KAG: Boosting LLMs in Professional Domains via Knowledge Augmented Generation

1 Introduction

• RAG’s similarity-based retrieval can return incomplete or repeated material, while specialized questions may require relational, numerical, or logical reasoning.The introduction also notes that language models can be weak at numerical and temporal reasoning.Source: §1 · p. 2
• KAG combines knowledge graphs and RAG, using semantic types, relationships, and KGQA logical forms in retrieval and generation.The framework targets professional-domain knowledge services.Source: §1 · p. 2
• LLMFriSPG supports schema-free information extraction and schema-constrained expert knowledge on shared entity and event types, with mutual indexing between graph structures and text chunks.This representation supports graph-structure-based inverted indexing and unified logical-form representation, reasoning, and retrieval.Source: §1 · p. 2
• The hybrid engine combines planning, reasoning, and retrieval operators to turn questions into processes using language and symbols.Its operators integrate text and exact-match retrieval, KG reasoning, language reasoning, and numerical computation.Source: §1 · p. 2
• KAG enhances general LLMs’ understanding, inference, and generation abilities for functions including retrieval, question understanding, semantic reasoning, and summarization.The paper also proposes semantic-reasoning-based knowledge alignment for relations such as synonyms, hypernyms, and inclusions.Source: §1 · pp. 2–3
• F1 scores rose by 19.6% on 2WikiMultiHopQA, 12.2% on MuSiQue, and 12.5% on HotpotQA versus HippoRAG; retrieval metrics also improved.The evaluation covered end-to-end Q&A and retrieval effectiveness.Source: §1 · p. 3
• KAG achieved significantly higher accuracy than traditional RAG in Ant Group’s E-Government and E-Health Q&A applications.The scenarios cover administrative-process questions and inquiries about diseases, symptoms, and treatments using provided resources.Source: §1 · p. 3

2 Approach

• KAG-Builder constructs offline indexes, while KAG-Solver integrates LLM, knowledge, and mathematical reasoning; KAG-Model optimizes capabilities used across the modules.Semantic-reasoning-based knowledge alignment is used in both Builder and Solver.Source: §2 · p. 3

2.1 LLM Friendly Knowledge Representation

• LLMFriSPG upgrades SPG through deep text-context awareness, dynamic properties, and knowledge stratification to make the representation more suitable for LLMs.It uses DIKW’s data, information, and knowledge hierarchy as a reference.Source: §2.1 · p. 3
• LLMFriSPG defines types, concepts, instance-to-concept relations, and executable rules as parts of its representation.Its type inventory includes entity and event types and predefined properties; concept types organize concepts and concept relations.Source: §2.1 · p. 3
• Type properties include expert-predefined, dynamically added, and system-built-in properties; supporting_chunks links an instance to the text chunks containing it.Static properties mainly serve professional decision-making, while dynamic properties mainly serve general information retrieval, allowing users to balance ease of use and professionalism.Source: §2.1 · pp. 4–5
• Concept nodes group instances that share conceptual terminology, enabling semantic alignment and providing navigation for retrieval.Figure 2 distinguishes instances from concepts in the representation.Source: Figure 2 · p. 4, §2.1 · p. 5
• The representation separates schema-constrained KGcs knowledge, extracted KGfr graph information, and semantically segmented raw chunks (RC), which supplement one another through graph-based indexing.KGcs offers high accuracy and logical rigor but requires more manual annotation; greater KGfr coverage can favor retrieval efficiency and lower expert effort when some information loss or error is acceptable.Source: §2.1 · p. 5
• KAG-Builder extracts phrases and triples, performs disambiguation and fusion through semantic alignment, then writes the constructed graph to storage.Figure 4 depicts this sequence for domain unstructured documents.Source: Figure 4 · p. 6

2.2 Mutual Indexing

• KAG-Builder’s three processes—structured information acquisition, knowledge semantic alignment, and graph storage—build mutual indexing between graph structures and text chunks.The indexing adds descriptive context to the graph and uses concept semantics to align knowledge granularities, reduce noise, and increase graph connectivity.Source: §2.2 · p. 6

