NodeRAG:

Structuring Graph-based RAG with Heterogeneous Nodes

Tianyang Xu1 , Haojie Zheng2 , Chengze Li1 , Haoxiang Chen1

Yixin Liu3 , Ruoxi Chen, Lichao Sun3
1
Columbia University, 2 University of Pennsylvania, 3 Lehigh University
tx2240@[Link], haojiez@[Link]

Abstract et al., 2024). Despite recent progress, current

RAG methods face notable shortcomings in
Retrieval-augmented generation (RAG) em-
handling multi-hop reasoning (Luo et al., 2023;
powers large language models to access ex-
Wang et al., 2024b) and summary-level queries
arXiv:2504.11544v1 [[Link]] 15 Apr 2025

ternal and private corpus, enabling factually

consistent responses in specific domains. By (Han et al., 2024a; Wen et al., 2023) due to
exploiting the inherent structure of the cor- their insufficient utilization of data structures
pus, graph-based RAG methods further enrich and lack of high-level understanding of the text
this process by building a knowledge graph corpus. Graph-based RAG methods (Tian et al.,
index and leveraging the structural nature of 2024; Park et al., 2023) have been proposed
graphs. However, current graph-based RAG ap- to enhance retrieval and question-answering
proaches seldom prioritize the design of graph
performance, specifically addressing the two main
structures. Inadequately designed graph not
only impede the seamless integration of di- challenges faced by traditional RAG approaches.
verse graph algorithms but also result in work- Leveraging LLMs to decompose raw data into
flow inconsistencies and degraded performance. graph structures (Jiménez Gutiérrez et al., 2024;
To further unleash the potential of graph for He et al., 2024) for utilizing structural information,
RAG, we propose NodeRAG, a graph-centric as well as employing LLMs for summary-based
framework introducing heterogeneous graph enhancements (Edge et al., 2024; Guo et al., 2024)
structures that enable the seamless and holis-
to derive insights beyond the original text, has
tic integration of graph-based methodologies
into the RAG workflow. By aligning closely
gradually become mainstream approaches.
with the capabilities of LLMs, this frame- However, previous Graph-based RAG works
work ensures a fully cohesive and efficient (Trajanoska et al., 2023; Jiménez Gutiérrez et al.,
end-to-end process. Through extensive experi- 2024) have rarely considered the critical role of
ments, we demonstrate that NodeRAG exhibits graph structures, i.e., what forms of graph better
performance advantages over previous meth- support RAG. Among existing approaches, knowl-
ods, including GraphRAG and LightRAG, not edge graphs (Sanmartin, 2024; Wang et al., 2024b)
only in indexing time, query time, and stor-
extract triples, with the graph containing only struc-
age efficiency but also in delivering superior
question-answering performance on multi-hop tural information, yet retrieval context remains con-
benchmarks and open-ended head-to-head eval- fined to text chunks, which often lack semantic co-
uations with minimal retrieval tokens. Our herence and include unrelated information. While
GitHub repository could be seen at this link. current methods attempt to incorporate more infor-
mation into the graph and extract deeper insights,
they suffer from inefficiencies and inconsistencies
1 Introduction
due to inadequately designed structures. For in-
Retrieval-augmented generation (RAG) has stance, as illustrated in Figure 1, GraphRAG (Edge
emerged as a solution to the challenges posed et al., 2024), adopt a tightly coupled entity-event
by the rapid evolution of real-world knowledge homogeneous structure, hindering the integration
domains (Fan et al., 2024), coupling large lan- of original context and summary information into
guage models (LLMs) with an external retrieval the graph. This results in inconsistencies in re-
mechanism to ensure the generation of factually trieval methods (separating local and global re-
consistent and contextually relevant information trieval) and leads to coarse-grained retrieval, where
(Tonmoy et al., 2024; Shrestha et al., 2024; Liu retrieving an entity indiscriminately includes all

1
 NaïveRAG GraphRAG HippoRAG Light RAG NodeRAG
Homogenous Graph Knowledge Triples Graph Homogenous Graph Heterogenous Graph
Community Node
N
Text Chunk 1 Node S N S N R
N R
S S Edge
S N
Text Chunk 2 R N R R N
Edge

R N R
N R N Node
Text Chunk 3
Community
Node
S S S Text
N
N
N R
S S
… S Edge
R N N Text S
Text Chunk 1

Text Chunk n
… Node
Community Summary
S S S
Text Chunk n N

N Entity Information Text Original Text Chunks S Summarization Information R Relationship

Figure 1: Comparsions between NodeRAG and other RAG systems. NaïveRAG retrieving fragmented text chunks, leads to
redundant information. HippoRAG introduces knowledge graphs but lacks high-level summarization. GraphRAG retrieves
community summaries but may still produce coarse-grained information. LightRAG incorporates one-hop neighbors but retrieves
redundant nodes. In contrast, NodeRAG utilizes multiple node types, including high-level elements, semantic units, and
relationships, enabling more precise, hierarchical retrieval while reducing irrelevant information.

associated events, adding irrelevant information. (1) Better Graph Structure for RAG The graph
To address these limitations, we propose structure serves as the foundation for graph-based
NodeRAG, which is built around a well-designed RAG where significance has been overlooked. Our
Heterogeneous Graph, comprehensively consider- work emphasizes its importance and introduces a
ing the entire process of graph indexing and search- graph structure that better supports RAG.
ing, enabling fine-grained retrieval. The hetero-
(2) Fine-grained and Explainable Retrieval The
graph adheres to the principle of unfolding and
heterograph enables fine-grained and functionally
flattening, decomposing different types of infor-
distinct nodes, allowing graph algorithms to effec-
mation to construct a heterogeneous fully nodal-
tively and reasonably identify key multi-hop nodes.
ized graph where nodes serve distinct functions
This leads to more relevant retrieval with minimal
and roles. This means that entities, relationships,
retrieval context, enhancing both precision and in-
original text chunks, independently decomposed
teroperability.
events from text chunks, and summaries extracted
by LLMs are all represented as nodes within the
(3) Unified-Level Information Retrieval Decom-
graph. The heterograph not only encapsulates infor-
posed information from documents and extracted
mation from the original corpus but also extends be-
insights from LLMs are not treated as separate
yond it, incorporating enriched insights such as key
layer but are instead unified as nodes within the
node attributes, and high-level discoveries. Each
heterograph. This integration allows for a cohesive
node in heterograph consists of unstructured con-
framework capable of handling information needs
tent, while preserving structural connections be-
across different levels.
tween nodes, striking a balance between structural
integrity and flexibility. As illustrated in Figure 1, In addition, extensive experiments demonstrate
for a multi-hop question, NodeRAG can retrieve a that NodeRAG not only outperforms previous
semantically coherent, independent event (seman- graph-based RAG methods on multi-hop tasks but
tic unit) and high-level discoveries (high-level ele- also exhibits superior performance in open-ended
ments) related to key entities such as Harry, Neville, head-to-head evaluations. With minimal retrieval
and the three-headed dog using graph algorithms, tokens, it achieves highly precise retrieval while
providing explainable and fine-grained retrievals as also demonstrating system-level efficiency advan-
well as high-level understanding. tages, including improvements in indexing time,
The key contributions of our work can be sum- query time, and storage efficiency, as shown in
marized in three main aspects. appendix A.

