Topic

Practices and Challenges of Landing a RAG Application in Tongyi Baijian

With the rapid development of big model applications, RAG as a plug-in brain implements injecting user's knowledge base into the process of big model usage, which plays an important complementary role to the big model capability. However, when applied in practice, complex knowledge formats and multimodal knowledge contents usually pose great challenges to the RAG system. In addition, with the rise of long text big models, the traditional RAG system based on chunking strategy has gradually failed to meet the new scene paradigm. Finally, when landing on the ground, how to evaluate the reliability and accuracy of the RAG system for the customer's scenario and give targeted solutions is also a key issue that affects the RAG technology to play a practical role. This presentation will introduce the main problems encountered by Tongyi Bailian RAG application in the past year's landing practice, and share our solutions with you. Outline: Overview of RAG technology  Requirements and background  Technology development in the past year  Practice Challenges Complex Graphic Solutions  Complexity of Knowledge Bases  Offline parsing in conjunction with multimodal LLMs  Technical Solution Introduction  long text RAG Long text agent  effect demonstration Future direction RAG Evaluation problem introduction Industry programme CoFE-RAG Assessment Framework  Effectiveness Demonstration