comparison")
+
+            print("\n⚡ Performance comparison between indexes...")
+            performance_metrics = evaluator.compare_index_performance(
+                non_compact_index_path, args.index, data[:10], complexity=complexity
+            )
+
+            # Step 5: Generation evaluation
+            test_samples = 20
+            print(f"\n🧪 Testing with first {test_samples} samples for generation analysis")
+
+            if args.llm_backend == "openai" and args.openai_api_key:
+                print("🔍🤖 Running OpenAI-based generation evaluation...")
+                evaluation_start = time.time()
+                timing_metrics = evaluator.evaluate_timing_breakdown(data[:test_samples])
+                evaluation_time = time.time() - evaluation_start
+            else:
+                print(
+                    f"🔍🤖 Running {args.llm_backend} generation evaluation with {args.model_name}..."
+                )
+                try:
+                    # Load LLM
+                    if args.llm_backend == "hf":
+                        tokenizer, model = load_hf_model(args.model_name)
+
+                        def llm_func(prompt):
+                            return generate_hf(tokenizer, model, prompt)
+                    else:  # vllm
+                        llm, sampling_params = load_vllm_model(args.model_name)
+
+                        def llm_func(prompt):
+                            return generate_vllm(llm, sampling_params, prompt)
+
+                    # Simple generation evaluation
+                    queries = [item["question"] for item in data[:test_samples]]
+                    gen_results = evaluate_rag(
+                        evaluator.searcher,
+                        llm_func,
+                        queries,
+                        domain="finance",
+                        complexity=complexity,
+                    )
+
+                    timing_metrics = {
+                        "total_questions": len(queries),
+                        "avg_search_time": gen_results["avg_search_time"],
+                        "avg_generation_time": gen_results["avg_generation_time"],
+                        "results": gen_results["results"],
+                    }
+                    evaluation_time = time.time()
+
+                except Exception as e:
+                    print(f"❌ Generation evaluation failed: {e}")
+                    timing_metrics = {
+                        "total_questions": 0,
+                        "avg_search_time": 0,
+                        "avg_generation_time": 0,
+                    }
+                    evaluation_time = 0
+
+            # Combine all metrics
+            combined_metrics = {
+                **timing_metrics,
+                "total_evaluation_time": evaluation_time,
+                "current_index": compact_size_metrics,
+                "non_compact_index": non_compact_size_metrics,
+                "performance_comparison": performance_metrics,
+                "storage_saving_percent": storage_saving,
+            }
+
+            # Print results
+            print("\n📊 Generation Results:")
+            print(f"  Total Questions: {timing_metrics.get('total_questions', 0)}")
+            print(f"  Avg Search Time: {timing_metrics.get('avg_search_time', 0):.3f}s")
+            print(f"  Avg Generation Time: {timing_metrics.get('avg_generation_time', 0):.3f}s")
+
+            # Save results if requested
+            if args.output:
+                print(f"\n💾 Saving results to {args.output}...")
+                with open(args.output, "w") as f:
+                    json.dump(combined_metrics, f, indent=2, default=str)
+                print(f"✅ Results saved to {args.output}")
+
+            evaluator.cleanup()
+            print("✅ Stage 4 completed!\n")
+
+        if args.stage == "all":
+            print("🎉 All evaluation stages completed successfully!")
+            print("\n📋 Summary:")
+            print("  Stage 2: ✅ Recall@3 evaluation completed")
+            print("  Stage 3: ✅ Optimal complexity found")
+            print("  Stage 4: ✅ Generation accuracy & timing evaluation completed")
+            print("\n🔧 Recommended next steps:")
+            print("  - Use optimal complexity for best speed/accuracy balance")
+            print("  - Review accuracy and timing breakdown for performance optimization")
+            print("  - Run full evaluation on complete dataset if needed")
+
+            # Clean up non-compact index after all stages complete
+            print("\n🧹 Cleaning up temporary non-compact index...")