2
2 NodeRAG For instance, V{N,R,S} represents the subset con-
taining only entity, relationship, and semantic unit
The NodeRAG pipeline is built on a foundational nodes.
graph structure defined as the heterograph, which
V{T,S,A,H} contain rich informational content
will be introduced in Section 2.1. The workflow
and are classified as retrievable nodes. In contrast,
is divided into two primary stages, graph indexing
V{N,O} , which represent names or titles, act solely
and graph searching. Graph indexing comprises
as critical linkage and entry points within the graph
three components, graph decomposition, graph aug-
but are not directly retrievable. For example, VH
mentation, and graph enrichment, which are dis-
provides detailed context for high-level concepts,
cussed in Sections 2.2, 2.3, and 2.4, respectively.
while VO represents the corresponding title and
This stage integrates various types of nodes and
keywords but does not contribute directly to the
edges into the heterograph by leveraging LLMs
retrieved content. Additionally, VR , is a nodalized
and graph algorithms. The subsequent stage, graph
edge, acting as connector nodes and secondary re-
searching, is detailed in Section 2.5 and combines
trievable nodes, contributing to the retrieval context
the structural advantages of the heterograph with
but not serving as graph entry points.
graph algorithms to efficiently retrieve relevant in-
formation. Moreover, the fundamental concepts 2.2 Graph Decomposition
and implementation details of the graph algorithms
used in the pipeline are provided in Appendix C, First, we define a null heterograph G 0 . The initial
while the prompting instructions for LLMs can be step involves employing a LLM to decompose text
found in Appendix E for reference. chunks from the source corpus into three primary
node types: semantic units (S), entities (N ), and
2.1 Heterograph relationships (R). These nodes are then intercon-
nected to construct the initial heterograph. This
The concept of the heterograph embodies the prin-
process can be formalized as:
ciple of comprehensive unfolding and flattening of
information into a fully nodalized structure. This G 1 = G 0 ∪{v ∈ V, ed , er ∈ E | Ψ(v) ∈ {S, N, R}},
structure achieves its granularity through the in- Where e represents the connecting edges between
tegration of seven hetero node types: entity (N ), semantic units and entity nodes, as well as between
relationship (R), semantic unit (S), attribute (A), relationship nodes and their corresponding source
high-level elements (H), high-level overview (O), and target entities. For instance, if “Hinton was
and text (T ). Each node type is tailored to represent awarded the Nobel Prize for inventing backprop-
specific roles and characteristics of the information, agation” serves as v ∈ VS derived from a text
enabling a fine-grained and functional decompo- chunk, then Hinton, Nobel Prize, and backpropaga-
sition of data. Mathematically, the heterograph is tion represent v ∈ VN nodes, with ed denoting their
defined as: connections to v ∈ VS . An example of v ∈ VR
G = (V, E, Ψ), would be “Hinton received Nobel Prize”, where
er represents the edge connecting the source node
where G is the heterograph, V represents the set of Hinton to the target node Nobel Prize.
nodes, E is the set of edges, and Ψ : V → Types is
a mapping function that assigns each node v ∈ V to Semantic unit (S) The semantic unit acts as a
a specific type. The set of node types, corresponds local summary, representing an independent event
to the seven predefined types: unit in a paraphrased form. It serves as the core
node for graph augmentation and improving search
Types = {N, R, S, A, H, O, T }.
quality. Since the division of text chunks is not
For any node v, Ψ(v) defines its type, with each
based on semantics, unrelated or unassociated con-
node type performing a distinct and well-defined
tent may coexist within a single chunk. This con-
function, as detailed in subsequent sections and
text noise increases entropy, leading to degraded
appendix C. For each e ∈ E, the default weight
quality when using text chunks for graph augmen-
of e is set to 1, representing a basic connection
tation or searching due to their coarse granularity
between two nodes. Furthermore, we define Vtypes
and irrelevant information.
as the subset of nodes corresponding to a subset set
types ⊆ Types, formally expressed as: Entity (N ) and Relationship (R) Entities (N )
Vtypes = {v ∈ V | Ψ(v) ∈ types}. are nodes that exclusively represent entity names,

3
 Step 2: Graph Augmentation
2.1 Node Importance Based Augmentation 2.2 Community Detection Based Aggregation
Leiden Algorithm O O
N Important entity Community 1 S H
H
R S A
R R A S A S
R R S

A Community 2 A
N Important entity O O
R R R S H H
R R R S A
S
High Level Elements Community 3
S H O
K-core U Betweenness Centrality
Attributes
High Level
R R R S
H Semantic matching
A Overview
Relationship O
Step 1: Graph Decomposition R
Text chunk
Step 3: Graph Enrichment
R Text Text insert S
R
R S Text S
R
Entity N
N Embedding nodes Non-Embedding nodes
N Embed
N N N S A N
N N O R
H Text
S
S S S S S S
Semantic Unit HNSW Semantic edges
Semantic edge A H
S
H
Text Text Text
Heterograph Text S
S H H

Figure 2: Main indexing workflow of NodeRAG. It illustrates the step-by-step construction of the heterograph,
including the process of graph decomposition, graph augmentation, and graph enrichment

while relationships (R) are also transformed into human reading behavior, where all relevant con-
nodes that connect source and target entities. These tent associated with a critical entity is reviewed
entities and relationships are directly connected to before synthesizing its attributes. The summariza-
semantic units (S), as v ∈ VS serves as the smallest, tion specifically focuses on the important entities
contextually meaningful representation of events identified within the corpus, rather than processing
within text chunks. This connection ensures that all entities, ensuring both precision and efficiency.
entities and relationships remain decoupled from The selection of important entities, N ∗ , is guided
specific events, allowing them to function inde- by two complementary metrics: K-core decom-
pendently while still being anchored to relevant position (Seidman, 1983; Kong et al., 2019) and
contexts. Such a design prevents redundant infor- betweenness centrality (Brandes, 2001). K-core
mation and enables a flexible graph structure. identifies nodes in densely connected subgraphs
that are critical to graph cohesion, while between-
2.3 Graph Augmentation ness centrality highlights nodes that act as bridges
The heterograph G 1 provides a foundational low- for information flow. These metrics are denoted as
level structure. However, it lacks high-level orga- K(G 1 ) and B(G 1 ), where K(·) and B(·) represent
nization and contextual insights. To further aug- the selected entity nodes from the graph. The final
ment the graph, we implement two primary meth- set of important entities is defined as:
ods: node importance-based augmentation and N ∗ = K(G 1 ) ∪ B(G 1 ).
community detection-based aggregation, which re- Entity attributes are constructed directly from
spectively capture the perspectives of individual relationships and semantic units, bypassing raw
node significance and structural cohesion within texts to avoid redundancy. Each generated attribute
the graph. node is added to the graph and connected to its
corresponding entity node via the edge ea . This
Node Importance Based Augmentation We pri- update to the graph is represented as:
oritize the selection of structurally significant and
G 2 = G 1 ∪ {v ∈ V, ea ∈ E | Ψ(v) ∈ {A}}.
functionally pivotal entities. These key entities,
along with their associated semantic units and re- Community Detection Based Aggregation We
lationships, are processed through LLMs to gen- first apply the Leiden algorithm (Traag et al., 2019)
erate attribute summaries. This approach mirrors to G 2 to perform community detection, segment-