2.2.1 Semantic Chunking

• Semantic chunking follows document structure and paragraph connections while keeping chunks within LLM context-window limits and thematically cohesive.Chunk identifiers encode article, paragraph, and within-paragraph order, while reciprocal links connect chunks to source documents.Source: §2.2.1 · p. 6

2.2.2 Information Extraction with More Descriptive Context

• Extraction builds graph instances from entities, events, and relations, adds hypernym links, and stores descriptive properties to support later knowledge alignment.LLMs generate properties including descriptions, summaries, semantic types, and SPG classes for each instance.Source: §2.2.2 · p. 6

2.2.3 Domain Knowledge Injection And Constraints

• Domain knowledge injection adds domain concepts and terms to extraction prompts, while schema constraints and document-type structures organize professional information.Government-affairs documents can be modeled with properties such as required materials, allowing the corresponding chunk to be retrieved directly.Source: §2.2.3 · p. 7

2.2.4 Mutual indexing between text chunk vectors and knowledge structures

• Mutual indexing links graph instances, text-chunk nodes, and a concept graph, with KG and vector stores holding graph data and text vectors.The concept graph links instances to fine-grained semantic types; graph instances can come from schema-free or schema-constrained extraction.Source: §2.2.4 · pp. 7–8

2.3 Logical Form Solver

• The solver decomposes a question into logical-form subquestions, performs hybrid reasoning, stores results in global memory, and iterates with supplementary questions when needed.Generation produces an answer from the accumulated memory; the example figure describes this loop from decomposition through a satisfaction check.Source: §2.3 · p. 8, Figure 6 · p. 9

2.3.1 Logical Form Planning

• Logical functions encode execution actions, and planning combinations of them decomposes complex problems into solvable expressions.Retrieval, sorting, mathematical calculation, and deduction are among the described operations.Source: §2.3.1 · pp. 9–10
• For multi-hop retrieval, later expressions reuse consistent variable names, while sorting can select results by direction and limit.A retrieved variable can also be used by later calculations or deductions.Source: §2.3.1 · p. 10

2.3.2 Logical Form for Reasoning

• Logical forms make distinctions such as “and” versus “or” explicit, and represent complex questions as linked actions for retrieval, numerical reasoning, and semantic deduction.GraphRetrieval applies the logical-form clause to obtain structured results, while HybridRetrieval combines the subproblem’s natural-language form with graph retrieval.Source: §2.3.2 · p. 10
• The reasoner returns retrieved results and answers for each subquery, while the logical-form language can directly output previously retrieved or calculated variables.Its symbolic representation is intended to make decomposition and reasoning more rigorous and interpretable.Source: §2.3 · p. 8, §2.3.2 · p. 10

2.3.3 Logical Form for Retrieval

• Sparse and dense retrieval capture complementary relevance features, but a single ensemble of their scores may not suit every question, especially multi-hop queries.The passage notes that proper nouns, times, numbers, and coordinates may need precise text indexes, while paragraph-like descriptions can benefit from coupled methods.Source: §2.3.3 · pp. 10–11
• Logical-form retrieval can use explicit keyword conditions as structured filters and choose sparse or dense indexes for precise or fuzzy searches.The disability-certificate example filters on location while retrieving supporting chunks.Source: §2.3.3 · p. 11
• When structured SPO knowledge cannot be retrieved, the solver can search semi-structured chunks or retrieve text passages for natural-language subproblems.This allows retrieval strategy to vary with the availability and form of the information.Source: §2.3.3 · p. 11

2.4 Knowledge Alignment

• Similarity-based retrieval can miss required relation types and direction, while openIE granularity noise and fragmented domain knowledge can weaken recall and professionalism.The stated alignment challenges include mismatched semantic relations, synonymous or similar nodes, and domain knowledge that should be connected but appears fragmented.Source: §2.4 · p. 11
• Concept-graph reasoning enhances offline indexing and online retrieval through instance standardization, instance-to-concept links, relation completion, and domain-knowledge injection.The approach adds descriptive context during extraction and supports domain concepts and terminology to address granularity mismatch.Source: §2.4 · p. 11

