16 New Types of Retrieval-Augmented Generation (RAG)

We’ve already done a post about 12 various types of Retrieval-Augmented Generation (RAG) before, but since then, many new enhanced RAG approaches have emerged. RAG techniques are always improving to achieve better accuracy of LLMs responses by retrieving relevant external documents or data during the response generation process. RAG systems are particularly useful for tasks requiring specific or up-to-date knowledge.

Here is a list of 16 latest types of RAG that can be used for different purposes:

1. HtmlRAG proposes using HTML directly to retain valuable information structure with headings or tables instead of plain text. To handle extra noise and size from HTML, HtmlRAG applies cleaning and pruning techniques. → Read more

2. FastRAG uses schema and script learning to process data efficiently without fully relying on AI models. It combines text search and knowledge graph querying, improving accuracy and cutting processing time by 90% and costs by 85%. → Read more

3. Auto-RAG introduces an autonomous approach, using LLMs to refine queries and plan retrievals through multi-turn dialogues until enough information is gathered. This fine-tuned system adapts to question difficulty and explains its process in natural language. → Read more

4. CORAG uses a Monte Carlo Tree Search (MCTS) framework to consider correlations between chunks and addresses the non-monotonic utility of adding more chunks. CORAG also adapts to different query types with a configuration agent. → Read more

5. MemoRAG introduces a dual-system approach with long-term memory: a lightweight LLM creates a "global memory" of the database and generates draft answers to guide retrieval tools, while a more powerful LLM generates the final answer using retrieved data. → Read more

6. RAG-Thief is designed to reveal privacy risks in RAG systems that integrate private knowledge bases with LLMs. It uses an automated approach to extract private data (over 70%) by starting with adversarial queries and learning from responses to generate more effective queries. → Read more

7. AssistRAG incorporates an intelligent assistant within LLMs to manage tools, memory, and plans. It uses a two-phase training process—Curriculum Assistant Learning and Reinforced Preference Optimization to enhance information retrieval and decision-making. → Read more

8. LaB-RAG generates image captions by combining image labels with RAG and LLMs. It converts X-ray images into text labels with simple classifiers. These labels guide pretrained LLMs to generate detailed radiology reports without training the LLM or using image data directly. → Read more

9. Video-RAG is a lightweight, cost-effective solution that enhances models' understanding of long videos without extensive fine-tuning. It uses a plug-and-play approach, employing open-source tools to extract aligned information from videos and integrates this data with LVLMs as auxiliary texts.→ Read more

10. Retrieval-Augmented Forecasting (RAF) enhances Time-Series Foundation Models (TSFMs), like Chronos, by dynamically retrieving relevant time-series examples to improve predictions. It addresses the event-driven and evolving nature of time-series data. → Read more

11. RuAG turns large datasets into interpretable logic rules with LLMs' commonsense to define predicates and Monte Carlo Tree Search (MCTS) to efficiently explore the data. The logic rules are then converted to natural language and integrated into LLM prompts, boosting their reasoning abilities. → Read more

12. MMed-RAG is a multimodal RAG system designed to improve the factual accuracy of medical vision-language models in tasks like diagnosis and report generation. It also uses adaptive context selection and preference fine-tuning to tackle issues like hallucination and misalignment. → Read more

13. Path-RAG is a framework that enhances cancer diagnosis by improving pathology image analysis for the PathVQA-Open tasks. It uses HistoCartography to extract knowledge from images, selecting key patches to incorporate expert insights, boosting accuracy from 38% to 47% and improving long-form question answering. → Read more

14. Multi-Reranker for financial tasks employs multiple reranker models to enhance retrieval accuracy. It tackles complex financial tasks like analyzing statements and disclosures using optimized query expansion, corpus refinement, and improved retrieval algorithms, securing 2nd place on the ACM-ICAIF '24 FinanceRAG competition. → Read more

15. G-RAG is designed for better information retrieval in fields like Material Science. It uses a graph database to map relationships between key entities (MatIDs) extracted from documents, which guide the retrieval of relevant information from external knowledge bases like Wikipedia. → Read more

16. RAGDiffusion is a RAG framework for generating accurate flat-lay clothing images from real-world data, addressing structural errors and hallucinations. It employs retrieval-based structure aggregation for spatial landmark guidance and omni-level garment generation for alignment in structure, patterns, and details. → Read more