Exclude macos-15-intel + Python 3.13 (no PyTorch wheels available)

The macos-15-intel runner runs macOS 15.7, so Homebrew libraries are
built for macOS 14+. Setting MACOSX_DEPLOYMENT_TARGET=13.0 causes
delocate to fail because system libraries require newer macOS.

Fix by setting deployment target to 15.0 for macos-15-intel, matching
the actual OS version. Intel Mac users will need macOS 15+.

.github/workflows/build-reusable.yml

@@ -35,8 +35,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - os: ubuntu-22.04
-            python: '3.9'
+          # Note: Python 3.9 dropped - uses PEP 604 union syntax (str | None)
+          # which requires Python 3.10+
           - os: ubuntu-22.04
             python: '3.10'
           - os: ubuntu-22.04
@@ -46,8 +46,6 @@ jobs:
           - os: ubuntu-22.04
             python: '3.13'
           # ARM64 Linux builds
-          - os: ubuntu-24.04-arm
-            python: '3.9'
           - os: ubuntu-24.04-arm
             python: '3.10'
           - os: ubuntu-24.04-arm
@@ -56,8 +54,6 @@ jobs:
             python: '3.12'
           - os: ubuntu-24.04-arm
             python: '3.13'
-          - os: macos-14
-            python: '3.9'
           - os: macos-14
             python: '3.10'
           - os: macos-14
@@ -66,8 +62,6 @@ jobs:
             python: '3.12'
           - os: macos-14
             python: '3.13'
-          - os: macos-15
-            python: '3.9'
           - os: macos-15
             python: '3.10'
           - os: macos-15
@@ -76,16 +70,24 @@ jobs:
             python: '3.12'
           - os: macos-15
             python: '3.13'
-          - os: macos-13
-            python: '3.9'
-          - os: macos-13
+          # Intel Mac builds (x86_64) - replaces deprecated macos-13
+          # Note: Python 3.13 excluded - PyTorch has no wheels for macOS x86_64 + Python 3.13
+          # (PyTorch <=2.4.1 lacks cp313, PyTorch >=2.5.0 dropped Intel Mac support)
+          - os: macos-15-intel
             python: '3.10'
-          - os: macos-13
+          - os: macos-15-intel
             python: '3.11'
-          - os: macos-13
+          - os: macos-15-intel
             python: '3.12'
-          # Note: macos-13 + Python 3.13 excluded due to PyTorch compatibility
-          # (PyTorch 2.5+ supports Python 3.13 but not Intel Mac x86_64)
+          # macOS 26 (beta) - arm64
+          - os: macos-26
+            python: '3.10'
+          - os: macos-26
+            python: '3.11'
+          - os: macos-26
+            python: '3.12'
+          - os: macos-26
+            python: '3.13'
     runs-on: ${{ matrix.os }}
 
     steps:
@@ -204,13 +206,16 @@ jobs:
             # Use system clang for better compatibility
             export CC=clang
             export CXX=clang++
-            # Homebrew libraries on each macOS version require matching minimum version
-            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=13.0
-            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=14.0
-            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+            # Set deployment target based on runner
+            # macos-15-intel runs macOS 15, so target 15.0 (system libraries require it)
+            if [[ "${{ matrix.os }}" == "macos-15-intel" ]]; then
               export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-14* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == macos-15* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-26* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=26.0
             fi
             uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
           else
@@ -224,14 +229,16 @@ jobs:
             # Use system clang for better compatibility
             export CC=clang
             export CXX=clang++
-            # DiskANN requires macOS 13.3+ for sgesdd_ LAPACK function
-            # But Homebrew libraries on each macOS version require matching minimum version
-            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=13.3
-            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=14.0
-            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+            # Set deployment target based on runner
+            # macos-15-intel runs macOS 15, so target 15.0 (system libraries require it)
+            if [[ "${{ matrix.os }}" == "macos-15-intel" ]]; then
               export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-14* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == macos-15* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-26* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=26.0
             fi
             uv build --wheel --python ${{ matrix.python }} --find-links ${GITHUB_WORKSPACE}/packages/leann-core/dist
           else
@@ -269,16 +276,19 @@ jobs:
         if: runner.os == 'macOS'
         run: |
           # Determine deployment target based on runner OS
-          # Must match the Homebrew libraries for each macOS version
-          if [[ "${{ matrix.os }}" == "macos-13" ]]; then
-            HNSW_TARGET="13.0"
-            DISKANN_TARGET="13.3"
-          elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-            HNSW_TARGET="14.0"
-            DISKANN_TARGET="14.0"
-          elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+          # macos-15-intel runs macOS 15, so target 15.0 (system libraries require it)
+          if [[ "${{ matrix.os }}" == "macos-15-intel" ]]; then
             HNSW_TARGET="15.0"
             DISKANN_TARGET="15.0"
+          elif [[ "${{ matrix.os }}" == macos-14* ]]; then
+            HNSW_TARGET="14.0"
+            DISKANN_TARGET="14.0"
+          elif [[ "${{ matrix.os }}" == macos-15* ]]; then
+            HNSW_TARGET="15.0"
+            DISKANN_TARGET="15.0"
+          elif [[ "${{ matrix.os }}" == macos-26* ]]; then
+            HNSW_TARGET="26.0"
+            DISKANN_TARGET="26.0"
           fi
 
           # Repair HNSW wheel
@@ -334,12 +344,15 @@ jobs:
           PY_TAG=$($UV_PY -c "import sys; print(f'cp{sys.version_info[0]}{sys.version_info[1]}')")
 
           if [[ "$RUNNER_OS" == "macOS" ]]; then
-            if [[ "${{ matrix.os }}" == "macos-13" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=13.3
-            elif [[ "${{ matrix.os }}" == "macos-14" ]]; then
-              export MACOSX_DEPLOYMENT_TARGET=14.0
-            elif [[ "${{ matrix.os }}" == "macos-15" ]]; then
+            # macos-15-intel runs macOS 15, so target 15.0 (system libraries require it)
+            if [[ "${{ matrix.os }}" == "macos-15-intel" ]]; then
               export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-14* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=14.0
+            elif [[ "${{ matrix.os }}" == macos-15* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=15.0
+            elif [[ "${{ matrix.os }}" == macos-26* ]]; then
+              export MACOSX_DEPLOYMENT_TARGET=26.0
             fi
           fi
 

README.md

@@ -36,7 +36,7 @@ LEANN is an innovative vector database that democratizes personal AI. Transform
 
 LEANN achieves this through *graph-based selective recomputation* with *high-degree preserving pruning*, computing embeddings on-demand instead of storing them all. [Illustration Fig →](#️-architecture--how-it-works) | [Paper →](https://arxiv.org/abs/2506.08276)
 