+            from pathlib import Path
+
+            if Path(non_compact_index_path).exists():
+                temp_index_dir = Path(non_compact_index_path).parent
+                temp_index_name = Path(non_compact_index_path).name
+                for temp_file in temp_index_dir.glob(f"{temp_index_name}*"):
+                    temp_file.unlink()
+                print(f"✅ Cleaned up {non_compact_index_path}")
+            else:
+                print("📝 No temporary index to clean up")
+    except KeyboardInterrupt:
+        print("\n⚠️  Evaluation interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Stage {args.stage} failed: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/financebench/setup_financebench.py

@@ -0,0 +1,462 @@
+#!/usr/bin/env python3
+"""
+FinanceBench Complete Setup Script
+Downloads all PDFs and builds full LEANN datastore
+"""
+
+import argparse
+import os
+import re
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from threading import Lock
+
+import pymupdf
+import requests
+from leann import LeannBuilder, LeannSearcher
+from tqdm import tqdm
+
+
+class FinanceBenchSetup:
+    def __init__(self, data_dir: str = "data"):
+        self.base_dir = Path(__file__).parent  # benchmarks/financebench/
+        self.data_dir = self.base_dir / data_dir
+        self.pdf_dir = self.data_dir / "pdfs"
+        self.dataset_file = self.data_dir / "financebench_merged.jsonl"
+        self.index_dir = self.data_dir / "index"
+        self.download_lock = Lock()
+
+    def download_dataset(self):
+        """Download the main FinanceBench dataset"""
+        print("📊 Downloading FinanceBench dataset...")
+        self.data_dir.mkdir(parents=True, exist_ok=True)
+
+        if self.dataset_file.exists():
+            print(f"✅ Dataset already exists: {self.dataset_file}")
+            return
+
+        url = "https://huggingface.co/datasets/PatronusAI/financebench/raw/main/financebench_merged.jsonl"
+        response = requests.get(url, stream=True)
+        response.raise_for_status()
+
+        with open(self.dataset_file, "wb") as f:
+            for chunk in response.iter_content(chunk_size=8192):
+                f.write(chunk)
+
+        print(f"✅ Dataset downloaded: {self.dataset_file}")
+
+    def get_pdf_list(self):
+        """Get list of all PDF files from GitHub"""
+        print("📋 Fetching PDF list from GitHub...")
+
+        response = requests.get(
+            "https://api.github.com/repos/patronus-ai/financebench/contents/pdfs"
+        )
+        response.raise_for_status()
+        pdf_files = response.json()
+
+        print(f"Found {len(pdf_files)} PDF files")
+        return pdf_files
+
+    def download_single_pdf(self, pdf_info, position):
+        """Download a single PDF file"""
+        pdf_name = pdf_info["name"]
+        pdf_path = self.pdf_dir / pdf_name
+
+        # Skip if already downloaded
+        if pdf_path.exists() and pdf_path.stat().st_size > 0:
+            return f"✅ {pdf_name} (cached)"
+
+        try:
+            # Download PDF
+            response = requests.get(pdf_info["download_url"], timeout=60)
+            response.raise_for_status()
+
+            # Write to file
+            with self.download_lock:
+                with open(pdf_path, "wb") as f:
+                    f.write(response.content)
+
+            return f"✅ {pdf_name} ({len(response.content) // 1024}KB)"
+
+        except Exception as e:
+            return f"❌ {pdf_name}: {e!s}"
+
+    def download_all_pdfs(self, max_workers: int = 5):
+        """Download all PDF files with parallel processing"""
+        self.pdf_dir.mkdir(parents=True, exist_ok=True)
+
+        pdf_files = self.get_pdf_list()
+
+        print(f"📥 Downloading {len(pdf_files)} PDFs with {max_workers} workers...")
