智通财经网·2025-12-31 01:39Group 1 - The core viewpoint is that the RAG (Retrieval-Augmented Generation) architecture provides long-term memory for large models, driven by enterprise and personalized needs, leading to increased demand for RAG storage [1] - RAG architecture allows LLMs (Large Language Models) to query vector databases before generating responses, enhancing the accuracy and timeliness of generated results [1] - RAG is penetrating both online scenarios (e-commerce, web search) and offline scenarios (enterprise, legal, engineering research), while personalized RAG retains user long-term memory and preferences, significantly boosting demand [1] Group 2 - The transition of vector database storage media from "in-memory retrieval" to "full SSD storage architecture" is driving the demand for high-bandwidth and large-capacity SSDs [2] - For a scale of 10 billion vectors, the required SSD capacity is 11.2TB, with 1.28TB for PQ vectors and 10TB for indexing [2] - The AiSAQ system offers a cost advantage of 4-7 times compared to DiskANN media, enhancing scalability and economic feasibility of RAG systems [2] Group 3 - The TOS (Tianyan) engine introduces a new paradigm for vector storage with the TOS Vector Bucket, which utilizes a self-developed Cloud-Native vector indexing library and a multi-level local caching architecture [3] - This architecture meets both high/low-frequency data storage needs and significantly lowers the barrier for enterprises to utilize large-scale vector data [3] - TOS Vectors can store massive semantic vectors while ensuring sustainable accumulation of long-term data, facilitating complex task processing [3]