-**Ready to RAG Everything?** Transform your laptop into a personal AI assistant that can semantic search your **[file system](#-personal-data-manager-process-any-documents-pdf-txt-md)**, **[emails](#-your-personal-email-secretary-rag-on-apple-mail)**, **[browser history](#-time-machine-for-the-web-rag-your-entire-browser-history)**, **[chat history](#-wechat-detective-unlock-your-golden-memories)** ([WeChat](#-wechat-detective-unlock-your-golden-memories), [iMessage](#-imessage-history-your-personal-conversation-archive)), **[agent memory](#-chatgpt-chat-history-your-personal-ai-conversation-archive)** ([ChatGPT](#-chatgpt-chat-history-your-personal-ai-conversation-archive), [Claude](#-claude-chat-history-your-personal-ai-conversation-archive)), **[live data](#mcp-integration-rag-on-live-data-from-any-platform)** ([Slack](#mcp-integration-rag-on-live-data-from-any-platform), [Twitter](#mcp-integration-rag-on-live-data-from-any-platform)), **[codebase](#-claude-code-integration-transform-your-development-workflow)**\* , or external knowledge bases (i.e., 60M documents) - all on your laptop, with zero cloud costs and complete privacy.
+**Ready to RAG Everything?** Transform your laptop into a personal AI assistant that can semantic search your **[file system](#-personal-data-manager-process-any-documents-pdf-txt-md)**, **[emails](#-your-personal-email-secretary-rag-on-apple-mail)**, **[browser history](#-time-machine-for-the-web-rag-your-entire-browser-history)**, **[chat history](#-wechat-detective-unlock-your-golden-memories)** ([WeChat](#-wechat-detective-unlock-your-golden-memories), [iMessage](#-imessage-history-your-personal-conversation-archive)), **[agent memory](#-chatgpt-chat-history-your-personal-ai-conversation-archive)** ([ChatGPT](#-chatgpt-chat-history-your-personal-ai-conversation-archive), [Claude](#-claude-chat-history-your-personal-ai-conversation-archive)), **[live data](#mcp-integration-rag-on-live-data-from-any-platform)** ([Slack](#slack-messages-search-your-team-conversations), [Twitter](#-twitter-bookmarks-your-personal-tweet-library)), **[codebase](#-claude-code-integration-transform-your-development-workflow)**\* , or external knowledge bases (i.e., 60M documents) - all on your laptop, with zero cloud costs and complete privacy.
 
 
 \* Claude Code only supports basic `grep`-style keyword search. **LEANN** is a drop-in **semantic search MCP service fully compatible with Claude Code**, unlocking intelligent retrieval without changing your workflow. 🔥 Check out [the easy setup →](packages/leann-mcp/README.md)
@@ -392,6 +392,54 @@ python -m apps.code_rag --repo-dir "./my_codebase" --query "How does authenticat
 
 </details>
 
+### 🎨 ColQwen: Multimodal PDF Retrieval with Vision-Language Models
+
+Search through PDFs using both text and visual understanding with ColQwen2/ColPali models. Perfect for research papers, technical documents, and any PDFs with complex layouts, figures, or diagrams.
+
+> **🍎 Mac Users**: ColQwen is optimized for Apple Silicon with MPS acceleration for faster inference!
+
+```bash
+# Build index from PDFs
+python -m apps.colqwen_rag build --pdfs ./my_papers/ --index research_papers
+
+# Search with text queries
+python -m apps.colqwen_rag search research_papers "How does attention mechanism work?"
+
+# Interactive Q&A
+python -m apps.colqwen_rag ask research_papers --interactive
+```
+
+<details>
+<summary><strong>📋 Click to expand: ColQwen Setup & Usage</strong></summary>
+
+#### Prerequisites
+```bash
+# Install dependencies
+uv pip install colpali_engine pdf2image pillow matplotlib qwen_vl_utils einops seaborn
+brew install poppler  # macOS only, for PDF processing
+```
+
+#### Build Index
+```bash
+python -m apps.colqwen_rag build \
+  --pdfs ./pdf_directory/ \
+  --index my_index \
+  --model colqwen2  # or colpali
+```
+
+#### Search
+```bash
+python -m apps.colqwen_rag search my_index "your question here" --top-k 5
+```
+
+#### Models
+- **ColQwen2** (`colqwen2`): Latest vision-language model with improved performance
+- **ColPali** (`colpali`): Proven multimodal retriever
+
+For detailed usage, see the [ColQwen Guide](docs/COLQWEN_GUIDE.md).
+
+</details>
+
 ### 📧 Your Personal Email Secretary: RAG on Apple Mail!
 
 > **Note:** The examples below currently support macOS only. Windows support coming soon.

apps/base_rag_example.py

@@ -6,7 +6,7 @@ Provides common parameters and functionality for all RAG examples.
 import argparse
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Any
+from typing import Any, Union
 
 import dotenv
 from leann.api import LeannBuilder, LeannChat
@@ -257,8 +257,8 @@ class BaseRAGExample(ABC):
         pass
 
     @abstractmethod
-    async def load_data(self, args) -> list[str]:
-        """Load data from the source. Returns list of text chunks."""
+    async def load_data(self, args) -> list[Union[str, dict[str, Any]]]:
+        """Load data from the source. Returns list of text chunks (strings or dicts with 'text' key)."""
         pass
 
     def get_llm_config(self, args) -> dict[str, Any]:
@@ -282,8 +282,8 @@ class BaseRAGExample(ABC):
 
         return config
 
-    async def build_index(self, args, texts: list[str]) -> str:
-        """Build LEANN index from texts."""
+    async def build_index(self, args, texts: list[Union[str, dict[str, Any]]]) -> str:
+        """Build LEANN index from texts (accepts strings or dicts with 'text' key)."""
         index_path = str(Path(args.index_dir) / f"{self.default_index_name}.leann")
 
         print(f"\n[Building Index] Creating {self.name} index...")
@@ -314,8 +314,14 @@ class BaseRAGExample(ABC):
         batch_size = 1000
         for i in range(0, len(texts), batch_size):
             batch = texts[i : i + batch_size]
-            for text in batch:
-                builder.add_text(text)
+            for item in batch:
+                # Handle both dict format (from create_text_chunks) and plain strings
+                if isinstance(item, dict):
+                    text = item.get("text", "")
+                    metadata = item.get("metadata")
+                    builder.add_text(text, metadata)
+                else:
+                    builder.add_text(item)
             print(f"Added {min(i + batch_size, len(texts))}/{len(texts)} texts...")
 
         print("Building index structure...")