+
+        with ThreadPoolExecutor(max_workers=max_workers) as executor:
+            # Submit all download tasks
+            future_to_pdf = {
+                executor.submit(self.download_single_pdf, pdf_info, i): pdf_info["name"]
+                for i, pdf_info in enumerate(pdf_files)
+            }
+
+            # Process completed downloads with progress bar
+            with tqdm(total=len(pdf_files), desc="Downloading PDFs") as pbar:
+                for future in as_completed(future_to_pdf):
+                    result = future.result()
+                    pbar.set_postfix_str(result.split()[-1] if "✅" in result else "Error")
+                    pbar.update(1)
+
+        # Verify downloads
+        downloaded_pdfs = list(self.pdf_dir.glob("*.pdf"))
+        print(f"✅ Successfully downloaded {len(downloaded_pdfs)}/{len(pdf_files)} PDFs")
+
+        # Show any failures
+        missing_pdfs = []
+        for pdf_info in pdf_files:
+            pdf_path = self.pdf_dir / pdf_info["name"]
+            if not pdf_path.exists() or pdf_path.stat().st_size == 0:
+                missing_pdfs.append(pdf_info["name"])
+
+        if missing_pdfs:
+            print(f"⚠️  Failed to download {len(missing_pdfs)} PDFs:")
+            for pdf in missing_pdfs[:5]:  # Show first 5
+                print(f"   - {pdf}")
+            if len(missing_pdfs) > 5:
+                print(f"   ... and {len(missing_pdfs) - 5} more")
+
+    def build_leann_index(
+        self,
+        backend: str = "hnsw",
+        embedding_model: str = "sentence-transformers/all-mpnet-base-v2",
+    ):
+        """Build LEANN index from all PDFs"""
+        print(f"🏗️  Building LEANN index with {backend} backend...")
+
+        # Check if we have PDFs
+        pdf_files = list(self.pdf_dir.glob("*.pdf"))
+        if not pdf_files:
+            raise RuntimeError("No PDF files found! Run download first.")
+
+        print(f"Found {len(pdf_files)} PDF files to process")
+
+        start_time = time.time()
+
+        # Initialize builder with standard compact configuration
+        builder = LeannBuilder(
+            backend_name=backend,
+            embedding_model=embedding_model,
+            embedding_mode="sentence-transformers",
+            graph_degree=32,
+            complexity=64,
+            is_recompute=True,  # Enable recompute (no stored embeddings)
+            is_compact=True,  # Enable compact storage (pruned)
+            num_threads=4,
+        )
+
+        # Process PDFs and extract text
+        total_chunks = 0
+        failed_pdfs = []
+
+        for pdf_path in tqdm(pdf_files, desc="Processing PDFs"):
+            try:
+                chunks = self.extract_pdf_text(pdf_path)
+                for chunk in chunks:
+                    builder.add_text(chunk["text"], metadata=chunk["metadata"])
+                    total_chunks += 1
+
+            except Exception as e:
+                print(f"❌ Failed to process {pdf_path.name}: {e}")
+                failed_pdfs.append(pdf_path.name)
+                continue
+
+        # Build index in index directory
+        self.index_dir.mkdir(parents=True, exist_ok=True)
+        index_path = self.index_dir / f"financebench_full_{backend}.leann"
+        print(f"🔨 Building index: {index_path}")
+        builder.build_index(str(index_path))
+
+        build_time = time.time() - start_time
+
+        print("✅ Index built successfully!")
+        print(f"   📁 Index path: {index_path}")
+        print(f"   📊 Total chunks: {total_chunks:,}")
+        print(f"   📄 Processed PDFs: {len(pdf_files) - len(failed_pdfs)}/{len(pdf_files)}")
+        print(f"   ⏱️  Build time: {build_time:.1f}s")
+
+        if failed_pdfs:
+            print(f"   ⚠️  Failed PDFs: {failed_pdfs}")
+
+        return str(index_path)
+
+    def build_faiss_flat_baseline(self, index_path: str, output_dir: str = "baseline"):
+        """Build FAISS flat baseline using the same embeddings as LEANN index"""
+        print("🔨 Building FAISS Flat baseline...")
