📄 PDF Search with Late Interaction

A fully local Retrieval-Augmented Generation (RAG) pipeline for PDF search and question answering.
It uses Late Interaction with ColPali / ColQwen / ColSmol family of models for encoding PDF page images, stores embeddings in Qdrant, and answers queries using a VLM with retrieved context.

This project includes an interactive Streamlit web app to index PDFs and ask questions.

---

🚀 Features

• Multi-Vector embedding of PDF pages.
• Vector search powered by Qdrant with multi-vector support
• Batch PDF indexing (folder or individual files).
• Question answering using retrieved context images.
• Local or remote LLMs via LiteLLM API compatibility.
• Streamlit UI for user-friendly PDF upload and querying.

---

🔍 Hybrid Search with Prefetch + Reranking

Uses a two-stage retrieval pipeline for faster queries without losing accuracy:

1. Prefetch (Fast Retrieval) – Query pooled embeddings to quickly fetch top-K candidates from Qdrant.

2. Rerank (High Accuracy) – Pass candidates to ColPali, ColQwen, or ColSmol models for late-interaction reranking.

Pros:

• Faster response times.
• Higher relevance with model-based reranking.

Cons:

• Requires extra storage for pooled embeddings.

📊 Visual Representation of the Pipeline

Sequence Diagram

sequenceDiagram
    %% Define participants in lanes
    participant User
    participant Indexing as Indexing Pipeline
    participant Embedder as Embedder
    participant VectorDB as Vector Database
    participant Retriever as Retriever
    participant Reranker as Reranker
    participant LLM as LLM

    par Offline Indexing
        User->>Indexing: Upload PDF
        Indexing->>Indexing: Extract Images
        Indexing->>Embedder: Create embeddings
        Embedder->>Indexing: Return embeddings
        Indexing->>Indexing: Pool embeddings for each image
        Indexing->>VectorDB: Store pooled & original embeddings
    and Online Retrieval + RAG
        User->>Retriever: Submit Query
        Retriever->>Embedder: Embed Query (ColPali / ColQwen / ColSmol)
        Embedder->>Retriever: Return query embeddings
        Retriever->>VectorDB: Prefetch Top-K results
        VectorDB->>Retriever: Return candidate results
        Retriever->>Reranker: Send candidates for reranking
        Reranker->>Retriever: Return ranked results
        Retriever->>LLM: Pass reranked context
        LLM->>User: Return final answer
    end

Loading

Flowchart

flowchart TD
    %% INDEXING STAGE
    subgraph Indexing[📥 Indexing Pipeline]
        direction LR
        A[Upload PDF / Images] --> B[Extract images & text]
        B --> C[Generate embeddings<br/>ColPali / ColQwen / ColSmol]
        C --> D[Mean-pool embeddings per image]
        D --> E[Store pooled & original embeddings in Qdrant]
    end

    %% RETRIEVAL + RAG STAGE
    subgraph Retrieval[🔍 Retrieval + RAG]
        direction LR
        F[User Query] --> G[Embed query]
        G --> H[Prefetch Top-K from Qdrant]
        H --> I[Rerank results<br/>ColPali / ColQwen / ColSmol]
        I --> J[Pass ranked results to LLM]
        J --> K[Generate contextual answer]
    end

    %% FLOW BETWEEN STAGES
    E --> H
    K --> L[Final Answer to User]

Loading

---

🛠 Requirements

• A supported GPU (CUDA or MPS) is recommended for faster performance.
• Installed system dependencies for:
  • poppler (required by pdf2image)

---

📦 Installation

1. Clone the repository

   git clone https://github.com/logxdx/contextualized-late-interation-with-pdfs.git
   cd contextualized-late-interation-with-pdfs

2. Create and activate a virtual environment

   python -m venv .venv
   source .venv/bin/activate   # On Windows: .venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

4. Install Poppler

   • Linux (Debian/Ubuntu):

     sudo apt-get install poppler-utils

   • macOS:

     brew install poppler

   • Windows:

     • Download from: https://github.com/oschwartz10612/poppler-windows/releases/
     • Add bin folder to your PATH.

5. Configure environment variables Create a .env file in the project root:

   # API key & URL for your LLM provider (ollama by default)
   API_KEY=your_api_key_here
   BASE_URL=http://localhost:11434/v1
   
   # Optional Hugging Face token if model access is gated
   HF_TOKEN=your_hf_token_here

---

📂 Project Structure

contextualized-late-interation-with-pdfs/
│
├── rag.py              # Core RAG backend logic
├── app.py              # Streamlit UI
├── requirements.txt    # Python dependencies
├── .env.example        # Example environment file
└── README.md           # Project documentation

---

⚙️ Usage

1️⃣ Command-line Example

You can use the backend without the UI.

python rag.py

Inside __main__, update:

rag = RAG("vidore/colpali-v1.3")
rag.index_file(pdf_path="attention_is_all_you_need.pdf", batch_size=1)
rag.answer(
    query="How does multi-headed attention work?",
    top_k=4,
    prefetch_limit=10
)
rag.close()

---

2️⃣ Running the Streamlit App

Start the web interface:

streamlit run app.py

Main Tabs:

• 📥 Index PDFs Upload PDF files or specify a folder to index.
• ❓ Ask Questions Enter your query and get an answer with retrieved page previews.

---

🧩 How It Works

1. PDF Processing

   • PDFs are converted to images (pdf2image).
   • Images are encoded into embeddings using ColPali / ColQwen / ColSmol models.

2. Vector Storage

   • Embeddings are stored in Qdrant with three vector fields:

     • original
     • mean_pooling_rows
     • mean_pooling_columns

3. Retrieval

   • For a query, text is embedded.
   • Qdrant searches across all three vector types.
   • Top matches (pages) are retrieved.

4. Answer Generation

   • Retrieved pages are sent as image+text context to an LLM.
   • The LLM answers based only on the given images.

---

⚡ Performance Notes

• Use GPU for significant speedup.
• Increase batch_size when indexing for faster throughput.
• Reduce dpi in _pdf_to_image for lower memory use.

---

🛡️ Troubleshooting

• CUDA out of memory: Reduce batch_size or use torch.float16.
• Poppler not found: Install via your OS package manager.
• Model access denied: Accept the model’s terms on Hugging Face and set HF_TOKEN.

---

📜 License

This project is released under the MIT License.

---

🙌 Acknowledgments

• ColPali Engine
• Qdrant
• LiteLLM
• Streamlit
• pdf2image