apps/colqwen_rag.py

@@ -0,0 +1,364 @@
+#!/usr/bin/env python3
+"""
+ColQwen RAG - Easy-to-use multimodal PDF retrieval with ColQwen2/ColPali
+
+Usage:
+    python -m apps.colqwen_rag build --pdfs ./my_pdfs/ --index my_index
+    python -m apps.colqwen_rag search my_index "How does attention work?"
+    python -m apps.colqwen_rag ask my_index --interactive
+"""
+
+import argparse
+import os
+import sys
+from pathlib import Path
+from typing import Optional, cast
+
+# Add LEANN packages to path
+_repo_root = Path(__file__).resolve().parents[1]
+_leann_core_src = _repo_root / "packages" / "leann-core" / "src"
+_leann_hnsw_pkg = _repo_root / "packages" / "leann-backend-hnsw"
+if str(_leann_core_src) not in sys.path:
+    sys.path.append(str(_leann_core_src))
+if str(_leann_hnsw_pkg) not in sys.path:
+    sys.path.append(str(_leann_hnsw_pkg))
+
+import torch  # noqa: E402
+from colpali_engine import ColPali, ColPaliProcessor, ColQwen2, ColQwen2Processor  # noqa: E402
+from colpali_engine.utils.torch_utils import ListDataset  # noqa: E402
+from pdf2image import convert_from_path  # noqa: E402
+from PIL import Image  # noqa: E402
+from torch.utils.data import DataLoader  # noqa: E402
+from tqdm import tqdm  # noqa: E402
+
+# Import the existing multi-vector implementation
+sys.path.append(str(_repo_root / "apps" / "multimodal" / "vision-based-pdf-multi-vector"))
+from leann_multi_vector import LeannMultiVector  # noqa: E402
+
+
+class ColQwenRAG:
+    """Easy-to-use ColQwen RAG system for multimodal PDF retrieval."""
+
+    def __init__(self, model_type: str = "colpali"):
+        """
+        Initialize ColQwen RAG system.
+
+        Args:
+            model_type: "colqwen2" or "colpali"
+        """
+        self.model_type = model_type
+        self.device = self._get_device()
+        # Use float32 on MPS to avoid memory issues, float16 on CUDA, bfloat16 on CPU
+        if self.device.type == "mps":
+            self.dtype = torch.float32
+        elif self.device.type == "cuda":
+            self.dtype = torch.float16
+        else:
+            self.dtype = torch.bfloat16
+
+        print(f"🚀 Initializing {model_type.upper()} on {self.device} with {self.dtype}")
+
+        # Load model and processor with MPS-optimized settings
+        try:
+            if model_type == "colqwen2":
+                self.model_name = "vidore/colqwen2-v1.0"
+                if self.device.type == "mps":
+                    # For MPS, load on CPU first then move to avoid memory allocation issues
+                    self.model = ColQwen2.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map="cpu",
+                        low_cpu_mem_usage=True,
+                    ).eval()
+                    self.model = self.model.to(self.device)
+                else:
+                    self.model = ColQwen2.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map=self.device,
+                        low_cpu_mem_usage=True,
+                    ).eval()
+                self.processor = ColQwen2Processor.from_pretrained(self.model_name)
+            else:  # colpali
+                self.model_name = "vidore/colpali-v1.2"
+                if self.device.type == "mps":
+                    # For MPS, load on CPU first then move to avoid memory allocation issues
+                    self.model = ColPali.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map="cpu",
+                        low_cpu_mem_usage=True,
+                    ).eval()
+                    self.model = self.model.to(self.device)
+                else:
+                    self.model = ColPali.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map=self.device,
+                        low_cpu_mem_usage=True,
+                    ).eval()
+                self.processor = ColPaliProcessor.from_pretrained(self.model_name)
+        except Exception as e:
+            if "memory" in str(e).lower() or "offload" in str(e).lower():
+                print(f"⚠️  Memory constraint on {self.device}, using CPU with optimizations...")
+                self.device = torch.device("cpu")
+                self.dtype = torch.float32
+
+                if model_type == "colqwen2":
+                    self.model = ColQwen2.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map="cpu",
+                        low_cpu_mem_usage=True,
+                    ).eval()
+                else:
+                    self.model = ColPali.from_pretrained(
+                        self.model_name,
+                        torch_dtype=self.dtype,
+                        device_map="cpu",
+                        low_cpu_mem_usage=True,
+                    ).eval()
+            else:
+                raise
+
+    def _get_device(self):
+        """Auto-select best available device."""
+        if torch.cuda.is_available():
+            return torch.device("cuda")
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            return torch.device("mps")
+        else:
+            return torch.device("cpu")
+
+    def build_index(self, pdf_paths: list[str], index_name: str, pages_dir: Optional[str] = None):
+        """
+        Build multimodal index from PDF files.
+
+        Args:
+            pdf_paths: List of PDF file paths
+            index_name: Name for the index
+            pages_dir: Directory to save page images (optional)
+        """
+        print(f"Building index '{index_name}' from {len(pdf_paths)} PDFs...")
+
+        # Convert PDFs to images
+        all_images = []
+        all_metadata = []
+
+        if pages_dir:
+            os.makedirs(pages_dir, exist_ok=True)
+
+        for pdf_path in tqdm(pdf_paths, desc="Converting PDFs"):
+            try:
+                images = convert_from_path(pdf_path, dpi=150)
+                pdf_name = Path(pdf_path).stem
+
+                for i, image in enumerate(images):
+                    # Save image if pages_dir specified
+                    if pages_dir:
+                        image_path = Path(pages_dir) / f"{pdf_name}_page_{i + 1}.png"
+                        image.save(image_path)
+
+                    all_images.append(image)
+                    all_metadata.append(
+                        {
+                            "pdf_path": pdf_path,
+                            "pdf_name": pdf_name,
+                            "page_number": i + 1,
+                            "image_path": str(image_path) if pages_dir else None,
+                        }
+                    )
+
+            except Exception as e:
+                print(f"❌ Error processing {pdf_path}: {e}")
+                continue
+
+        print(f"📄 Converted {len(all_images)} pages from {len(pdf_paths)} PDFs")
+        print(f"All metadata: {all_metadata}")
+
+        # Generate embeddings
+        print("🧠 Generating embeddings...")
+        embeddings = self._embed_images(all_images)