+
+        import os
+        import pickle
+
+        import numpy as np
+        from leann.api import compute_embeddings
+        from leann_backend_hnsw import faiss
+
+        os.makedirs(output_dir, exist_ok=True)
+        baseline_path = os.path.join(output_dir, "faiss_flat.index")
+        metadata_path = os.path.join(output_dir, "metadata.pkl")
+
+        if os.path.exists(baseline_path) and os.path.exists(metadata_path):
+            print(f"✅ Baseline already exists at {baseline_path}")
+            return baseline_path
+
+        # Read metadata from the built index
+        meta_path = f"{index_path}.meta.json"
+        with open(meta_path) as f:
+            import json
+
+            meta = json.loads(f.read())
+
+        embedding_model = meta["embedding_model"]
+        passage_source = meta["passage_sources"][0]
+        passage_file = passage_source["path"]
+
+        # Convert relative path to absolute
+        if not os.path.isabs(passage_file):
+            index_dir = os.path.dirname(index_path)
+            passage_file = os.path.join(index_dir, os.path.basename(passage_file))
+
+        print(f"📊 Loading passages from {passage_file}...")
+        print(f"🤖 Using embedding model: {embedding_model}")
+
+        # Load all passages for baseline
+        passages = []
+        passage_ids = []
+        with open(passage_file, encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    data = json.loads(line)
+                    passages.append(data["text"])
+                    passage_ids.append(data["id"])
+
+        print(f"📄 Loaded {len(passages)} passages")
+
+        # Compute embeddings using the same method as LEANN
+        print("🧮 Computing embeddings...")
+        embeddings = compute_embeddings(
+            passages,
+            embedding_model,
+            mode="sentence-transformers",
+            use_server=False,
+        )
+
+        print(f"📐 Embedding shape: {embeddings.shape}")
+
+        # Build FAISS flat index
+        print("🏗️  Building FAISS IndexFlatIP...")
+        dimension = embeddings.shape[1]
+        index = faiss.IndexFlatIP(dimension)
+
+        # Add embeddings to flat index
+        embeddings_f32 = embeddings.astype(np.float32)
+        index.add(embeddings_f32.shape[0], faiss.swig_ptr(embeddings_f32))
+
+        # Save index and metadata
+        faiss.write_index(index, baseline_path)
+        with open(metadata_path, "wb") as f:
+            pickle.dump(passage_ids, f)
+
+        print(f"✅ FAISS baseline saved to {baseline_path}")
+        print(f"✅ Metadata saved to {metadata_path}")
+        print(f"📊 Total vectors: {index.ntotal}")
+
+        return baseline_path
+
+    def extract_pdf_text(self, pdf_path: Path) -> list[dict]:
+        """Extract and chunk text from a PDF file"""
+        chunks = []
+        doc = pymupdf.open(pdf_path)
+
+        for page_num in range(len(doc)):
+            page = doc[page_num]
+            text = page.get_text()  # type: ignore
+
+            if not text.strip():
+                continue
+
+            # Create metadata
+            metadata = {
+                "source_file": pdf_path.name,
+                "page_number": page_num + 1,
+                "document_type": "10K" if "10K" in pdf_path.name else "10Q",
+                "company": pdf_path.name.split("_")[0],
+                "doc_period": self.extract_year_from_filename(pdf_path.name),
+            }
+
+            # Use recursive character splitting like LangChain
+            if len(text.split()) > 500:
+                # Split by double newlines (paragraphs)
+                paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
+
+                current_chunk = ""
+                for para in paragraphs:
+                    # If adding this paragraph would make chunk too long, save current chunk
+                    if current_chunk and len((current_chunk + " " + para).split()) > 300:
+                        if current_chunk.strip():
+                            chunks.append(
+                                {
+                                    "text": current_chunk.strip(),
+                                    "metadata": {
+                                        **metadata,
+                                        "chunk_id": f"page_{page_num + 1}_chunk_{len(chunks)}",
+                                    },
+                                }
+                            )
+                        current_chunk = para
+                    else:
+                        current_chunk = (current_chunk + " " + para).strip()
+
+                # Add the last chunk
+                if current_chunk.strip():
+                    chunks.append(
+                        {
+                            "text": current_chunk.strip(),
+                            "metadata": {
+                                **metadata,
+                                "chunk_id": f"page_{page_num + 1}_chunk_{len(chunks)}",
+                            },
+                        }
+                    )
+            else:
+                # Page is short enough, use as single chunk
+                chunks.append(
+                    {
+                        "text": text.strip(),
+                        "metadata": {**metadata, "chunk_id": f"page_{page_num + 1}"},
+                    }
+                )
+
+        doc.close()
+        return chunks
+
+    def extract_year_from_filename(self, filename: str) -> str:
+        """Extract year from PDF filename"""
+        # Try to find 4-digit year in filename
+
+        match = re.search(r"(\d{4})", filename)
+        return match.group(1) if match else "unknown"
+
+    def verify_setup(self, index_path: str):
+        """Verify the setup by testing a simple query"""
+        print("🧪 Verifying setup with test query...")