4
ing the graph into closely related substructures, information remains searchable within graph.
denoted as communities. Each node v ∈ G 2 is G 4 = G 3 ∪ {v, es | Ψ(v) = T },
assigned to a specific community Cn , where Cn where es denotes the edges connecting text chunks
represents the n-th community identified by the to their relevant semantic units.
algorithm. Within each community Cn , an LLM is
utilized to analyze the aggregated content, extract- Embedding As mentioned in Section 2.1, v ∈
ing high-level elements (H) that encapsulate the V{T,A,S,H} contains rich informational context
core information of the community, such as sum- where vector similarity is highly effective. Con-
maries, sentiment analysis, and other significant versely, v ∈ V{N,O} , which includes names and
insights. For each generated high-level element titles represented as words or phrase, is less suit-
node v ∈ VH , it is essential to establish meaning- able for vector similarity methods. To address this
ful connections eh with relevant nodes from G 2 limitation, we developed a dual search mechanism.
to preserve the graph’s structural coherence. To During the embedding process, we selectively em-
accomplish this, we propose semantic matching bed only a subset of the graph’s data, specifically
within community algorithm. This algorithm iden- v ∈ V{T,A,S,H} . This targeted embedding step is
tifies the most semantically related nodes within the crucial for reducing storage overhead while pre-
same community Cn for each high-level element serving efficient search capabilities.
node. To achieve this, K-means clustering (Mac-
HNSW Semantic Edges The Hierarchical Navi-
Queen et al., 1967) is applied to the embeddings
gable Small World (HNSW) algorithm (Malkov
of v ∈ V{S,A,H} . The number of clusters K is
q and Yashunin, 2018) is an approximate nearest
determined as |V{S,A,H} |, where |V{S,A,H} | rep- neighbor search method that organizes data into
resents the total number of nodes labeled S, A, or a multi-layer graph structure to efficiently retrieve
H. An edge eh (v, v ′ ) exists between v ∈ V{S,A,H} semantically similar nodes. It represents the data
and v ′ ∈ VH if both v and v ′ belong to the same as a layered graph H = {L0 , L1 , . . . , Lm }, where
semantic cluster Sk and the same community Cn . L0 is the base layer containing the densest se-
Additionally, the LLM can extract a keyword title mantic similarity connections, and higher layers
for each high-level element, referred to the high- (Li , i > 0) are sparsely connected to facilitate
level overview (O), which is used for dual search coarse-grained navigation. H is built iteratively.
as elaborated in Section 2.5. Each v ∈ VH and When a new node is added, it is inserted into a ran-
v ∈ VO will have a corresponding connection eo . dom level and all layers below it, connecting to sim-
The updated graph G 3 incorporates high-level el- ilar neighbors based on cosine similarity. Higher
ements (H) and their corresponding connections layers remain sparse with long-range connections,
(eh , eo ). It is defined as: while L0 focuses on dense local relationships. The
G 3 = G 2 ∪ {v ∈ V, eh , eo ∈ E | ψ(v) = {H, O}}. search starts at the sparsely connected top layer,
and progressively descends to L0 . In our work,
2.4 Graph Enrichment the base layer L0 of the HNSW graph, which en-
codes semantic relations between nodes, is inte-
In the previous process of generating the hetero-
grated with the heterograph G. The updated graph,
graph, G 3 already contains a wealth of information.
denoted as G 5 , is expressed as:
However, certain unique and additional details can
still further enrich the heterograph, enabling it to G 5 = G 4 ∪ L0 .
not only preserve the entirety of the original text’s The inclusion of L0 enhances the heterograph’s
information but also gain enhanced features and search capabilities by incorporating semantic dense
insights that go far beyond the source material. proximity edges, augmenting its structural infor-
mation in the graph. When an edge already ex-
Text Insertion As mentioned earlier, text chunks ists in G 4 , adding the corresponding edge from L0
are not directly incorporated into G during graph increases its weight by 1, reinforcing frequently
augmentation due to their semantic incoherent na- occurring connections.
ture. However, original text chunks hold significant
value as they contain detailed information, which is 2.5 Graph Searching
often lost during the LLM transformation process. We first apply a dual search mechanism to identify
Therefore, it is essential to ensure that the original entry points within the heterograph. Subsequently,

5
 Query: How did HNSW
Weight flowing Cross nodes
Harry Potter embedding Entry points Entry points + Cross nodes
first learn about H
R H O
Hogwarts and S A
who delivered H N
A R S N Retrievals
the letter to him? Accurate search S O
N S
Harry Text Text A H S R A Text
Decompose N N
Potter O
Hogwarts
O
Text
Letter
Dual Search Shallow PPR Filtering retrieval nodes

Figure 3: This figure focuses on the querying process, where entry points are extracted from the original query,
followed by searching for related nodes that need to be retrieved in the heterograph.