+
+        # Build LEANN index
+        print("🔍 Building LEANN index...")
+        leann_mv = LeannMultiVector(
+            index_path=index_name,
+            dim=embeddings.shape[-1],
+            embedding_model_name=self.model_type,
+        )
+
+        # Create collection and insert data
+        leann_mv.create_collection()
+        for i, (embedding, metadata) in enumerate(zip(embeddings, all_metadata)):
+            data = {
+                "doc_id": i,
+                "filepath": metadata.get("image_path", ""),
+                "colbert_vecs": embedding.numpy(),  # Convert tensor to numpy
+            }
+            leann_mv.insert(data)
+
+        # Build the index
+        leann_mv.create_index()
+        print(f"✅ Index '{index_name}' built successfully!")
+
+        return leann_mv
+
+    def search(self, index_name: str, query: str, top_k: int = 5):
+        """
+        Search the index with a text query.
+
+        Args:
+            index_name: Name of the index to search
+            query: Text query
+            top_k: Number of results to return
+        """
+        print(f"🔍 Searching '{index_name}' for: '{query}'")
+
+        # Load index
+        leann_mv = LeannMultiVector(
+            index_path=index_name,
+            dim=128,  # Will be updated when loading
+            embedding_model_name=self.model_type,
+        )
+
+        # Generate query embedding
+        query_embedding = self._embed_query(query)
+
+        # Search (returns list of (score, doc_id) tuples)
+        search_results = leann_mv.search(query_embedding.numpy(), topk=top_k)
+
+        # Display results
+        print(f"\n📋 Top {len(search_results)} results:")
+        for i, (score, doc_id) in enumerate(search_results, 1):
+            # Get metadata for this doc_id (we need to load the metadata)
+            print(f"{i}. Score: {score:.3f} | Doc ID: {doc_id}")
+
+        return search_results
+
+    def ask(self, index_name: str, interactive: bool = False):
+        """
+        Interactive Q&A with the indexed documents.
+
+        Args:
+            index_name: Name of the index to query
+            interactive: Whether to run in interactive mode
+        """
+        print(f"💬 ColQwen Chat with '{index_name}'")
+
+        if interactive:
+            print("Type 'quit' to exit, 'help' for commands")
+            while True:
+                try:
+                    query = input("\n🤔 Your question: ").strip()
+                    if query.lower() in ["quit", "exit", "q"]:
+                        break
+                    elif query.lower() == "help":
+                        print("Commands: quit/exit/q (exit), help (this message)")
+                        continue
+                    elif not query:
+                        continue
+
+                    self.search(index_name, query, top_k=3)
+
+                    # TODO: Add answer generation with Qwen-VL
+                    print("\n💡 For detailed answers, we can integrate Qwen-VL here!")
+
+                except KeyboardInterrupt:
+                    print("\n👋 Goodbye!")
+                    break
+        else:
+            query = input("🤔 Your question: ").strip()
+            if query:
+                self.search(index_name, query)
+
+    def _embed_images(self, images: list[Image.Image]) -> torch.Tensor:
+        """Generate embeddings for a list of images."""
+        dataset = ListDataset(images)
+        dataloader = DataLoader(dataset, batch_size=1, shuffle=False, collate_fn=lambda x: x)
+
+        embeddings = []
+        with torch.no_grad():
+            for batch in tqdm(dataloader, desc="Embedding images"):
+                batch_images = cast(list, batch)
+                batch_inputs = self.processor.process_images(batch_images).to(self.device)
+                batch_embeddings = self.model(**batch_inputs)
+                embeddings.append(batch_embeddings.cpu())
+
+        return torch.cat(embeddings, dim=0)
+
+    def _embed_query(self, query: str) -> torch.Tensor:
+        """Generate embedding for a text query."""
+        with torch.no_grad():
+            query_inputs = self.processor.process_queries([query]).to(self.device)
+            query_embedding = self.model(**query_inputs)
+            return query_embedding.cpu()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="ColQwen RAG - Easy multimodal PDF retrieval")
+    subparsers = parser.add_subparsers(dest="command", help="Available commands")
+
+    # Build command
+    build_parser = subparsers.add_parser("build", help="Build index from PDFs")
+    build_parser.add_argument("--pdfs", required=True, help="Directory containing PDF files")
+    build_parser.add_argument("--index", required=True, help="Index name")
+    build_parser.add_argument(
+        "--model", choices=["colqwen2", "colpali"], default="colqwen2", help="Model to use"
+    )
+    build_parser.add_argument("--pages-dir", help="Directory to save page images")
+
+    # Search command
+    search_parser = subparsers.add_parser("search", help="Search the index")
+    search_parser.add_argument("index", help="Index name")
+    search_parser.add_argument("query", help="Search query")
+    search_parser.add_argument("--top-k", type=int, default=5, help="Number of results")
+    search_parser.add_argument(
+        "--model", choices=["colqwen2", "colpali"], default="colqwen2", help="Model to use"
+    )
+
+    # Ask command
+    ask_parser = subparsers.add_parser("ask", help="Interactive Q&A")
+    ask_parser.add_argument("index", help="Index name")
+    ask_parser.add_argument("--interactive", action="store_true", help="Interactive mode")
+    ask_parser.add_argument(
+        "--model", choices=["colqwen2", "colpali"], default="colqwen2", help="Model to use"
+    )
+
+    args = parser.parse_args()
+
+    if not args.command:
+        parser.print_help()
+        return
+
+    # Initialize ColQwen RAG
+    if args.command == "build":
+        colqwen = ColQwenRAG(args.model)
+
+        # Get PDF files
+        pdf_dir = Path(args.pdfs)
+        if pdf_dir.is_file() and pdf_dir.suffix.lower() == ".pdf":
+            pdf_paths = [str(pdf_dir)]
+        elif pdf_dir.is_dir():
+            pdf_paths = [str(p) for p in pdf_dir.glob("*.pdf")]
+        else:
+            print(f"❌ Invalid PDF path: {args.pdfs}")
+            return
+
+        if not pdf_paths:
+            print(f"❌ No PDF files found in {args.pdfs}")
+            return
+
+        colqwen.build_index(pdf_paths, args.index, args.pages_dir)
+
+    elif args.command == "search":
+        colqwen = ColQwenRAG(args.model)
+        colqwen.search(args.index, args.query, args.top_k)
+
+    elif args.command == "ask":
+        colqwen = ColQwenRAG(args.model)
+        colqwen.ask(args.index, args.interactive)
+
+
+if __name__ == "__main__":
+    main()