+
+        try:
+            searcher = LeannSearcher(index_path)
+
+            # Test query
+            test_query = "What is the capital expenditure for 3M in 2018?"
+            results = searcher.search(test_query, top_k=3)
+
+            print(f"✅ Test query successful! Found {len(results)} results:")
+            for i, result in enumerate(results, 1):
+                company = result.metadata.get("company", "Unknown")
+                year = result.metadata.get("doc_period", "Unknown")
+                page = result.metadata.get("page_number", "Unknown")
+                print(f"   {i}. {company} {year} (page {page}) - Score: {result.score:.3f}")
+                print(f"      {result.text[:100]}...")
+
+            searcher.cleanup()
+            print("✅ Setup verification completed successfully!")
+
+        except Exception as e:
+            print(f"❌ Setup verification failed: {e}")
+            raise
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Setup FinanceBench with full PDF datastore")
+    parser.add_argument("--data-dir", default="data", help="Data directory")
+    parser.add_argument(
+        "--backend", choices=["hnsw", "diskann"], default="hnsw", help="LEANN backend"
+    )
+    parser.add_argument(
+        "--embedding-model",
+        default="sentence-transformers/all-mpnet-base-v2",
+        help="Embedding model",
+    )
+    parser.add_argument("--max-workers", type=int, default=5, help="Parallel download workers")
+    parser.add_argument("--skip-download", action="store_true", help="Skip PDF download")
+    parser.add_argument("--skip-build", action="store_true", help="Skip index building")
+    parser.add_argument(
+        "--build-baseline-only",
+        action="store_true",
+        help="Only build FAISS baseline from existing index",
+    )
+
+    args = parser.parse_args()
+
+    print("🏦 FinanceBench Complete Setup")
+    print("=" * 50)
+
+    setup = FinanceBenchSetup(args.data_dir)
+
+    try:
+        if args.build_baseline_only:
+            # Only build baseline from existing index
+            index_path = setup.index_dir / f"financebench_full_{args.backend}"
+            index_file = f"{index_path}.index"
+            meta_file = f"{index_path}.leann.meta.json"
+
+            if not os.path.exists(index_file) or not os.path.exists(meta_file):
+                print("❌ Index files not found:")
+                print(f"   Index: {index_file}")
+                print(f"   Meta: {meta_file}")
+                print("💡 Run without --build-baseline-only to build the index first")
+                exit(1)
+
+            print(f"🔨 Building baseline from existing index: {index_path}")
+            baseline_path = setup.build_faiss_flat_baseline(str(index_path))
+            print(f"✅ Baseline built at {baseline_path}")
+            return
+
+        # Step 1: Download dataset
+        setup.download_dataset()
+
+        # Step 2: Download PDFs
+        if not args.skip_download:
+            setup.download_all_pdfs(max_workers=args.max_workers)
+        else:
+            print("⏭️  Skipping PDF download")
+
+        # Step 3: Build LEANN index
+        if not args.skip_build:
+            index_path = setup.build_leann_index(
+                backend=args.backend, embedding_model=args.embedding_model
+            )
+
+            # Step 4: Build FAISS flat baseline
+            print("\n🔨 Building FAISS flat baseline...")
+            baseline_path = setup.build_faiss_flat_baseline(index_path)
+            print(f"✅ Baseline built at {baseline_path}")
+
+            # Step 5: Verify setup
+            setup.verify_setup(index_path)
+        else:
+            print("⏭️  Skipping index building")
+
+        print("\n🎉 FinanceBench setup completed!")
