This repository serves as a code and resource compilation for the book "Practical RAG for Large Models." It targets developers and researchers interested in advancing Retrieval-Augmented Generation (RAG) systems beyond basic implementations, offering practical techniques and code examples for building sophisticated RAG pipelines.

How It Works

The project categorizes RAG advancements into three stages: Basic, Advanced, and Super. Basic RAG involves simple pipelines with fixed-length chunking and retrieval. Advanced RAG focuses on optimizing retrieval and generation through techniques like query enhancement (rewriting, step-back, sub-questions), chunk enhancement (compression, guessing questions), retrieval enhancement (mixed search, metadata filtering), and generation enhancement (traceability). Super RAG explores Agentic RAG, multimodal RAG, structured RAG, and GraphRAG, integrating LLMs as decision-making engines and leveraging graph structures for richer context.

Quick Start & Requirements

• Installation: Primarily involves cloning the repository and setting up Python environments. Specific commands for running individual components (e.g., training embedding models, finetuning LLMs) are detailed within respective directories.
• Prerequisites: Python 3.x, and potentially specific libraries like Langchain. GPU and CUDA are likely required for model training and finetuning. Some sections may require API keys for closed-source models.
• Resources: Training embedding models and finetuning LLMs can be resource-intensive, requiring significant GPU memory and time.
• Links: The repository structure implies code examples for various RAG techniques, with specific chapters dedicated to Langchain demos and data preparation.

Highlighted Details

• Explores advanced RAG techniques including query and chunk enhancement, retrieval strategies (mixed search, metadata filtering), and generation improvements (traceability).
• Covers Agentic RAG, where LLMs act as engines to plan and execute tasks using retrieval as tools.
• Includes sections on GraphRAG, modifying existing implementations for Chinese prompts and private models, and multimodal RAG.
• Details different model training approaches: Independent Training, Sequential Training (LLM First/Retriever First), and Joint Training for RAG systems.

Maintenance & Community

• The project is associated with the book "大模型RAG实战" (Practical RAG for Large Models), suggesting a structured development and potential for community engagement around the book's content. Specific community links (Discord, Slack) are not explicitly mentioned in the provided README excerpt.

Licensing & Compatibility

• The licensing information is not detailed in the provided README excerpt. Compatibility for commercial use would depend on the specific licenses of the underlying models and libraries used.

Limitations & Caveats

• Some sections are marked as "待更新~" (to be updated), indicating that certain advanced RAG techniques (e.g., retrieval enhancement, generation enhancement) are still under development or documentation. The project's focus on specific Chinese models might limit direct applicability for users relying solely on English-centric ecosystems.