apps/document_rag.py

@@ -5,6 +5,7 @@ Supports PDF, TXT, MD, and other document formats.
 
 import sys
 from pathlib import Path
+from typing import Any, Union
 
 # Add parent directory to path for imports
 sys.path.insert(0, str(Path(__file__).parent))
@@ -51,7 +52,7 @@ class DocumentRAG(BaseRAGExample):
             help="Enable AST-aware chunking for code files in the data directory",
         )
 
-    async def load_data(self, args) -> list[str]:
+    async def load_data(self, args) -> list[Union[str, dict[str, Any]]]:
         """Load documents and convert to text chunks."""
         print(f"Loading documents from: {args.data_dir}")
         if args.file_types:

apps/image_rag.py

@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+"""
+CLIP Image RAG Application
+
+This application enables RAG (Retrieval-Augmented Generation) on images using CLIP embeddings.
+You can index a directory of images and search them using text queries.
+
+Usage:
+    python -m apps.image_rag --image-dir ./my_images/ --query "a sunset over mountains"
+    python -m apps.image_rag --image-dir ./my_images/ --interactive
+"""
+
+import argparse
+import pickle
+import tempfile
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+from sentence_transformers import SentenceTransformer
+from tqdm import tqdm
+
+from apps.base_rag_example import BaseRAGExample
+
+
+class ImageRAG(BaseRAGExample):
+    """
+    RAG application for images using CLIP embeddings.
+
+    This class provides a complete RAG pipeline for image data, including
+    CLIP embedding generation, indexing, and text-based image search.
+    """
+
+    def __init__(self):
+        super().__init__(
+            name="Image RAG",
+            description="RAG application for images using CLIP embeddings",
+            default_index_name="image_index",
+        )
+        # Override default embedding model to use CLIP
+        self.embedding_model_default = "clip-ViT-L-14"
+        self.embedding_mode_default = "sentence-transformers"
+        self._image_data: list[dict] = []
+
+    def _add_specific_arguments(self, parser: argparse.ArgumentParser):
+        """Add image-specific arguments."""
+        image_group = parser.add_argument_group("Image Parameters")
+        image_group.add_argument(
+            "--image-dir",
+            type=str,
+            required=True,
+            help="Directory containing images to index",
+        )
+        image_group.add_argument(
+            "--image-extensions",
+            type=str,
+            nargs="+",
+            default=[".jpg", ".jpeg", ".png", ".gif", ".bmp", ".webp"],
+            help="Image file extensions to process (default: .jpg .jpeg .png .gif .bmp .webp)",
+        )
+        image_group.add_argument(
+            "--batch-size",
+            type=int,
+            default=32,
+            help="Batch size for CLIP embedding generation (default: 32)",
+        )
+
+    async def load_data(self, args) -> list[str]:
+        """Load images, generate CLIP embeddings, and return text descriptions."""
+        self._image_data = self._load_images_and_embeddings(args)
+        return [entry["text"] for entry in self._image_data]
+
+    def _load_images_and_embeddings(self, args) -> list[dict]:
+        """Helper to process images and produce embeddings/metadata."""
+        image_dir = Path(args.image_dir)
+        if not image_dir.exists():
+            raise ValueError(f"Image directory does not exist: {image_dir}")
+
+        print(f"📸 Loading images from {image_dir}...")
+
+        # Find all image files
+        image_files = []
+        for ext in args.image_extensions:
+            image_files.extend(image_dir.rglob(f"*{ext}"))
+            image_files.extend(image_dir.rglob(f"*{ext.upper()}"))
+
+        if not image_files:
+            raise ValueError(
+                f"No images found in {image_dir} with extensions {args.image_extensions}"
+            )
+
+        print(f"✅ Found {len(image_files)} images")
+
+        # Limit if max_items is set
+        if args.max_items > 0:
+            image_files = image_files[: args.max_items]
+            print(f"📊 Processing {len(image_files)} images (limited by --max-items)")
+
+        # Load CLIP model
+        print("🔍 Loading CLIP model...")
+        model = SentenceTransformer(self.embedding_model_default)
+
+        # Process images and generate embeddings
+        print("🖼️  Processing images and generating embeddings...")
+        image_data = []
+        batch_images = []
+        batch_paths = []
+
+        for image_path in tqdm(image_files, desc="Processing images"):
+            try:
+                image = Image.open(image_path).convert("RGB")
+                batch_images.append(image)
+                batch_paths.append(image_path)
+
+                # Process in batches
+                if len(batch_images) >= args.batch_size:
+                    embeddings = model.encode(
+                        batch_images,
+                        convert_to_numpy=True,
+                        normalize_embeddings=True,
+                        batch_size=args.batch_size,
+                        show_progress_bar=False,
+                    )
+
+                    for img_path, embedding in zip(batch_paths, embeddings):
+                        image_data.append(
+                            {
+                                "text": f"Image: {img_path.name}\nPath: {img_path}",
+                                "metadata": {
+                                    "image_path": str(img_path),
+                                    "image_name": img_path.name,
+                                    "image_dir": str(image_dir),
+                                },
+                                "embedding": embedding.astype(np.float32),
+                            }
+                        )
+
+                    batch_images = []
+                    batch_paths = []
+
+            except Exception as e:
+                print(f"⚠️  Failed to process {image_path}: {e}")
+                continue
+
+        # Process remaining images
+        if batch_images:
+            embeddings = model.encode(
+                batch_images,
+                convert_to_numpy=True,
+                normalize_embeddings=True,
+                batch_size=len(batch_images),
+                show_progress_bar=False,
+            )
+
+            for img_path, embedding in zip(batch_paths, embeddings):
+                image_data.append(
+                    {
+                        "text": f"Image: {img_path.name}\nPath: {img_path}",
+                        "metadata": {
+                            "image_path": str(img_path),
+                            "image_name": img_path.name,
+                            "image_dir": str(image_dir),
+                        },
+                        "embedding": embedding.astype(np.float32),
+                    }
+                )
+
+        print(f"✅ Processed {len(image_data)} images")
+        return image_data
+
+    async def build_index(self, args, texts: list[str]) -> str:
+        """Build index using pre-computed CLIP embeddings."""
+        from leann.api import LeannBuilder
+
+        if not self._image_data or len(self._image_data) != len(texts):
+            raise RuntimeError("No image data found. Make sure load_data() ran successfully.")
+
+        print("🔨 Building LEANN index with CLIP embeddings...")
+        builder = LeannBuilder(
+            backend_name=args.backend_name,
+            embedding_model=self.embedding_model_default,
+            embedding_mode=self.embedding_mode_default,
+            is_recompute=False,
+            distance_metric="cosine",
+            graph_degree=args.graph_degree,
+            build_complexity=args.build_complexity,
+            is_compact=not args.no_compact,
+        )
+
+        for text, data in zip(texts, self._image_data):
+            builder.add_text(text=text, metadata=data["metadata"])
+
+        ids = [str(i) for i in range(len(self._image_data))]
+        embeddings = np.array([data["embedding"] for data in self._image_data], dtype=np.float32)
+
+        with tempfile.NamedTemporaryFile(mode="wb", suffix=".pkl", delete=False) as f:
+            pickle.dump((ids, embeddings), f)
+            pkl_path = f.name
+
+        try:
+            index_path = str(Path(args.index_dir) / f"{self.default_index_name}.leann")
+            builder.build_index_from_embeddings(index_path, pkl_path)
+            print(f"✅ Index built successfully at {index_path}")
+            return index_path
+        finally:
+            Path(pkl_path).unlink()
+
+
+def main():
+    """Main entry point for the image RAG application."""
+    import asyncio
+
+    app = ImageRAG()
+    asyncio.run(app.run())
+
+
+if __name__ == "__main__":
+    main()

apps/multimodal/vision-based-pdf-multi-vector/leann_multi_vector.py

@@ -1,5 +1,7 @@
 import concurrent.futures
+import glob
 import json
+import logging
 import os
 import re
 import sys
@@ -11,6 +13,8 @@ import numpy as np
 from PIL import Image
 from tqdm import tqdm
 