2.4.1 Enhance Indexing

• Index enhancement predicts synonymous instances for fusion, links instances to concepts, and completes hypernym and isA relations among concepts.The described completion can connect a concept through intermediate types to a broader class, strengthening graph connectivity.Source: §2.4.1 · p. 12

2.4.2 Enhance Retrieval

• At retrieval time, semantic-relation reasoning combines with similarity retrieval, using precise type matching and entity linking before reasoning when type matching fails.The examples connect Political Party with Political Faction and cataract patients with visually impaired people to find relevant material through semantic relations.Source: §2.4.2 · pp. 12–13

2.5 KAG-Model

• KAG-Model targets NLU, NLI, and NLG capabilities to reduce the complexity and cascading losses associated with multi-stage pipelines.The paper frames indexing and QA as involving similar classify, mention-detection, semantic-alignment, embedding, and generation steps.Source: §2.5 · p. 13

2.5.1 Natural Language Understanding

• Instruction reconstruction creates over 20,000 diverse NLU instructions, and the paper reports improved NLU benchmark performance after supervised fine-tuning.Strategies vary labels, input-output formats, and task descriptions; the reported evaluation uses OpenCompass benchmarks.Source: §2.5.1 · p. 14, Table 3 · p. 15

2.5.2 Natural Language Inference

• KAG-Model uses a conceptual knowledge set of 8,000 concepts and their relations to train semantic-reasoning instructions, with reported gains on hypernym discovery and general reasoning tasks.The reported comparisons include better performance than Llama3 on CMNLI, OCNLI, and SIQA, and outperformance of Llama3 and ChatGPT-3.5 on Hypernym Discovery.Source: §2.5.2 · pp. 14–15, Table 5 · p. 15, Table 4 · p. 15

2.5.3 Natural Language Generation

• K-LoRA trains on triples-to-text examples to recognize KG information formats and domain-specific linguistic style, while AKGF uses KG-scored response feedback.AKGF scores answers by correctly matched knowledge triples, with the reward also incorporating entity matches.Source: §2.5.3 · p. 16
• On CMedQA and BioASQ, the paper reports generation-performance enhancement using 1,000 randomly selected test instances from each dataset.The knowledge corpora come from non-selected CMedQA answers and BioASQ reference passages; Table 6 compares performance.Source: Table 6 · p. 16, §2.5.3 · p. 16

2.5.4 Onepass Inference

• OneGen assigns retrieval to added retrieval tokens and trains them with contrastive learning while other output tokens use language-model objectives.At inference, retrieval tokens retrieve on demand in the same model’s forward pass, avoiding a separate retriever.Source: §2.5.4 · p. 16
• OneGen outperforms original Self-RAG across datasets, with gains of 3.1pt on Pub and 2.8pt on ARC, and achieves comparable performance using 60K noisy samples.The passage also reports efficacy in retrieval-to-generation and performance gains from jointly training retrieval and generation.Source: Table 7 · p. 17, §2.5.4 · p. 17

3.1 Experimental Settings

• Experiments used 1,000 questions from each of three multi-hop QA validation sets, measured EM, F1, Recall@2, and Recall@5, and compared NativeRAG, HippoRAG, and IRCoT.The datasets were HotpotQA, 2WikiMultiHopQA, and MuSiQue.Source: §3.1 · p. 17

3.2.1 Overall Results

• F1 improved by 12.5%, 19.1%, and 12.2% over IRCoT + HippoRAG on HotpotQA, 2WikiMultiHopQA, and MuSiQue, respectively.EM increased by 11.5%, 19.8%, and 10.5% on those datasets, respectively.Source: §3.2.1 · p. 18, Table 8 · p. 18

3.3 Ablation Studies

• The ablations replace knowledge-alignment and logical-form-solving modules with alternative methods to assess their effects on final results.The study covers graph indexing and reasoning/retrieval alternatives.Source: §3.3 · p. 18, §3.3.1 · p. 18

