Agentic RAG

Theoretical review

Published by Meta AI in NeurIPS 2020 -> Paper.

1. Overview:

Hybrid Model: Combines the strengths of retrieval-based and generation-based models.
Purpose: Enhances LLM text generation by incorporating relevant external information, improving accuracy and context.

2. Structure:

Retriever:
- Function: Searches a large corpus (e.g., documents, articles) to find relevant information based on the input query.
- Techniques: Uses methods like BM25, dense retrieval, or neural retrievers.
Generator:
- Function: Generates a response using the information retrieved.
- Model: Typically a transformer-based model like GPT-4 or Llama.

3. Workflow:

Input Query: User provides a query or prompt.
Document Retrieval:
- The retriever fetches a set of relevant documents or passages.
- These documents provide context and factual information.
Response Generation:
- The generator uses the retrieved documents to produce a coherent and contextually accurate response.
- Ensures the generated text is informed by the most relevant information available.

4. Benefits:

Enhanced Accuracy: By grounding responses in real-world data, RAG models significantly improve the accuracy of generated content.
Reduced Hallucinations: The integration of external knowledge helps mitigate the risk of generating incorrect or nonsensical responses.
Scalability: RAG systems can handle vast amounts of data, making them suitable for enterprise-level applications.

1. Overview:

Enhanced RAG: Extends RAG by adding agent-like capabilities.
Purpose: Designed to perform tasks autonomously, interacting with various tools and APIs to achieve specific goals.

2. Structure:

Retriever:
- Function: Similar to RAG, it fetches relevant documents based on the input query.
Generator:
- Function: Generates an initial response using the retrieved documents.
Agent Module:
- Function: Evaluates the generated response, cross-references it with the knowledge base, and makes corrections if discrepancies are found.

3. Workflow:

Input Query: User provides a query/question/task.
Document Retrieval: The retriever fetches relevant documents to provide context.
Initial Response Generation: The generator creates a preliminary answer using the retrieved information.
Response Verification: The agent system assesses the initial response against the knowledge base to ensure accuracy.
Response Correction (if needed): If inaccuracies are detected, the agent system refines the response to align with verified information.

4. Benefits:

Improved Reliability: The agent system's verification process ensures responses are accurate and trustworthy.
Dynamic Correction: Enables real-time adjustments to responses, enhancing the system's adaptability to new information.
User Trust: By providing verified answers, the system builds greater user confidence in its outputs.