+logger = logging.getLogger(__name__)
+
 
 def _ensure_repo_paths_importable(current_file: str) -> None:
     """Make local leann packages importable without installing (mirrors multi-vector-leann.py)."""
@@ -96,12 +100,63 @@ def _natural_sort_key(name: str) -> int:
     return int(m.group()) if m else 0
 
 
-def _load_images_from_dir(pages_dir: str) -> tuple[list[str], list[Image.Image]]:
-    filenames = [n for n in os.listdir(pages_dir) if n.lower().endswith((".png", ".jpg", ".jpeg"))]
-    filenames = sorted(filenames, key=_natural_sort_key)
-    filepaths = [os.path.join(pages_dir, n) for n in filenames]
-    images = [Image.open(p) for p in filepaths]
-    return filepaths, images
+def _load_images_from_dir(
+    pages_dir: str, recursive: bool = False
+) -> tuple[list[str], list[Image.Image]]:
+    """
+    Load images from a directory.
+
+    Args:
+        pages_dir: Directory path containing images
+        recursive: If True, recursively search subdirectories (default: False)
+
+    Returns:
+        Tuple of (filepaths, images)
+    """
+
+    # Supported image extensions
+    extensions = ("*.png", "*.jpg", "*.jpeg", "*.PNG", "*.JPG", "*.JPEG", "*.webp", "*.WEBP")
+
+    if recursive:
+        # Recursive search
+        filepaths = []
+        for ext in extensions:
+            pattern = os.path.join(pages_dir, "**", ext)
+            filepaths.extend(glob.glob(pattern, recursive=True))
+    else:
+        # Non-recursive search (only top-level directory)
+        filepaths = []
+        for ext in extensions:
+            pattern = os.path.join(pages_dir, ext)
+            filepaths.extend(glob.glob(pattern))
+
+    # Sort files naturally
+    filepaths = sorted(filepaths, key=lambda x: _natural_sort_key(os.path.basename(x)))
+
+    # Load images with error handling
+    images = []
+    valid_filepaths = []
+    failed_count = 0
+
+    for filepath in filepaths:
+        try:
+            img = Image.open(filepath)
+            # Convert to RGB if necessary (handles RGBA, P, etc.)
+            if img.mode != "RGB":
+                img = img.convert("RGB")
+            images.append(img)
+            valid_filepaths.append(filepath)
+        except Exception as e:
+            failed_count += 1
+            print(f"Warning: Failed to load image {filepath}: {e}")
+            continue
+
+    if failed_count > 0:
+        print(
+            f"Warning: Failed to load {failed_count} image(s) out of {len(filepaths)} total files"
+        )
+
+    return valid_filepaths, images
 
 
 def _maybe_convert_pdf_to_images(pdf_path: Optional[str], pages_dir: str, dpi: int = 200) -> None:
@@ -151,6 +206,8 @@ def _select_device_and_dtype():
 
 
 def _load_colvision(model_choice: str):
+    import os
+
     import torch
     from colpali_engine.models import (
         ColPali,
@@ -162,6 +219,16 @@ def _load_colvision(model_choice: str):
     from colpali_engine.models.paligemma.colpali.processing_colpali import ColPaliProcessor
     from transformers.utils.import_utils import is_flash_attn_2_available
 
+    # Force HuggingFace Hub to use HF endpoint, avoid Google Drive
+    # Set environment variables to ensure models are downloaded from HuggingFace
+    os.environ.setdefault("HF_ENDPOINT", "https://huggingface.co")
+    os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
+
+    # Log model loading info
+    logger.info(f"Loading ColVision model: {model_choice}")
+    logger.info(f"HF_ENDPOINT: {os.environ.get('HF_ENDPOINT', 'not set')}")
+    logger.info("Models will be downloaded from HuggingFace Hub, not Google Drive")
+
     device_str, device, dtype = _select_device_and_dtype()
 
     # Determine model name and type
@@ -202,29 +269,36 @@ def _load_colvision(model_choice: str):
         "flash_attention_2" if (device_str == "cuda" and is_flash_attn_2_available()) else "eager"
     )
 
+    # Load model from HuggingFace Hub (not Google Drive)
+    # Use local_files_only=False to ensure download from HF if not cached
     if model_type == "colqwen2.5":
         model = ColQwen2_5.from_pretrained(
             model_name,
             torch_dtype=torch.bfloat16,
             device_map=device,
             attn_implementation=attn_implementation,
+            local_files_only=False,  # Ensure download from HuggingFace Hub
         ).eval()
-        processor = ColQwen2_5_Processor.from_pretrained(model_name)
+        processor = ColQwen2_5_Processor.from_pretrained(model_name, local_files_only=False)
     elif model_type == "colqwen2":
         model = ColQwen2.from_pretrained(
             model_name,
             torch_dtype=torch.bfloat16,
             device_map=device,
             attn_implementation=attn_implementation,
+            local_files_only=False,  # Ensure download from HuggingFace Hub
         ).eval()
-        processor = ColQwen2Processor.from_pretrained(model_name)
+        processor = ColQwen2Processor.from_pretrained(model_name, local_files_only=False)
     else:  # colpali
         model = ColPali.from_pretrained(
             model_name,
             torch_dtype=torch.bfloat16,
             device_map=device,
+            local_files_only=False,  # Ensure download from HuggingFace Hub
         ).eval()
-        processor = cast(ColPaliProcessor, ColPaliProcessor.from_pretrained(model_name))
+        processor = cast(
+            ColPaliProcessor, ColPaliProcessor.from_pretrained(model_name, local_files_only=False)
+        )
 
     return model_name, model, processor, device_str, device, dtype
 

apps/multimodal/vision-based-pdf-multi-vector/multi-vector-leann-similarity-map.py