a shallow Personalized PageRank (PPR) algorithm that relevance remains localized to the neighbor-
is employed to extract cross nodes. The combina- hoods of the entry points. This early stop strategy
tion of entry point nodes and cross nodes is then prevents excessive diffusion to distant or irrelevant
filtered to produce the final retrieval. parts of the graph, focusing instead on multi-hop
nodes near the entry points. Let P be the normal-
Dual Search Dual search combines exact match- ized adjacency matrix of G, where Pij represents
ing on title nodes and vector similarity search on the transition probability from node i to node j.
rich information nodes to identify entry points in The PPR process starts with a personalization vec-
the heterograph G. Given a query, the LLM extracts tor p ∈ R|V| , where pi = 1/|Ventry | if vi ∈ Ventry ,
entities N q and embeds the query into vector (q). and pi = 0 otherwise. The PPR score vector π (t)
The entry points are defined as: after t iterations is computed iteratively as:
Ventry = {v ∈ V | Φ(v, N q , q)}, π (t) = αp + (1 − α)P ⊤ π (t−1) , π (0) = p,
where the condition function Φ(v, N q , q) is de- where α ∈ (0, 1) is the teleport probability that
fined as: ( balances restarting at entry points and propagat-
q v ∈ V{N,O} ∧ M(N q , v), ing through the graph. After t iterations, the top-k
Φ(v, N , q) =
v ∈ V{S,A,H} ∧ R(q, v, k). nodes with the highest PPR scores for each type are
Here, the exact matching function M(v ∗ , v) re- selected as cross nodes, denoted as Vcross . In our de-
turns true if a node matches one of the extracted en- fault setting, we use α = 0.5 and t = 2 to achieve
tities by word level string matching. Additionally, a balance between exploration and convergence.
the similarity-ranking function R(q, v, k) returns Filter Retrieval Nodes Finally, the retrieval
true if a node ranks among the top-k most similar nodes are filtered from the union of entry nodes
to q based on the HNSW algorithm. By lever- and cross nodes to include only retrievable nodes
aging the non-retrievability of v ∈ V{N,O} , they of v ∈ V{T,A,S,H,R} . v ∈ V{N,O} , which contain
serve exclusively as entry points to the graph with- only keywords without informational content, are
out contributing directly to the retrievable content. excluded from the retrieval context. The final set
Only nodes identified through the shallow PPR as of retrieval nodes is therefore defined as:
closely related to all entry points are included in the Vretrieval = { v ∈ Ventry ∪ Vcross |
retrieval results as cross nodes. This ensures that
ψ(v) ∈ {T, S, A, H, R}}
the effects of noisy or ambiguous queries, which
may lead to errors in exact matching, do not di- 3 Evaluation
rectly impact the retrieval process. Any indirect
effects are further minimized by the graph algo- We evaluate NodeRAG’s performance across three
rithm, enhancing the robustness of the retrieval different multihop benchmarks, HotpotQA (Yang
system. et al., 2018), MuSiQue (Trivedi et al., 2022b),
MultiHop-RAG (Tang and Yang, 2024), and an
Shallow PPR Personalized PageRank (PPR) open-ended head to head evaluation RAG-QA
identifies relevant nodes in the heterograph G by Arena (Han et al., 2024b) across six domains. And
simulating a biased random walk starting from a we compare our method against several strong and
set of entry points. In our approach, we use shallow widely used RAG methods as baseline models, in-
PPR, limiting the number of iterations t to ensure cluding NaiveRAG (Lewis et al., 2020), HyDE

6
 Part I: General comparisons

HotpotQA MuSiQue MultiHop Arena-Writing Arena-Tech Arena-Science Arena-Recreation Arena-Lifestyle Arena-FiQA

Methods
Acc.↑ #Token↓ Acc.↑ #Token↓ Sco.↑ #Token↓ W+T↑ #Token↓ W+T↑ #Token↓ W+T↑ #Token↓ W+T↑ #Token↓ W+T↑ #Token↓ W+T↑ #Token↓

NaiveRAG 87.50% 9.8k 39.43% 9.6k 0.56 8.9k 0.663 9.4k 0.689 9.1k 0.526 9.0k 0.720 9.3k 0.817 9.1k 0.926 9.1k
HyDE 73.00% 10.0k 33.14% 9.8k 0.53 9.4k 0.789 9.6k 0.863 9.3k 0.823 9.3k 0.777 9.5k 0.829 9.3k 0.949 9.3k

LightRAG 79.00% 7.1k 36.00% 7.4k 0.50 7.9k 0.754 6.3k 0.937 6.9k 0.840 7.1k 0.800 6.2k 0.817 6.8k 0.937 7.7k
GraphRAG 89.00% 6.6k 41.71% 6.6k 0.53 7.4k 0.749 6.4k 0.943 6.7k 0.863 6.7k 0.806 6.6k 0.863 6.8k 0.960 6.8k
NodeRAG 89.50% 5.0k 46.29% 5.9k 0.57 6.1k 0.794 3.3k 0.949 3.8k 0.903 4.2k 0.886 3.4k 0.949 3.3k 0.977 3.4k

Part II: Pairwise Comparisons

Domain M1 vs M2 Win (M1) Tie Win (M2) Domain M1 vs M2 Win (M1) Tie Win (M2) Domain M1 vs M2 Win (M1) Tie Win (M2)

NodeRAG vs GraphRAG 0.520 0.126 0.354 NodeRAG vs GraphRAG 0.531 0.126 0.343 NodeRAG vs GraphRAG 0.691 0.120 0.189
NodeRAG vs LightRAG 0.486 0.103 0.411 NodeRAG vs LightRAG 0.526 0.143 0.331 NodeRAG vs LightRAG 0.651 0.115 0.234
NodeRAG vs NaiveRAG 0.749 0.034 0.217 NodeRAG vs NaiveRAG 0.800 0.017 0.183 NodeRAG vs NaiveRAG 0.851 0.018 0.131
NodeRAG vs HyDE 0.531 0.155 0.314 NodeRAG vs HyDE 0.440 0.189 0.371 NodeRAG vs HyDE 0.349 0.228 0.423
GraphRAG vs LightRAG 0.320 0.303 0.377 GraphRAG vs LightRAG 0.406 0.154 0.440 GraphRAG vs LightRAG 0.297 0.303 0.400
FiQA Recreation Writing
GraphRAG vs NaiveRAG 0.754 0.092 0.154 GraphRAG vs NaiveRAG 0.714 0.080 0.206 GraphRAG vs NaiveRAG 0.691 0.092 0.217
GraphRAG vs HyDE 0.491 0.132 0.377 GraphRAG vs HyDE 0.377 0.137 0.486 GraphRAG vs HyDE 0.177 0.126 0.697
LightRAG vs NaiveRAG 0.711 0.106 0.183 LightRAG vs NaiveRAG 0.691 0.063 0.246 LightRAG vs NaiveRAG 0.731 0.080 0.189
LightRAG vs HyDE 0.514 0.143 0.343 LightRAG vs HyDE 0.349 0.171 0.480 LightRAG vs HyDE 0.211 0.178 0.611
NaiveRAG vs HyDE 0.611 0.063 0.326 NaiveRAG vs HyDE 0.674 0.069 0.257 HyDE vs NaiveRAG 0.857 0.040 0.103

NodeRAG vs GraphRAG 0.640 0.114 0.246 NodeRAG vs GraphRAG 0.497 0.200 0.303 NodeRAG vs GraphRAG 0.543 0.154 0.303
NodeRAG vs LightRAG 0.623 0.131 0.246 NodeRAG vs LightRAG 0.538 0.208 0.254 NodeRAG vs LightRAG 0.497 0.137 0.366
NodeRAG vs NaiveRAG 0.800 0.040 0.160 NodeRAG vs NaiveRAG 0.829 0.085 0.086 NodeRAG vs NaiveRAG 0.777 0.046 0.177
NodeRAG vs HyDE 0.526 0.205 0.269 NodeRAG vs HyDE 0.423 0.280 0.297 NodeRAG vs HyDE 0.543 0.160 0.297
GraphRAG vs LightRAG 0.429 0.120 0.451 GraphRAG vs LightRAG 0.361 0.343 0.296 GraphRAG vs LightRAG 0.400 0.234 0.366
Lifestyle Science Tech
GraphRAG vs NaiveRAG 0.680 0.074 0.246 GraphRAG vs NaiveRAG 0.829 0.108 0.063 GraphRAG vs NaiveRAG 0.657 0.097 0.246
GraphRAG vs HyDE 0.354 0.097 0.549 GraphRAG vs HyDE 0.354 0.172 0.474 GraphRAG vs HyDE 0.463 0.143 0.394
LightRAG vs NaiveRAG 0.663 0.046 0.291 LightRAG vs NaiveRAG 0.828 0.119 0.053 LightRAG vs NaiveRAG 0.691 0.075 0.234
LightRAG vs HyDE 0.349 0.120 0.531 LightRAG vs HyDE 0.308 0.189 0.503 LightRAG vs HyDE 0.463 0.097 0.440
HyDE vs NaiveRAG 0.709 0.028 0.263 HyDE vs NaiveRAG 0.840 0.074 0.086 HyDE vs NaiveRAG 0.606 0.051 0.343

