From RAG to KAG, understanding the evolution of knowledge enhancement generation technology (Part 1)

Written by
Jasper Cole
Updated on:June-27th-2025
Recommendation

In-depth analysis of knowledge enhancement generation technology, insight into the evolution path of RAG and KAG.

Core content:
1. RAG technical limitations and KAG innovative architecture
2. Knowledge retrieval fuzziness and lack of logical reasoning ability
3. KAG application effect and future trend in professional field question answering

Yang Fangxian
Founder of 53A/Most Valuable Expert of Tencent Cloud (TVP)

Retrieval-enhanced generation  (RAG)  and knowledge-enhanced generation  (KAG)  represent two generations of technical paradigms that combine large models with external knowledge. As large models are increasingly used in professional fields, traditional RAG technology has exposed many limitations. KAG, as a new generation framework,  has significantly improved the accuracy, logic , and interpretability of professional question-answering  by deeply integrating knowledge graphs  with  large language models .

Today's article will systematically analyze the limitations of RAG, the innovative architecture of KAG, the comparison of their core technologies, actual application effects, and future development trends. We hope that it will help you fully understand the inherent logic and  practical value of this technological evolution .

Limitations of RAG Technology

Since its introduction, RAG technology has become  an important bridge connecting large language models with external knowledge, but it has gradually exposed many deficiencies in professional field applications  . Understanding these limitations is the premise for recognizing the value of KAG.

① The ambiguity of knowledge retrieval is the most fundamental problem of RAG.

RAG relies on vector similarity for retrieval, but semantic similarity does not mean knowledge relevance . In professional fields, many questions require precise matching of professional terms and concepts, and vector retrieval often returns content that is semantically similar but knowledge-irrelevant .

For example, in the legal field, when searching for "sentencing standards for negligent homicide", content about "intentional homicide" may be retrieved because the two are close in position in the vector space. This "approximate rather than precise" retrieval mechanism leads to a significant decrease in the accuracy of answers in professional fields.

② Weak logical reasoning ability is another major flaw.

RAG is essentially still based on  the next-token prediction mechanism of a large language model , and performs poorly on problems that require numerical calculations, time series analysis, or complex logical reasoning. Tests show that traditional RAG can only solve about 16% of the problems in the enterprise knowledge base, especially complex problems that require multi-step reasoning.

For example, in the medical field, for problems such as "judging the progression of a disease based on the changing trends of indicators in three consecutive examinations of the patient", the RAG system often finds it difficult to give an accurate analysis that conforms to medical logic.

③ The cascading errors of multi-hop question answering are particularly obvious in complex queries.

When a question requires connecting multiple knowledge points, any retrieval error in the process will cause the final answer to deviate from the correct direction. However, RAG lacks a logical verification mechanism for each retrieval result, and errors will continue to accumulate .

For example, for "the mechanism by which a drug interaction leads to increased side effects", RAG needs to first search for the characteristics of drug A, then search for the characteristics of drug B, and finally analyze the interaction.

④ Knowledge updating and consistency issues also restrict the performance of RAG.

Knowledge in professional fields is frequently updated and requires strict consistency , such as the revision of clinical guidelines or the update of legal provisions. It is difficult for the RAG system to ensure that all relevant text blocks are updated synchronously, which may result in contradictory information on the same issue in different document blocks  .

⑤ Poor domain adaptability and insufficient handling of domain-specific logic.

Professional knowledge such as compound interest calculation in the financial field and rules for citing articles in the legal field  are difficult to accurately grasp with general language models alone. Ant Group found in the e-government Q&A that traditional RAG’s answers to questions such as government service processes and material requirements were not professional enough.

Table: Comparison of RAG's performance on different types of questions

Question TypeExampleRAG ApplicabilityMain Challenges
Explicit fact query"Where will the 2016 Olympics be held?"excellentRetrieval accuracy optimization
Implicit fact query"Stores with the highest revenue growth rate in the past month"generallyMulti-step search and calculation
Explainable Reasoning"Why did the revenue in South China drop by 5%?"PoorBusiness rules integration
Implicit Reasoning"Root Cause Analysis of IT System Failures"Almost ineffectiveLogic mining and induction

These limitations have prompted the industry to explore more powerful  knowledge enhancement solutions . KAG  came into being in this context, and  systematically solved the above problems through the two-way enhancement of knowledge graphs and large models.