3.3.2 Reasoning and Retrieval Ablation

• The reasoning/retrieval ablations compare chunk retrieval weighted by vector similarity and personalized PageRank with logical-form solving and hybrid retrieval variants.The logical-form solver uses graph reasoning and supporting-chunk retrieval when reasoning returns no result; the hybrid variant disables graph retrieval for direct reasoning.Source: §3.3.2 · p. 19

3.3.1 Knowledge Graph Indexing Ablation

• Top-5 recall for CRref3 increased by 9.2%, 28.4%, and 9.5% after knowledge alignment, averaging 15.7%.The accompanying graph analyses report increased relation density and rightward shifts in 1-hop, 2-hop, and 3-hop neighbor distributions after alignment.Source: Figure 9 · p. 20, §3.3.3 · p. 21

3.3.3 Experimental Results and Discussion

• Top-5 recall for LFSHref3 exceeded CRref3 by 15%, 32.2%, and 12.7% across the three datasets, averaging 19.9%.The analysis attributes the recall gains to multi-step query decomposition and retrieval, alongside integration of execution-derived SPO triples into chunk computation; it also reports increased computation time as a trade-off.Source: §3.3.3 · p. 21
• LFSref3 answered 33%, 34%, and 18% of questions from retrieved subgraphs and had F1 decreases of 2.2%, 0.1%, and 0.5%, while reducing computation time by 12%, 22%, and 18%.Because it answers from subgraphs, LFSref3 is not directly comparable in recall-rate analysis and is evaluated using F1.Source: Figure 8 · p. 20, §3.3.3 · p. 21
• Increasing the maximum reflection iterations helped re-plan from existing information when LFSrefn could not complete a solution, while LFSref1 had lower F1 than LFSref3.The reported F1 decreases were 0.6%, 1.6%, and 4.8%; errors or incomplete SPOs during construction could also produce incorrect sub-query answers, and the study frames slight metric decreases against reduced computation time as a feasible industrial trade-off.Source: §3.3.3 · p. 21

4.1 KAG for E-Goverment

• The E-Government application was built from 11,000 government-service documents and supports questions about service methods, materials, conditions, and locations.Its implemented modules include index building, logical-form-guided reasoning, semantic enhancement, and conditional summary generation.Source: §4.1 · p. 22
• Government-service reasoning generates a logical function from the question and graph index, locates the user's administrative area, searches by item name or group, and finds the requested service chunk.Documents are semantically chunked into matters and properties such as region, process, materials, location, and target audience.Source: §4.1 · p. 22
• For E-Government Q&A, semantic enhancement links items by synonymy and hypernymy, including differently named equivalent services and services under a shared broader category.Examples include differently named social-security-card services and two housing-provident-fund loan services sharing a hypernym.Source: §4.1 · p. 22
• KAG showed significant improvements in completeness and accuracy compared with NaiveRAG in the E-Government Q&A evaluation.The supplied table caption identifies Table 12 as ablation experiments, but provides no ablation outcomes.Source: §4.1 · p. 22

4.2 KAG for E-Health

• The E-Health application uses strongly constrained schemas for entities and translates query-derived logical forms into DSL queries over the knowledge graph.Logical forms retain query structure such as city, gender, age, and indicator values; expert DSL rules can also be triggered during graph queries.Source: §4.2 · pp. 22–23
• Recall reached 60.67% and precision reached 81.32% on sampled online Q&A queries; end-to-end accuracy was 77.2% for medical insurance inquiries and exceeded 94% for popular-science and 93% for indicator-interpretation intents.The medical application uses a term set to align entities with medical terms and retrieve hyponyms when users ask about a broader concept.Source: §4.2 · p. 23

5.1 DIKW Pyramid

• LLMFriSPG represents data, information, and knowledge hierarchically and adds reverse context-enhanced mutual indexing, addressing context support absent from SPG.The passage also notes that RDF/OWL uses binary triples and LPG lacks support for knowledge semantics and classification.Source: §5.1 · p. 23