Table 1: Part I: General Comparisons evaluates NaiveRAG, HyDE, LightRAG, GraphRAG, and NodeRAG on
HotpotQA and MuSiQue (accuracy and average tokens) and in the Arena using Win+Tie ratios and average tokens.
Part II: Pairwise Comparisons shows the fraction of “wins" (Win(M1)), “ties" (Tie), and “losses" (Win(M2))
when comparing one RAG method against another (e.g., NodeRAG vs. GraphRAG). Bold values highlight the best
performance.

(Gao et al., 2022a), GraphRAG (Edge et al., 2024), ation, Writing, Lifestyle, Science, and Technology.
LightRAG (Guo et al., 2024). The details of these We conduct comprehensive pairwise comparisons
datasets and baseline models are introduced in Ap- among all method combinations and calculate the
pendix B. corresponding win and tie rates for each matchup,
thereby identifying the better RAG system.
3.1 Metrics
General Comparison In the first part, we eval-
uate NaiveRAG, HyDE, LightRAG, GraphRAG, 3.2 Implementation details
and NodeRAG across four benchmark datasets. For
HotpotQA and MuSiQue benchmarks, we assess
accuracy (Acc) to measure effectiveness and the By default, all these RAG methods are imple-
average number of retrieved tokens (#Token) to mented with GPT 4o-mini, and the temperature
evaluate efficiency. For the MultiHop-RAG bench- is set to 0 across the entire evaluation. Meanwhile,
mark, we adopt its original evaluation metric, Score we identify a potential unfairness in the current
(Sco), while still using #Token to gauge retrieval evaluation setup, evident in several key areas. No-
efficiency. Lastly, for the RAG-QA Arena bench- tably, the baselines vary in their choice of prompts
mark, we continue to track #Token for efficiency used to synthesis the final response based on re-
and employ a win and tie ratio (W+T) against gold trieved information. Therefore, we standardized re-
responses as a measure of performance across dif- sponse prompts for every method. Our initiative to
ferent methods. standardize these settings also benefits other meth-
ods like GraphRAG, improving their performance
Pairwise Comparsion In this part, the evalua- compared to their default setting, underscoring the
tion focuses exclusively on the RAG-QA Arena broader value of establishing fair and consistent
benchmark, covering six domains: FiQA, Recre- evaluation standards.

7
3.3 Results both accuracy and computational efficiency makes
General Comparison As shown in Part I of Ta- NodeRAG the optimal choice for a wide range of
ble 1, NodeRAG consistently outperforms compet- RAG tasks, from research applications to deploy-
ing methods on HotpotQA, MuSiQue, and Mul- ments in resource-constrained environments.
tiHopRAG, demonstrating the highest accuracy
while retrieving noticeably fewer tokens. For ex-
4 Ablation experiments
ample, for MuSiQue, NodeRAG attains an accu-
18
racy of 46.29%, surpassing GraphRAG (41.71%) Accuracy 7.2K
48% 16
and LightRAG (36.00%). In HotpotQA, while Avg Tokens 7.0K
NodeRAG achieves a slightly higher accuracy 46% Avg Time (s) 14
11.3 6.8K
12
(89.50% vs. 89.00% for GraphRAG), it does so 44% 9.7
6.6K
10

with only 5k retrieved tokens, which is 1.6k fewer 7.5 8

42% 6.2 6.4K
than GraphRAG. In the RAG-QA Arena bench- 5.4 5.8 5.4 5.7 5.5 6

40% 6.2K
mark, graph-enhanced RAG systems exhibit a clear 4

advantage over traditional approaches. Notably, 6.0K

2
38%
100 200 500 1000 5.8K
0
NodeRAG achieves the highest win and tie ra- 1 2 3 10 15
tio in each of the five domains while keeping re- PPR Max Iterations
trieval costs minimal. For example, it attains a ratio Figure 4: Ablation analysis on PPR iterations.
of 94.9%, notably surpassing GraphRAG’s 86.3%
We conducted ablation experiments on the
and LightRAG’s 81.7% in the Lifestyle domain,
MuSiQue dataset, adhering to the same settings
and does so with less than half the retrieved tokens
and evaluation metrics described earlier. We specif-
compared to the other models. It can also be no-
ically examined the impact of four key submodules:
ticed that graph-enhanced RAG systems generally
shallow PPR, cross-node interactions, HNSW se-
retrieve fewer tokens than traditional RAG across
mantic edges, and dual search.
all benchmarks. These results confirm NodeRAG’s
We first investigated the variation in PPR itera-
remarkable effectiveness and efficiency, demon-
tions and examined whether shallow PPR offers ad-
strating that our heterograph can significantly boost
vantages. PPR, with a few iterations, performs bet-
RAG performance across diverse tasks.
ter than deep PPR because it highlights important
Pairwise Comparsion Across all the six do- nodes that are closer to the entry points. Moreover,
mains, NodeRAG consistently achieves higher win early stopping reduces unnecessary computational
ratios against GraphRAG, LightRAG, NaiveRAG, overhead, leading to improved retrieval efficiency.
and HyDE, demonstrating notable dominance, Moreover, we evaluate the performance of ap-
for instance, in the Lifestyle domain, NodeRAG plying the top-k vector similarity method to all
achieves 0.640 win rate against GraphRAG, 0.623 node data in the graph. Although increasing the
against LightRAG, 0.800 against NaiveRAG and retrieval context, its performance remains lower
0.526 against HyDE. GraphRAG, LightRAG, than the basic version. This confirms the necessity
NaiveRAG, and HyDE show scattered successes, of cross-nodes in our method, as they help identify
such as LightRAG edging out NaiveRAG (0.649 vs. important multi-hop nodes. Second, performing
0.246) in Recreation, GraphRAG beats LightRAG vector similarity solely on node data consistently
(0.361 vs. 0.296) in Science, yet their overall win outperforms the naive RAG approach of similar-
rates remain lower when compared to NodeRAG. ity on text chunks, demonstrating the advantages
Notably, these trends persist across other domains brought by graph-based data augmentation.
like Writing, Recreation, Science, and Tech, fur- In addition, without integration of accurate
ther underscoring NodeRAG’s leading position, fol- search in dual search, accuracy drops to 44.57%,
lowed by LightRAG and GraphRAG, showing the and the token count increases to 9.7k. This is be-
superiority of our method. cause losing entity and high-level overview nodes
In general, NodeRAG not only achieves the as entry points causes nodes with long texts, such
highest accuracy rate and the lowest retrieval to- as text nodes, to have higher weights after shallow
ken count in general benchmarks but also outper- PPR. Since vector similarity entry nodes are more
forms all other baselines in preference evaluation frequently connected to T nodes, while accurate
comparisons. This unparalleled performance in entry nodes are more connected to S, A, and H

