Knowledge‑Enhanced Large Model Service Framework (KAG): Integrating Knowledge Graphs with LLMs for Vertical Domain Applications

At the 2024 Inclusion·外滩大会 forum, Ant Group’s knowledge‑graph lead Liang Lei introduced the Knowledge‑Enhanced Large Model Service Framework (KAG), a system that uses graph‑logic symbols to guide decision‑making and retrieval, markedly improving precision and logical rigor in vertical domains.

KAG mitigates knowledge‑graph sparsity by leveraging information retrieval and the generative capabilities of large language models, lowering the barrier to constructing domain‑specific knowledge graphs.

The framework’s effectiveness is demonstrated in real‑world deployments such as Alipay’s AI‑native app “支小宝,” where government‑question answering accuracy reached 91% and medical Q&A accuracy exceeded 90%.

Ant Group plans to open KAG to the community, integrating it natively into the open‑source OpenSPG project and inviting collaborative development.

Key technical enhancements of KAG include:

01 LLM‑friendly Knowledge Representation: upgraded semantic representations that support multi‑modal, dynamic structures and schema‑constrained modeling.

02 Mutual Indexing: a graph‑based inverted index that captures entities, concepts, and text blocks for richer semantic retrieval.

03 Hybrid Reasoning: a mixed engine that combines symbolic decision‑making, vector retrieval, and LLM inference, generating executable logic forms (e.g., KGDSL, GQL).

04 Semantic Alignment: balances flexible information retrieval with strict professional decision‑making through concept‑level alignment and schema constraints.

05 KAG Model: defines collaborative tasks between LLMs and knowledge graphs, using instruction synthesis to let smaller models approach the performance of larger ones.

KAG’s architecture relies on Ant’s TuGraph‑DB as the underlying graph engine, providing efficient storage and Cypher‑based retrieval for knowledge‑graph‑augmented LLM responses.

In practice, KAG has been applied to vertical scenarios such as government services, medical health queries, and report generation, consistently reducing hallucination rates and improving factual accuracy.

Future work includes releasing a full technical report, adding user‑adjustable parameters for accuracy‑vs‑recall trade‑offs, and deepening collaboration with Zhejiang University’s OpenKG to build the OneGraph knowledge‑enhanced LLM platform.