+        print(f"📁 Data directory: {setup.data_dir.absolute()}")
+        print("\nNext steps:")
+        print(
+            "1. Run evaluation: python evaluate_financebench.py --index data/index/financebench_full_hnsw.leann"
+        )
+        print(
+            "2. Or test manually: python -c \"from leann import LeannSearcher; s = LeannSearcher('data/index/financebench_full_hnsw.leann'); print(s.search('3M capital expenditure 2018'))\""
+        )
+
+    except KeyboardInterrupt:
+        print("\n⚠️  Setup interrupted by user")
+        exit(1)
+    except Exception as e:
+        print(f"\n❌ Setup failed: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/financebench/verify_recall.py

@@ -0,0 +1,214 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.9"
+# dependencies = [
+#     "faiss-cpu",
+#     "numpy",
+#     "sentence-transformers",
+#     "torch",
+#     "tqdm",
+# ]
+# ///
+
+"""
+Independent recall verification script using standard FAISS.
+Creates two indexes (HNSW and Flat) and compares recall@3 at different complexities.
+"""
+
+import json
+import time
+from pathlib import Path
+
+import faiss
+import numpy as np
+from sentence_transformers import SentenceTransformer
+from tqdm import tqdm
+
+
+def compute_embeddings_direct(chunks: list[str], model_name: str) -> np.ndarray:
+    """
+    Direct embedding computation using sentence-transformers.
+    Copied logic to avoid dependency issues.
+    """
+    print(f"Loading model: {model_name}")
+    model = SentenceTransformer(model_name)
+
+    print(f"Computing embeddings for {len(chunks)} chunks...")
+    embeddings = model.encode(
+        chunks,
+        show_progress_bar=True,
+        batch_size=32,
+        convert_to_numpy=True,
+        normalize_embeddings=False,
+    )
+
+    return embeddings.astype(np.float32)
+
+
+def load_financebench_queries(dataset_path: str, max_queries: int = 200) -> list[str]:
+    """Load FinanceBench queries from dataset"""
+    queries = []
+    with open(dataset_path, encoding="utf-8") as f:
+        for line in f:
+            if line.strip():
+                data = json.loads(line)
+                queries.append(data["question"])
+                if len(queries) >= max_queries:
+                    break
+    return queries
+
+
+def load_passages_from_leann_index(index_path: str) -> tuple[list[str], list[str]]:
+    """Load passages from LEANN index structure"""
+    meta_path = f"{index_path}.meta.json"
+    with open(meta_path) as f:
+        meta = json.load(f)
+
+    passage_source = meta["passage_sources"][0]
+    passage_file = passage_source["path"]
+
+    # Convert relative path to absolute
+    if not Path(passage_file).is_absolute():
+        index_dir = Path(index_path).parent
+        passage_file = index_dir / Path(passage_file).name
+
+    print(f"Loading passages from {passage_file}")
+
+    passages = []
+    passage_ids = []
+    with open(passage_file, encoding="utf-8") as f:
+        for line in tqdm(f, desc="Loading passages"):
+            if line.strip():
+                data = json.loads(line)
+                passages.append(data["text"])
+                passage_ids.append(data["id"])
+
+    print(f"Loaded {len(passages)} passages")
+    return passages, passage_ids
+
+
+def build_faiss_indexes(embeddings: np.ndarray) -> tuple[faiss.Index, faiss.Index]:
+    """Build FAISS indexes: Flat (ground truth) and HNSW"""
+    dimension = embeddings.shape[1]
+
+    # Build Flat index (ground truth)
+    print("Building FAISS IndexFlatIP (ground truth)...")
+    flat_index = faiss.IndexFlatIP(dimension)
+    flat_index.add(embeddings)
+
+    # Build HNSW index
+    print("Building FAISS IndexHNSWFlat...")