8
nodes, the absence of accurate search disrupts this process extensive external documents holistically
balance. (Jiang et al., 2024b). RAG has been applied to
Finally, we investigate the effect of HNSW. various domain-specific knowledge bases, such as
HNSW introduces semantic edges to the hetero- BioRAG and MedicalRAG (Wang et al., 2024a;
graph, and removing this integration results in per- Wu et al., 2024; Jiang et al., 2024a). RAG also
formance degradation. This is because HNSW en- struggles with corpus-wide understanding tasks,
hances connectivity between semantically related like query-focused abstractive summarization,
nodes, enabling more efficient and meaningful re- which require synthesizing knowledge across large
trieval. datasets.

Method Accuracy Time (s) Tokens (k)

NodeRAG (Ours) 46.29% 4.05 5.96
w/o HNSW 41.71% 4.92 6.78 RAG over Hierarchical Index To overcome the
w/o Dual Search 44.57% 4.72 9.70 limitations of traditional RAG, advanced systems
w/o Cross Node
Top-k = 10 41.71% 4.15 4.27
integrate hierarchical indexing to incorporate doc-
Top-k = 20 43.43% 4.70 7.89 ument summaries and enhance retrieval perfor-
Top-k = 30 42.29% 4.80 11.62 mance. Dense Hierarchical Retrieval (DHR) (Liu
et al., 2021) improves passage representations by
Table 2: Ablation study of NodeRAG components.
combining macroscopic document semantics with
microscopic passage details. Expanding on this,
5 Related Works Hybrid Hierarchical Retrieval (HHR) (Arivazha-
gan et al., 2023) fuses sparse and dense retrieval
Retrieval-augmented generation Retrieval- techniques for both document- and passage-level
Augmented Generation (RAG) systems (Gupta retrieval, achieving greater precision. Other meth-
et al., 2024) enhance the performance of large ods leverage hierarchical data structures to facil-
language models (LLM) by retrieving relevant itate complex document summarization. For in-
information from external documents, grounding stance, RAPTOR (Sarthi et al., 2024) employs
responses in domain-specific knowledge. Tradi- tree-based structures to integrate knowledge across
tional RAG approaches (Zhao et al., 2024) embed lengthy documents, synthesizing information at
user queries and entries from a knowledge base various levels of abstraction. Graph-based RAGs
into a shared vector space and then compare query (Trajanoska et al., 2023; Zhang et al., 2024) ex-
vectors to knowledge base vectors to retrieve tend this by constructing knowledge graphs (KGs)
the top-K most similar contexts based on cosine (Chen et al., 2020) at the indexing stage and ap-
similarity or similar variants (Fan et al., 2024; plying graph algorithms during querying (Haveli-
Lewis et al., 2020). While effective, naive RAG wala et al., 2003). Notable examples include Hip-
methods face several limitations, prompting vari- poRAG (Jiménez Gutiérrez et al., 2024) and KAP-
ous enhancements in subsequent works. JPR (Min ING (Baek et al., 2023), which refine knowledge
et al., 2021) improves multi-answer retrieval by organization and retrieval efficiency. Similarly,
refining passage selection, while IR-CoT (Trivedi GraphRAG (Edge et al., 2024) introduces graph-
et al., 2022a) integrates chain-of-thought reasoning based text indexing using LLMs and generates
for multi-hop question answering. Similarly, Tree community-based summaries (Blondel et al., 2008;
of Clarifications (Kim et al., 2023) constructs a Traag et al., 2019), inspiring subsequent works
tree-based disambiguation structure to resolve such as LightRAG (Guo et al., 2024), which inte-
ambiguous queries. HyDE (Gao et al., 2022b) also grates both high- and low-level information while
enhances the performance of dense retrieval by optimizing indexing costs. While these approaches
generating hypothetical documents. Other works effectively leverage hierarchical data structures,
examine how different document types influence they do not fully exploit the synergy between
RAG effectiveness and LLM performance (Hsia LLMs and graph-based methods. Our proposed
et al., 2024). Despite these advancements, tradi- framework addresses these gaps by refining graph
tional RAG systems still face significant challenges. structures and incorporating advanced graph algo-
The context window limitations(Cheng et al., 2024; rithms, leading to superior retrieval accuracy and
Su et al., 2024) of LLMs constrain their ability to efficiency.

9
6 Conclusion and Discussion Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang,
Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing
In this paper, we introduce NodeRAG, a novel Li. 2024. A survey on rag meeting llms: Towards
framework designed to enhance RAG performance retrieval-augmented large language models. In Inter-
by optimizing graph structures in indexing for more national Conference on Knowledge Discovery and
Data Mining (KDD), pages 6491–6501.
effective and fine-grained retrieval. NodeRAG con-
structs a well-defined heterograph with function- Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan.
ally distinct nodes, balancing fine-grained under- 2022a. Precise zero-shot dense retrieval without rele-
standing with a global perspective of the knowl- vance labels. arXiv preprint arXiv:2212.10496.
edge corpus. Experimental results demonstrate that
Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan.
NodeRAG outperforms existing methods across 2022b. Precise zero-shot dense retrieval without rele-
multi-hop reasoning benchmarks and open-ended vance labels. arXiv preprint arXiv:2212.10496.
retrieval tasks. As the saying goes,“A strong foun-
dation supports a higher structure". In the realm of Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao
Huang. 2024. Lightrag: Simple and fast retrieval-
graph-based RAG, the graph structure serves as this
augmented generation. arXiv preprint 2410.05779.
very foundation. The introduction of NodeRAG
underscores the critical role of graph structures, en- Shailja Gupta, Rajesh Ranjan, and Surya Narayan
couraging a renewed emphasis on their design and Singh. 2024. A comprehensive survey of retrieval-
optimization. augmented generation (rag): Evolution, current
landscape and future directions. arXiv preprint
2410.12837.
References Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan
Manoj Ghuhan Arivazhagan, Lan Liu, Peng Qi, Xinchi Ding, Yongjia Lei, Mahantesh Halappanavar, Ryan A
Chen, William Yang Wang, and Zhiheng Huang. Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al.
2023. Hybrid hierarchical retrieval for open-domain 2024a. Retrieval-augmented generation with graphs
question answering. In Findings of the Association (graphrag). arXiv preprint arXiv:2501.00309.
for Computational Linguistics: ACL 2023, pages
10680–10689. Rujun Han, Yuhao Zhang, Peng Qi, Yumo Xu, Jenyuan
Wang, Lan Liu, William Yang Wang, Bonan Min,
Jinheon Baek, Alham Fikri Aji, and Amir Saffari. 2023. and Vittorio Castelli. 2024b. Rag-qa arena: Eval-
Knowledge-augmented language model prompting uating domain robustness for long-form retrieval
for zero-shot knowledge graph question answering. augmented question answering. arXiv preprint
arXiv preprint arXiv:2306.04136. arXiv:2407.13998.

