Tencent Large Language Model Applications: RAG, GraphRAG, and Agent Technologies

Tencent's large language models are applied in many business contexts such as content generation, intelligent customer service, document assistance, code copilot, and role‑playing NPC interactions, aiming to boost automation and user experience across the WeChat ecosystem, video platforms, office tools, and games.

Retrieval‑Augmented Generation (RAG) combines external knowledge bases with generative models to improve answer accuracy, reduce hallucinations, and keep knowledge up‑to‑date; the article outlines its data preparation, knowledge recall, and generation enhancement steps, as well as challenges like document format diversity and relevance filtering.

GraphRAG extends RAG by integrating knowledge graphs, enabling both local entity‑level retrieval and global community‑level reasoning; the workflow includes knowledge extraction, graph construction, indexed retrieval, and generation that provides traceable, context‑rich answers for complex, long‑form texts.

Agent technology equips the LLM with planning, reasoning, and tool‑calling capabilities, allowing multi‑step task execution such as travel planning, external API calls, and dynamic replanning; the system defines roles (User, Planner, Tool) and iteratively refines actions based on feedback.

Model improvement methods cover Supervised Fine‑Tuning (SFT) on domain‑specific data, prompt engineering, few‑shot examples, and evaluation metrics like correctness and relevance, ensuring the models meet specific business requirements.

The article concludes with a Q&A session addressing embedding strategies for QA pairs, Chinese semantic splitting, evaluation of answers, when to use SFT versus RAG, and the impact of fine‑tuning on model generality and reasoning ability.