@@ -62,7 +62,7 @@ DATASET_NAME: str = "weaviate/arXiv-AI-papers-multi-vector"
 # DATASET_NAMES: Optional[list[str | tuple[str, Optional[str]]]] = None
 DATASET_NAMES = [
     "weaviate/arXiv-AI-papers-multi-vector",
-    ("lmms-lab/DocVQA", "DocVQA"),  # Specify config name for datasets with multiple configs
+    # ("lmms-lab/DocVQA", "DocVQA"),  # Specify config name for datasets with multiple configs
 ]
 # Load multiple splits to get more data (e.g., ["train", "test", "validation"])
 # Set to None to try loading all available splits automatically
@@ -75,6 +75,11 @@ MAX_DOCS: Optional[int] = None  # limit number of pages to index; None = all
 # Local pages (used when USE_HF_DATASET == False)
 PDF: Optional[str] = None  # e.g., "./pdfs/2004.12832v2.pdf"
 PAGES_DIR: str = "./pages"
+# Custom folder path (takes precedence over USE_HF_DATASET and PAGES_DIR)
+# If set, images will be loaded directly from this folder
+CUSTOM_FOLDER_PATH: Optional[str] = None  # e.g., "/home/ubuntu/dr-tulu/agent/screenshots"
+# Whether to recursively search subdirectories when loading from custom folder
+CUSTOM_FOLDER_RECURSIVE: bool = False  # Set to True to search subdirectories
 
 # Index + retrieval settings
 # Use a different index path for larger dataset to avoid overwriting existing index
@@ -83,7 +88,7 @@ INDEX_PATH: str = "./indexes/colvision_large.leann"
 # These are now command-line arguments (see CLI overrides section)
 TOPK: int = 3
 FIRST_STAGE_K: int = 500
-REBUILD_INDEX: bool = True
+REBUILD_INDEX: bool = False  # Set to True to force rebuild even if index exists
 
 # Artifacts
 SAVE_TOP_IMAGE: Optional[str] = "./figures/retrieved_page.png"
@@ -128,12 +133,33 @@ parser.add_argument(
     default=TOPK,
     help=f"Number of top results to retrieve. Default: {TOPK}",
 )
+parser.add_argument(
+    "--custom-folder",
+    type=str,
+    default=None,
+    help="Path to a custom folder containing images to search. Takes precedence over dataset loading. Default: None",
+)
+parser.add_argument(
+    "--recursive",
+    action="store_true",
+    default=False,
+    help="Recursively search subdirectories when loading images from custom folder. Default: False",
+)
+parser.add_argument(
+    "--rebuild-index",
+    action="store_true",
+    default=False,
+    help="Force rebuild the index even if it already exists. Default: False (reuse existing index if available)",
+)
 cli_args, _unknown = parser.parse_known_args()
 SEARCH_METHOD: str = cli_args.search_method
 QUERY = cli_args.query  # Override QUERY with CLI argument if provided
 USE_FAST_PLAID: bool = cli_args.use_fast_plaid
 FAST_PLAID_INDEX_PATH: str = cli_args.fast_plaid_index_path
 TOPK: int = cli_args.topk  # Override TOPK with CLI argument if provided
+CUSTOM_FOLDER_PATH = cli_args.custom_folder if cli_args.custom_folder else CUSTOM_FOLDER_PATH  # Override with CLI argument if provided
+CUSTOM_FOLDER_RECURSIVE = cli_args.recursive if cli_args.recursive else CUSTOM_FOLDER_RECURSIVE  # Override with CLI argument if provided
+REBUILD_INDEX = cli_args.rebuild_index  # Override REBUILD_INDEX with CLI argument
 
 # %%
 
@@ -180,7 +206,23 @@ else:
 # Step 2: Load data only if we need to build the index
 if need_to_build_index:
     print("Loading dataset...")
-    if USE_HF_DATASET:
+    # Check for custom folder path first (takes precedence)
+    if CUSTOM_FOLDER_PATH:
+        if not os.path.isdir(CUSTOM_FOLDER_PATH):
+            raise RuntimeError(f"Custom folder path does not exist: {CUSTOM_FOLDER_PATH}")
+        print(f"Loading images from custom folder: {CUSTOM_FOLDER_PATH}")
+        if CUSTOM_FOLDER_RECURSIVE:
+            print("  (recursive mode: searching subdirectories)")
+        filepaths, images = _load_images_from_dir(CUSTOM_FOLDER_PATH, recursive=CUSTOM_FOLDER_RECURSIVE)
+        print(f"  Found {len(filepaths)} image files")
+        if not images:
+            raise RuntimeError(
+                f"No images found in {CUSTOM_FOLDER_PATH}. Ensure the folder contains image files (.png, .jpg, .jpeg, .webp)."
+            )
+        print(f"  Successfully loaded {len(images)} images")
+        # Use filenames as identifiers instead of full paths for cleaner metadata
+        filepaths = [os.path.basename(fp) for fp in filepaths]
+    elif USE_HF_DATASET:
         from datasets import load_dataset, concatenate_datasets, DatasetDict
 
         # Determine which datasets to load
@@ -621,7 +663,6 @@ else:
             except Exception:
                 print(f"Saved retrieved page (rank {rank}) to: {out_path}")
 