Vincent D Blondel, Jean-Loup Guillaume, Renaud Taher Haveliwala, Sepandar Kamvar, and Glen Jeh.
Lambiotte, and Etienne Lefebvre. 2008. Fast un- 2003. An analytical comparison of approaches to
folding of communities in large networks. Jour- personalizing pagerank. Technical report, Stanford.
nal of statistical mechanics: theory and experiment,
2008(10):P10008. Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla,
Thomas Laurent, Yann LeCun, Xavier Bresson, and
Ulrik Brandes. 2001. A faster algorithm for between- Bryan Hooi. 2024. G-retriever: Retrieval-augmented
ness centrality. Journal of mathematical sociology, generation for textual graph understanding and ques-
25(2):163–177. tion answering. arXiv preprint arXiv:2402.07630.
Zhe Chen, Yuehan Wang, Bin Zhao, Jing Cheng, Xin
Zhao, and Zongtao Duan. 2020. Knowledge graph Jennifer Hsia, Afreen Shaikh, Zhiruo Wang, and Gra-
completion: A review. Ieee Access, 8:192435– ham Neubig. 2024. Ragged: Towards informed
192456. design of retrieval augmented generation systems.
arXiv preprint 2403.09040.
Xin Cheng, Di Luo, Xiuying Chen, Lemao Liu,
Dongyan Zhao, and Rui Yan. 2024. Lift yourself Xinke Jiang, Yue Fang, Rihong Qiu, Haoyu Zhang,
up: Retrieval-augmented text generation with self- Yongxin Xu, Hao Chen, Wentao Zhang, Ruizhe
memory. Advances in Neural Information Processing Zhang, Yuchen Fang, Xu Chu, et al. 2024a. Tc-
Systems, 36. rag: Turing-complete rag’s case study on medical llm
systems. arXiv preprint arXiv:2408.09199.
Darren Edge, Ha Trinh, Newman Cheng, Joshua
Bradley, Alex Chao, Apurva Mody, Steven Truitt, Ziyan Jiang, Xueguang Ma, and Wenhu Chen. 2024b.
and Jonathan Larson. 2024. From local to global: A Longrag: Enhancing retrieval-augmented gener-
graph rag approach to query-focused summarization. ation with long-context llms. arXiv preprint
arXiv preprint 2404.16130. arXiv:2406.15319.

10
Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michi- Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh
hiro Yasunaga, and Yu Su. 2024. Hipporag: Neu- Khanna, Anna Goldie, and Christopher D Man-
robiologically inspired long-term memory for large ning. 2024. Raptor: Recursive abstractive pro-
language models. arXiv preprint arXiv:2405.14831. cessing for tree-organized retrieval. arXiv preprint
arXiv:2401.18059.
Gangwoo Kim, Sungdong Kim, Byeongguk Jeon, Joon-
suk Park, and Jaewoo Kang. 2023. Tree of clarifica- Stephen B Seidman. 1983. Network structure and mini-
tions: Answering ambiguous questions with retrieval- mum degree. Social networks, 5(3):269–287.
augmented large language models. arXiv preprint
arXiv:2310.14696. Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li,
Yusheng Xie, and Siqi Deng. 2024. Fairrag: Fair
Yi-Xiu Kong, Gui-Yuan Shi, Rui-Jie Wu, and Yi-Cheng human generation via fair retrieval augmentation. In
Zhang. 2019. k-core: Theories and applications. Proceedings of the IEEE/CVF Conference on Com-
Physics Reports, 832:1–32. puter Vision and Pattern Recognition, pages 11996–
12005.
Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu,
Petroni, Vladimir Karpukhin, Naman Goyal, Hein- and Yiqun Liu. 2024. Dragin: Dynamic retrieval aug-
rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- mented generation based on the real-time informa-
täschel, et al. 2020. Retrieval-augmented generation tion needs of large language models. arXiv preprint
for knowledge-intensive nlp tasks. Advances in Neu- arXiv:2403.10081.
ral Information Processing Systems, 33:9459–9474.
Yixuan Tang and Yi Yang. 2024. Multihop-rag: Bench-
Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paran- marking retrieval-augmented generation for multi-
jape, Michele Bevilacqua, Fabio Petroni, and Percy hop queries. arXiv preprint arXiv:2401.15391.
Liang. 2024. Lost in the middle: How language mod-
els use long contexts. Transactions of the Association Yijun Tian, Huan Song, Zichen Wang, Haozhu Wang,
for Computational Linguistics, 12:157–173. Ziqing Hu, Fang Wang, Nitesh V Chawla, and Pan-
pan Xu. 2024. Graph neural prompting with large
Ye Liu, Kazuma Hashimoto, Yingbo Zhou, Semih language models. In Proceedings of the AAAI Con-
Yavuz, Caiming Xiong, and Philip S Yu. 2021. Dense ference on Artificial Intelligence, volume 38, pages
hierarchical retrieval for open-domain question an- 19080–19088.
swering. arXiv preprint arXiv:2110.15439.
SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vip-
Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and ula Rawte, Aman Chadha, and Amitava Das. 2024.
Shirui Pan. 2023. Reasoning on graphs: Faithful and A comprehensive survey of hallucination mitigation
interpretable large language model reasoning. arXiv techniques in large language models. arXiv preprint
preprint arXiv:2310.01061. arXiv:2401.01313.

James MacQueen et al. 1967. Some methods for clas- Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck.
sification and analysis of multivariate observations. 2019. From louvain to leiden: guaranteeing well-
In Proceedings of the fifth Berkeley symposium on connected communities. Scientific reports, 9(1):1–
mathematical statistics and probability, volume 1, 12.
pages 281–297. Oakland, CA, USA.
Milena Trajanoska, Riste Stojanov, and Dimitar Tra-
Yu A Malkov and Dmitry A Yashunin. 2018. Efficient janov. 2023. Enhancing knowledge graph construc-
and robust approximate nearest neighbor search us- tion using large language models. arXiv preprint
ing hierarchical navigable small world graphs. IEEE arXiv:2305.04676.
transactions on pattern analysis and machine intelli- Harsh Trivedi, Niranjan Balasubramanian, Tushar
gence, 42(4):824–836. Khot, and Ashish Sabharwal. 2022a. Interleav-
ing retrieval with chain-of-thought reasoning for
Sewon Min, Kenton Lee, Ming-Wei Chang, Kristina knowledge-intensive multi-step questions. arXiv
Toutanova, and Hannaneh Hajishirzi. 2021. Joint preprint arXiv:2212.10509.
passage ranking for diverse multi-answer retrieval.
arXiv preprint arXiv:2104.08445. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot,
and Ashish Sabharwal. 2022b. Musique: Multi-
Jinyoung Park, Ameen Patel, Omar Zia Khan, Hyun- hop questions via single-hop question composition.
woo J Kim, and Joo-Kyung Kim. 2023. Graph- Transactions of the Association for Computational
guided reasoning for multi-hop question answer- Linguistics, 10:539–554.
ing in large language models. arXiv preprint
arXiv:2311.09762. Chengrui Wang, Qingqing Long, Meng Xiao, Xunxin
Cai, Chengjun Wu, Zhen Meng, Xuezhi Wang,
Diego Sanmartin. 2024. Kg-rag: Bridging the gap and Yuanchun Zhou. 2024a. Biorag: A rag-llm
between knowledge and creativity. arXiv preprint framework for biological question reasoning. arXiv
arXiv:2405.12035. preprint arXiv:2408.01107.

