Bridging the Gap Between Large Models and Real‑World Applications with RAG and Agents

Large language models still face a distance between technology and practical use; Retrieval‑Augmented Generation (RAG) and Agent techniques are proposed to bridge this gap, addressing the "last mile" problem and offering new solutions for multi‑Agent collaboration, intelligent data governance, financial risk control, and operations automation.

During the preparation of the 2025 DA Digital Technology Conference (Shanghai), industry experts were interviewed to provide technical and practical insights. The article compiles these interviews, discussing the practice and application of these technologies, analyzing their advantages and challenges in various scenarios, and looking ahead to future development directions.

Core Challenges and Key Technologies for Large‑Model Deployment – RAG enhances retrieval capabilities, compensating for the limitations of large models in directly solving application problems, while Agents simulate human behavior to complete specific tasks from input to output. In operations, Agents can generate actionable commands from alerts, and multi‑Agent collaboration further improves task accuracy and reliability.

Large‑Model Practice in Operations – Operations of large enterprises involve maintaining tens of thousands of databases or data nodes, where traditional manual methods are error‑prone and inefficient. Agents enable automated pre‑ and post‑checks and rollback operations, reducing human intervention and operational risk. Multi‑Agent collaboration generates answers from different perspectives, with a master Agent selecting the consensus answer, mitigating hallucinations and enhancing reliability.

Large‑Model Applications in Finance – In finance, large models support intelligent customer service to improve Q&A efficiency, while in intelligent advisory they must comply with licensing and professional standards. Models assist risk identification through probabilistic analysis, alleviating talent shortages and lowering operational costs, though final decisions still require human oversight.

Intelligent Data Governance – Data governance addresses security, compliance, and cross‑scenario data usage. RAG excels at unstructured data retrieval but needs improvement for structured data. Automated data validation and verification during migration can significantly boost efficiency and reduce labor costs, while ethical and security considerations remain critical.

Database Localization – As a strategic demand, database localization benefits from large‑model automation in data migration and verification, enhancing safety and reducing manual effort.

Edge Models and Inference Acceleration – Edge models run locally on devices, solving latency and resource constraints, while cloud‑side inference acceleration lowers cost and improves response speed. Combining edge and cloud inference expands large‑model applicability across scenarios.

Exploration of Intelligent Applications – Large models automate office tasks such as generating weekly reports from emails, and in consumer products they can provide travel‑related recommendations based on weather and user preferences, illustrating the growing potential of AI‑driven solutions.

Overall, despite remaining challenges, the integration of RAG, Agents, and multi‑Agent collaboration demonstrates early success in operations, finance, and data governance, and promises broader impact as technologies mature.