-## TODO stange results of second page of DeepSeek-V2 rather than the first page
 
 # %%
 # Step 6: Similarity maps for top-K results

docs/COLQWEN_GUIDE.md

@@ -0,0 +1,200 @@
+# ColQwen Integration Guide
+
+Easy-to-use multimodal PDF retrieval with ColQwen2/ColPali models.
+
+## Quick Start
+
+> **🍎 Mac Users**: ColQwen is optimized for Apple Silicon with MPS acceleration for faster inference!
+
+### 1. Install Dependencies
+```bash
+uv pip install colpali_engine pdf2image pillow matplotlib qwen_vl_utils einops seaborn
+brew install poppler  # macOS only, for PDF processing
+```
+
+### 2. Basic Usage
+```bash
+# Build index from PDFs
+python -m apps.colqwen_rag build --pdfs ./my_papers/ --index research_papers
+
+# Search with text queries
+python -m apps.colqwen_rag search research_papers "How does attention mechanism work?"
+
+# Interactive Q&A
+python -m apps.colqwen_rag ask research_papers --interactive
+```
+
+## Commands
+
+### Build Index
+```bash
+python -m apps.colqwen_rag build \
+  --pdfs ./pdf_directory/ \
+  --index my_index \
+  --model colqwen2 \
+  --pages-dir ./page_images/  # Optional: save page images
+```
+
+**Options:**
+- `--pdfs`: Directory containing PDF files (or single PDF path)
+- `--index`: Name for the index (required)
+- `--model`: `colqwen2` (default) or `colpali`
+- `--pages-dir`: Directory to save page images (optional)
+
+### Search Index
+```bash
+python -m apps.colqwen_rag search my_index "your question here" --top-k 5
+```
+
+**Options:**
+- `--top-k`: Number of results to return (default: 5)
+- `--model`: Model used for search (should match build model)
+
+### Interactive Q&A
+```bash
+python -m apps.colqwen_rag ask my_index --interactive
+```
+
+**Commands in interactive mode:**
+- Type your questions naturally
+- `help`: Show available commands
+- `quit`/`exit`/`q`: Exit interactive mode
+
+## 🧪 Test & Reproduce Results
+
+Run the reproduction test for issue #119:
+```bash
+python test_colqwen_reproduction.py
+```
+
+This will:
+1. ✅ Check dependencies
+2. 📥 Download sample PDF (Attention Is All You Need paper)
+3. 🏗️ Build test index
+4. 🔍 Run sample queries
+5. 📊 Show how to generate similarity maps
+
+## 🎨 Advanced: Similarity Maps
+
+For visual similarity analysis, use the existing advanced script:
+```bash
+cd apps/multimodal/vision-based-pdf-multi-vector/
+python multi-vector-leann-similarity-map.py
+```
+
+Edit the script to customize:
+- `QUERY`: Your question
+- `MODEL`: "colqwen2" or "colpali"
+- `USE_HF_DATASET`: Use HuggingFace dataset or local PDFs
+- `SIMILARITY_MAP`: Generate heatmaps
+- `ANSWER`: Enable Qwen-VL answer generation
+
+## 🔧 How It Works
+
+### ColQwen2 vs ColPali
+- **ColQwen2** (`vidore/colqwen2-v1.0`): Latest vision-language model
+- **ColPali** (`vidore/colpali-v1.2`): Proven multimodal retriever
+
+### Architecture
+1. **PDF → Images**: Convert PDF pages to images (150 DPI)
+2. **Vision Encoding**: Process images with ColQwen2/ColPali
+3. **Multi-Vector Index**: Build LEANN HNSW index with multiple embeddings per page
+4. **Query Processing**: Encode text queries with same model
+5. **Similarity Search**: Find most relevant pages/regions
+6. **Visual Maps**: Generate attention heatmaps (optional)
+
+### Device Support
+- **CUDA**: Best performance with GPU acceleration
+- **MPS**: Apple Silicon Mac support
+- **CPU**: Fallback for any system (slower)
+
+Auto-detection: CUDA > MPS > CPU
+
+## 📊 Performance Tips
+
+### For Best Performance:
+```bash
+# Use ColQwen2 for latest features
+--model colqwen2
+
+# Save page images for reuse
+--pages-dir ./cached_pages/
+
+# Adjust batch size based on GPU memory
+# (automatically handled)
+```
+
+### For Large Document Sets:
+- Process PDFs in batches
+- Use SSD storage for index files
+- Consider using CUDA if available
+
+## 🔗 Related Resources
+
+- **Fast-PLAID**: https://github.com/lightonai/fast-plaid
+- **Pylate**: https://github.com/lightonai/pylate
+- **ColBERT**: https://github.com/stanford-futuredata/ColBERT
+- **ColPali Paper**: Vision-Language Models for Document Retrieval
+- **Issue #119**: https://github.com/yichuan-w/LEANN/issues/119
+
+## 🐛 Troubleshooting
+
+### PDF Conversion Issues (macOS)
+```bash
+# Install poppler
+brew install poppler
+which pdfinfo && pdfinfo -v
+```
+
+### Memory Issues
+- Reduce batch size (automatically handled)
+- Use CPU instead of GPU: `export CUDA_VISIBLE_DEVICES=""`
+- Process fewer PDFs at once
+
+### Model Download Issues
+- Ensure internet connection for first run
+- Models are cached after first download
+- Use HuggingFace mirrors if needed
+
+### Import Errors
+```bash
+# Ensure all dependencies installed
+uv pip install colpali_engine pdf2image pillow matplotlib qwen_vl_utils einops seaborn
+
+# Check PyTorch installation
+python -c "import torch; print(torch.__version__)"
+```
+
+## 💡 Examples
+
+### Research Paper Analysis
+```bash
+# Index your research papers
+python -m apps.colqwen_rag build --pdfs ~/Papers/AI/ --index ai_papers
+
+# Ask research questions
+python -m apps.colqwen_rag search ai_papers "What are the limitations of transformer models?"
+python -m apps.colqwen_rag search ai_papers "How does BERT compare to GPT?"
+```
+
+### Document Q&A
+```bash
+# Index business documents
+python -m apps.colqwen_rag build --pdfs ~/Documents/Reports/ --index reports
+
+# Interactive analysis
+python -m apps.colqwen_rag ask reports --interactive
+```
+
+### Visual Analysis
+```bash
+# Generate similarity maps for specific queries
+cd apps/multimodal/vision-based-pdf-multi-vector/
+# Edit multi-vector-leann-similarity-map.py with your query
+python multi-vector-leann-similarity-map.py
+# Check ./figures/ for generated heatmaps
+```
+
+---
+
+**🎯 This integration makes ColQwen as easy to use as other LEANN features while maintaining the full power of multimodal document understanding!**

packages/leann-core/pyproject.toml

@@ -7,7 +7,7 @@ name = "leann-core"
 version = "0.3.5"
 description = "Core API and plugin system for LEANN"
 readme = "README.md"
-requires-python = ">=3.9"
+requires-python = ">=3.10"
 license = { text = "MIT" }
 
 # All required dependencies included

packages/leann-mcp/README.md

@@ -53,6 +53,11 @@ leann build my-project --docs $(git ls-files)
 # Start Claude Code
 claude
 ```
+**Performance tip**: For maximum speed when storage space is not a concern, add the `--no-recompute` flag to your build command. This materializes all tensors and stores them on disk, avoiding recomputation on subsequent builds:
+
+```bash
+leann build my-project --docs $(git ls-files) --no-recompute
+```
 
 ## 🚀 Advanced Usage Examples to build the index
 

packages/leann/pyproject.toml

@@ -7,7 +7,7 @@ name = "leann"
 version = "0.3.5"
 description = "LEANN - The smallest vector index in the world. RAG Everything with LEANN!"
 readme = "README.md"
-requires-python = ">=3.9"
+requires-python = ">=3.10"
 license = { text = "MIT" }
 authors = [
     { name = "LEANN Team" }
@@ -18,10 +18,10 @@ classifiers = [
     "Intended Audience :: Developers",
     "License :: OSI Approved :: MIT License",
     "Programming Language :: Python :: 3",
-    "Programming Language :: Python :: 3.9",
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
 ]
 
 # Default installation: core + hnsw + diskann

pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "leann-workspace"
 version = "0.1.0"
-requires-python = ">=3.9"
+requires-python = ">=3.10"
 
 dependencies = [
     "leann-core",