+    M = 32  # Same as LEANN default
+    hnsw_index = faiss.IndexHNSWFlat(dimension, M, faiss.METRIC_INNER_PRODUCT)
+    hnsw_index.hnsw.efConstruction = 200  # Same as LEANN default
+    hnsw_index.add(embeddings)
+
+    print(f"Built indexes with {flat_index.ntotal} vectors, dimension {dimension}")
+    return flat_index, hnsw_index
+
+
+def evaluate_recall_at_k(
+    query_embeddings: np.ndarray,
+    flat_index: faiss.Index,
+    hnsw_index: faiss.Index,
+    passage_ids: list[str],
+    k: int = 3,
+    ef_search: int = 64,
+) -> float:
+    """Evaluate recall@k comparing HNSW vs Flat"""
+
+    # Set search parameters for HNSW
+    hnsw_index.hnsw.efSearch = ef_search
+
+    total_recall = 0.0
+    num_queries = query_embeddings.shape[0]
+
+    for i in range(num_queries):
+        query = query_embeddings[i : i + 1]  # Keep 2D shape
+
+        # Get ground truth from Flat index (standard FAISS API)
+        flat_distances, flat_indices = flat_index.search(query, k)
+        ground_truth_ids = {passage_ids[idx] for idx in flat_indices[0]}
+
+        # Get results from HNSW index (standard FAISS API)
+        hnsw_distances, hnsw_indices = hnsw_index.search(query, k)
+        hnsw_ids = {passage_ids[idx] for idx in hnsw_indices[0]}
+
+        # Calculate recall
+        intersection = ground_truth_ids.intersection(hnsw_ids)
+        recall = len(intersection) / k
+        total_recall += recall
+
+        if i < 3:  # Show first few examples
+            print(f"  Query {i + 1}: Recall@{k} = {recall:.3f}")
+            print(f"    Flat: {list(ground_truth_ids)}")
+            print(f"    HNSW: {list(hnsw_ids)}")
+            print(f"    Intersection: {list(intersection)}")
+
+    avg_recall = total_recall / num_queries
+    return avg_recall
+
+
+def main():
+    # Configuration
+    dataset_path = "data/financebench_merged.jsonl"
+    index_path = "data/index/financebench_full_hnsw.leann"
+    embedding_model = "sentence-transformers/all-mpnet-base-v2"
+
+    print("🔍 FAISS Recall Verification")
+    print("=" * 50)
+
+    # Check if files exist
+    if not Path(dataset_path).exists():
+        print(f"❌ Dataset not found: {dataset_path}")
+        return
+    if not Path(f"{index_path}.meta.json").exists():
+        print(f"❌ Index metadata not found: {index_path}.meta.json")
+        return
+
+    # Load data
+    print("📖 Loading FinanceBench queries...")
+    queries = load_financebench_queries(dataset_path, max_queries=50)
+    print(f"Loaded {len(queries)} queries")
+
+    print("📄 Loading passages from LEANN index...")
+    passages, passage_ids = load_passages_from_leann_index(index_path)
+
+    # Compute embeddings
+    print("🧮 Computing passage embeddings...")
+    passage_embeddings = compute_embeddings_direct(passages, embedding_model)
+
+    print("🧮 Computing query embeddings...")
+    query_embeddings = compute_embeddings_direct(queries, embedding_model)
+
+    # Build FAISS indexes
+    print("🏗️ Building FAISS indexes...")
+    flat_index, hnsw_index = build_faiss_indexes(passage_embeddings)
+
+    # Test different efSearch values (equivalent to LEANN complexity)
+    print("\n📊 Evaluating Recall@3 at different efSearch values...")
+    ef_search_values = [16, 32, 64, 128, 256]
+
+    for ef_search in ef_search_values:
+        print(f"\n🧪 Testing efSearch = {ef_search}")
+        start_time = time.time()
+
+        recall = evaluate_recall_at_k(
+            query_embeddings, flat_index, hnsw_index, passage_ids, k=3, ef_search=ef_search
+        )
+
+        elapsed = time.time() - start_time
+        print(
+            f"📈 efSearch {ef_search}: Recall@3 = {recall:.3f} ({recall * 100:.1f}%) in {elapsed:.2f}s"
+        )
+
+    print("\n✅ Verification completed!")
+    print("\n📋 Summary:")
+    print("  - Built independent FAISS Flat and HNSW indexes")
+    print("  - Compared recall@3 at different efSearch values")
+    print("  - Used same embedding model as LEANN")
+    print("  - This validates LEANN's recall measurements")
+
+
+if __name__ == "__main__":
+    main()