Skip to content 
flowchart TD
    subgraph "1. Document Processing"
        A[Documents] --> B[Split Text into Chunks]
        B --> C1[Chunk-1]
        B --> C2[Chunk-2]
        B --> C3[Chunk-n]
    end

    subgraph "2. Dual Embedding"
        C1 & C2 & C3 --> DEM{{Dense Embedding Model}}
        C1 & C2 & C3 --> SEM{{Sparse Embedding Model}}
        DEM --> DE1[Dense Embedding-1] & DE2[Dense Embedding-2] & DE3[Dense Embedding-n]
        SEM --> SE1[Sparse Embedding-1] & SE2[Sparse Embedding-2] & SE3[Sparse Embedding-n]
    end

    subgraph "3. Unified Indexing"
        DE1 & SE1 --> UI1[Unified Index Entry-1]
        DE2 & SE2 --> UI2[Unified Index Entry-2]
        DE3 & SE3 --> UI3[Unified Index Entry-n]
        UI1 & UI2 & UI3 --> UDB[(Unified VectorDB)]
    end

    subgraph "4. Query Processing"
        Q[Query] --> QDEM{{Dense Embedding Model}}
        Q --> QSEM{{Sparse Embedding Model}}
        QDEM --> QDE[Query Dense Embedding]
        QSEM --> QSE[Query Sparse Embedding]
    end

    subgraph "5. Two-Step Hybrid Retrieval"
        QSE --> SS{Sparse Search}
        SS --> UDB
        UDB -->|Top K1 Sparse Results| SR[Sparse Retrieved Chunks]
        SR --> DS{Dense Search}
        QDE --> DS
        DS -->|Top K2 Dense Results| DR[Final Retrieved Chunks]
    end

    subgraph "6. Context Formation"
        DR --> CF[Query + Retrieved Chunks]
    end

    subgraph "7. Generation"
        CF --> LLM[LLM]
        LLM --> R[Response]
    end

    Q --> CF

Sentence Window Retriever-Based RAG Approach

Introduction

The Sentence Window Retriever-Based RAG (Retrieval-Augmented Generation) approach is an advanced implementation of the RAG framework, designed to enhance the context-awareness and coherence of AI-generated responses. This method combines the power of large language models with efficient information retrieval techniques, providing a robust solution for generating high-quality, context-rich responses.

Motivation

Traditional RAG systems often struggle with maintaining coherence across larger contexts or when dealing with information that spans multiple chunks of text. The Sentence Window Retriever-Based approach addresses this limitation by preserving the contextual relationships between chunks during the indexing process and leveraging this information during retrieval and generation.

Method Details

Document Preprocessing and Vector Store Creation

  1. Document Splitting: The input document is split into sentences.
  2. Chunk Creation: Sentences are grouped into manageable chunks.
  3. Embedding: Each chunk is processed through an embedding model to create vector representations.
  4. Vector Database Indexing: Chunk IDs, text, and embeddings are stored in a vector database for efficient similarity search.
  5. Document Structure Indexing: A separate database stores the relationships between chunks, including references to previous and next k chunks for each chunk.

Retrieval-Augmented Generation Workflow

  1. Query Processing: The user query is embedded using the same embedding model used for chunks.
  2. Similarity Search: The query embedding is used to find the most relevant chunks in the vector database.
  3. Context Expansion: For each retrieved chunk, the system fetches the previous and next k chunks using the document structure database.
  4. Context Formation: The retrieved chunks and their expanded context are combined with the original query.
  5. Generation: The expanded context and query are passed to a large language model to generate a response.

Flow Chart

The following flow chart illustrates the Sentence Window Retriever-Based RAG approach:

flowchart TD
    subgraph "1. Document Processing"
        A[Document] --> B[Split into Sentences]
        B --> C[Group Sentences into Chunks]
    end

    subgraph "2. Chunk Processing"
        C --> D1[Chunk 1]
        C --> D2[Chunk 2]
        C --> D3[Chunk 3]
        C --> D4[...]
        C --> Dn[Chunk n]
    end

    subgraph "3. Embedding"
        D1 & D2 & D3 & D4 & Dn --> E{Embedding Model}
        E --> F1[Embedding 1]
        E --> F2[Embedding 2]
        E --> F3[Embedding 3]
        E --> F4[...]
        E --> Fn[Embedding n]
    end

    subgraph "4. Indexing"
        F1 & F2 & F3 & F4 & Fn --> G[(Vector Database)]
        G -->|Store| H[Chunk ID]
        G -->|Store| I[Chunk Text]
        G -->|Store| J[Embedding]
    end

    subgraph "5. Document Structure Store"
        C --> K[(Document Structure DB)]
        K -->|Store| L[Chunk ID]
        K -->|Store| M[Previous k Chunk IDs]
        K -->|Store| N[Next k Chunk IDs]
    end

    subgraph "6. Retrieval"
        O[Query] --> P{Embedding Model}
        P --> Q[Query Embedding]
        Q --> R{Similarity Search}
        R --> G
        G --> S[Retrieved Chunks]
    end

    subgraph "7. Context Expansion"
        S --> T{Expand Context}
        T --> K
        K --> U[Previous k Chunks]
        K --> V[Next k Chunks]
        S & U & V --> W[Expanded Context]
    end

    subgraph "8. Generation"
        W --> X[Query + Expanded Context]
        X --> Y[LLM]
        Y --> Z[Response]
    end

Key Features of RAG

  • Efficient Retrieval: Utilizes vector similarity search for fast and accurate information retrieval.
  • Context Preservation: Maintains document structure and chunk relationships during indexing.
  • Flexible Context Window: Allows for adjustable context expansion at retrieval time.
  • Scalability: Capable of handling large document collections and diverse query types.

Benefits of this Approach

  1. Improved Coherence: By including surrounding chunks, the system can generate more coherent and contextually accurate responses.
  2. Reduced Hallucination: Access to expanded context helps the model ground its responses in retrieved information, reducing the likelihood of generating false or irrelevant content.
  3. Efficient Storage: Only stores necessary information in the vector database, optimizing storage usage.
  4. Adaptable Context: The size of the context window can be adjusted based on the specific needs of different queries or applications.
  5. Preservation of Document Structure: Maintains the original structure and flow of the document, allowing for more nuanced understanding and generation.

Conclusion

The Sentence Window Retriever-Based RAG approach offers a powerful solution for enhancing the quality and contextual relevance of AI-generated responses. By preserving document structure and allowing for flexible context expansion, this method addresses key limitations of traditional RAG systems. It provides a robust framework for building advanced question-answering, document analysis, and content generation applications.