Retrieval-Augmented Generation for Large Language Models: A Survey

Abstract

Large Language Models (LLMs) :

• Encounter challenges such as :
  • Hallucination
  • Outdated knowledge
  • Non-transparent, untraceable reasoning processes.

 

Solution :

Retrieval-Augmented Generation (RAG) : Incorporating knowledge from external databases.

 

 

 

 

 

II. OVERVIEW OF RAG

A typical application of RAG

Typical Application of RAG

• A common scenario is illustrated where RAG is used to answer a user's query by retrieving relevant document chunks from external databases.
• This approach bridges the gap when LLMs lack the capacity to provide up-to-date or domain-specific information.

 

Components of RAG

• Indexing:
  • Cleaning and extraction
  • raw data (in diverse formats like PDF, HTML, Word, and Markdown) → uniform plain text → chunks → (encode using an embedding model) →  vector representations
  • This step is crucial for enabling efficient similarity searches in the subsequent retrieval phase
• Retrieval:
  • Transforms user queries into vector representations.
  • Computes similarity scores between query vectors and document vectors to retrieve relevant chunks.
• Generation:
  • Synthesizes retrieved documents and user queries into coherent prompts.
  • LLMs use these prompts to generate responses, drawing on both intrinsic and retrieved knowledge.

 

Naive RAG

• Introduction
  • Represents the earliest methodology.
  • Follows a traditional process of indexing, retrieval, and generation.
  • Characterized as a "Retrieve-Read" framework.
• Challenges and Limitations
  • Retrieval Challenges : Faces issues with precision and recall during retrieval.
  • Generation Difficulties : Generation phase can suffer from hallucinations and irrelevance.( irrelevance, toxicity, or bias)
  • Augmentation hurdles :
    • Integrating retrieved information with the different task → Include disjointed or incoherent outputs.
    • Similar information is retrieved from multiple sources → redundant outputs and difficulty integrating retrieved information.
    • various passages and ensuring stylistic and tonal consistency → complexity
    • Facing complex issues, a single retrieval based on the original query may not suffice to acquire adequate context information

Advanced RAG

• Introduction
  • Introduces improvements to overcome the limitations of Naive RAG.
  • Focuses on enhancing retrieval quality using pre-retrieval and post-retrieval strategies.
  • Optimizes indexing and retrieval processes for better performance.

Modular RAG

• Introduction
  • Advances beyond Naive and Advanced RAG, offering greater adaptability and versatility.
  • Incorporates diverse strategies for improving retrieval and processing.
  • Supports sequential processing and integrated end-to-end training across components.

 

Challenges and Limitations

• Naive RAG:
  •  
• Advanced and Modular RAG:
  • Address these limitations with refined indexing techniques, enhanced retrieval methods, and flexible module configurations.

Comparison of RAG Paradigms

• Naive RAG:
  • Simple, sequential process with basic indexing, retrieval, and generation.
• Advanced RAG:
  • Incorporates optimization strategies around pre-retrieval and post-retrieval processes.
  • Still follows a chain-like structure.
• Modular RAG:
  • Greater flexibility with multiple specific functional modules.
  • Supports iterative and adaptive retrieval methods, not limited to sequential processing.

Illustrative Examples

• Figures in the paper illustrate the technology tree of RAG research and comparisons between the three paradigms of RAG.
• Examples show how RAG can be applied to enhance the accuracy and relevance of LLM-generated answers.