5.2 Vector Similarity-based RAG

• Vector-similarity retrieval struggles to track long-distance knowledge associations and focus on key relations in complex questions, leaving information density and remote-knowledge association problems.The passage mentions fine-grained segmentation, chain-of-thought, and interactive retrieval as prior optimizations that do not remove these difficulties.Source: §5.2 · p. 23

5.3 Information Retrieval-based GraphRAG

• GraphRAG, ToG 2.0, and HippoRAG use OpenIE-derived graph structure, while GraphRAG's hierarchical summaries are described as difficult to use for multi-hop Q&A and incremental document updates.The passage also notes noise in OpenIE-built KGs and limited attention to semantic directionality and logical sensitivity.Source: §5.3 · p. 24

5.4 KG-based Question and Answering

• KG-based QA methods include edge retrieval and semantic parsing, but complex S-expressions can be hard for LLMs to understand and inconvenient for combining KBQA with RAG.KAG proposes decomposed logical forms that break multi-hop queries into sub-queries and support integrated SPO and chunk retrieval.Source: §5.4 · p. 24

5.5 Bidirectional-enhancement of LLMs and KGs

• Prior work explores LLMs for KG tasks and KGs as resources for LLM factual answers, planning, and knowledge editing; bidirectional enhancement has been partially achieved.Examples include LLM-based triple completion and KG resources for mitigating hallucinations.Source: §5.5 · p. 25

Overhead

• KAG requires multiple LLM calls and substantial intermediate-token generation, increasing computational, economic, and runtime overhead.Planning outputs logical functions as well as sub-problems, producing approximately twice as many tokens as sub-problem decomposition alone; the authors suggest smaller domain-specific models for some extraction and standardization tasks.Source: §6 · p. 25

Planning

• Complex-problem decomposition and planning require accuracy, stability, and solvability that the current framework does not yet optimize.The example shows that comparing ages of deceased people should use birth dates rather than age-at-death answers; its generated answer initially names the wrong older director.Source: §6 · p. 25

Knowledge extraction and alignment

• OpenIE lowers the threshold for building KGs but increases alignment challenges, while accuracy of extracting multiple knowledge types and consistency across extraction rounds remain unresolved.The authors also identify expert schema-constrained extraction as rigorous but labor-intensive, and propose using both approaches collaboratively.Source: §6 · pp. 25–26
• Incorrect extraction can impair response accuracy, particularly when comprehensive extraction is needed to answer questions such as the total number of team members.The passage frames this as a reason to balance rigor in vertical-domain decision-making with convenient information retrieval.Source: §6 · p. 26

Conclusion

• KAG combines symbolic KGs and parameterized LLMs, introducing LLM-friendly representation, mutual indexing, semantic alignment, logical-form-guided hybrid reasoning, and a KAG model.The paper reports improvements over competitive methods on HotpotQA, 2wiki, and musique, alongside E-Government and E-Health case verifications.Source: §7 · p. 26

Future Work

• Future work includes reducing KG construction cost and improving reasoning interpretability and transparency, with multiple-knowledge extraction and domain knowledge injection among the listed topics.The passage also lists OneGraph-based alignment, large-scale instruction synthesis, and suppression of knowledge-logic hallucinations.Source: §7 · p. 26
• The study does not enhance models for complex-problem decomposition and planning, and proposes using KAG as a reward model to assess planning-result accuracy, stability, and solvability.The authors identify pre-training, SFT, and CoT strategies as future avenues for improving model adaptability to logical forms and planning and reasoning capabilities.Source: §7 · p. 26
• The authors plan continued collaboration with OpenKG to address technical issues in LLM–KG collaboration.Source: §7 · p. 26

• The example uses logical form to guide question planning, reasoning, retrieval, and answer generation, with multiple reflection rounds rephrasing questions.Source: Table 13 · p. 32

• The table presents cases of reasoning with logical form.Source: Table 14 · p. 33