11
Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi
Zhang, and Tyler Derr. 2024b. Knowledge graph
prompting for multi-document question answering.
In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 38, pages 19206–19214.
Yilin Wen, Zifeng Wang, and Jimeng Sun. 2023.
Mindmap: Knowledge graph prompting sparks graph
of thoughts in large language models. arXiv preprint
arXiv:2308.09729.
Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min
Xu, Filippo Menolascina, and Vicente Grau. 2024.
Medical graph rag: Towards safe medical large lan-
guage model via graph retrieval-augmented genera-
tion. arXiv preprint arXiv:2408.04187.
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Ben-
gio, William W Cohen, Ruslan Salakhutdinov, and
Christopher D Manning. 2018. Hotpotqa: A dataset
for diverse, explainable multi-hop question answer-
ing. arXiv preprint arXiv:1809.09600.
Yuzhe Zhang, Yipeng Zhang, Yidong Gan, Lina Yao,
and Chen Wang. 2024. Causal graph discovery with
retrieval-augmented generation based large language
models. arXiv preprint arXiv:2402.15301.
Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren
Wang, Yunteng Geng, Fangcheng Fu, Ling Yang,
Wentao Zhang, and Bin Cui. 2024. Retrieval-
augmented generation for ai-generated content: A
survey. arXiv preprint arXiv:2402.19473.

12
A Comparison of RAG System Performance

Datasets Corpus Size Index Time Storage Usage Query Time Average Retrieval Tokens
Graph Light Node Graph Light Node Graph-L Graph-G Light Node Graph-L Graph-G Light Node
HotpotQA 1.93M 66min 39min 21min 227MB 461MB 214MB 2.66s 26.69s 5.58s 3.98s 6680.65 810529 7176.73 5079.40
Musique 1.84M 76min 90min 25min 255MB 492MB 250MB 2.94s 22.65s 6.53s 4.05s 6616.84 1111073 7458.34 5960.25
MultiHop 1.41M 50min 58min 24min 141MB 276MB 137MB 4.15s 34.45s 7.10s 4.89s 7367.54 920780 8920.00 5259.99
Arena-Fiqa 1.65M 45min 49min 19min 112MB 240MB 117MB 8.95s 28.94s 13.35s 8.86s 6819.45 713560 7721.73 3381.72
Arena-Lifestyle 1.64M 52min 59min 18min 138MB 278MB 125MB 7.54s 33.09s 10.43s 6.79s 6860.26 895964 6822.32 3350.35
Arena-Recreation 0.93M 34min 33min 10min 89MB 172MB 80MB 5.10s 23.10s 8.01s 6.90s 6669.95 564636 6249.31 3448.38
Arena-Science 1.43M 43min 46min 17min 116MB 236MB 111MB 8.05s 35.79s 14.28s 8.85s 6759.15 778051 7111.80 4284.13
Arena-Tech 1.72M 54min 54min 14min 133MB 276MB 139MB 7.35s 28.64s 8.89s 6.74s 6755.46 741690 6922.55 3821.78
Arena-Writing 1.82M 50min 71min 13min 151MB 309MB 157MB 5.65s 40.12s 10.70s 5.40s 6477.72 877354 6364.59 3373.34

Table 3: Performance metrics for RAG methods, including Index Time, Storage Usage, Query Time, and Retrieval
Tokens across various datasets. Graph denotes GraphRAG, with Graph-l representing its local mode and Graph-G
its global mode. Light refers to LightRAG in hybrid mode, while Node represents our proposed method.

The table 3 presents the system performance of mainstream graph-based RAG methods and our proposed
approach. Compared to previous work, our method demonstrates superior performance across multiple
datasets and in open-ended head-to-head evaluations, while also achieving better system-level efficiency.
All evaluations in the table were conducted using the default indexing settings of each RAG method, with
the query settings and the prompt details provided in Appendix B.2. Notably, our method demonstrates
a significant advantage in indexing time, which is crucial for practical deployment. This advantage is
attributed to the construction process of our Hetero Graph, which not only creates a more fine-grained and
semantically meaningful graph structure but also carefully considers the algorithmic complexity of the
retrieval process.
NodeRAG also exhibits relatively better storage efficiency. Although the total number of nodes in our
expanded graph is significantly larger than in previous graph structures, the combination of selective
embedding and dual search effectively reduces the number of embedded nodes, leading to a more efficient
storage strategy. Moreover, our unified information retrieval approach results in reduced query time.
While the GraphRAG local search (Graph-l) relies purely on vector similarity—similar to our "without
cross-node" setting mentioned in Section 4—and achieves faster search speeds, its global mode (Graph-G)
experiences significantly higher query times, exceeding 20 seconds with a concurrency of 16. This is due
to its reliance on LLM-based traversal of all community information, leading to a substantial number
of retrieval tokens. Given the considerable time and computational overhead associated with Graph-G
queries, we conducted a full evaluation only on the MuSiQue dataset. For other datasets, query time and
retrieval token statistics were estimated based on a sample of 20 selected queries. Further details on the
ablation study of GraphRAG can be found in the Appendix B.4.
In contrast, our method leverages the heterograph and graph algorithms to achieve unified information
retrieval, effectively capturing meaningful information needs across multiple levels within a single
framework while maintaining efficient query speed. Finally, the nodes within the heterograph are
connected in a fine-grained structure, ensuring that more relevant text is retrieved with relatively fewer
retrieval tokens.

B Experiment details

B.1 Datasets

We evaluate Node RAG’s performance across four different benchmarks: HotpotQA, MuSiQue, MultiHop-
RAG and RAG-QA Arena. However, the original question formats of HotpotQA and MuSiQue required
selecting the most relevant passages from multiple documents, incorporating multi-hop reasoning details.
This setup no longer aligns with mainstream RAG methods, as modern approaches perform indexing over
an entire corpus and subsequently retrieve information from the indexed data. To adapt to this paradigm,
we concatenate all passages into a unified corpus, transforming the task into retrieving multi-hop relevant
information from the entire corpus. This modification makes the task more challenging compared to the

13
original setting. The evaluation metrics for HotpotQA and MuSiQue are divided into two aspects: the
quality of the retrieved documents and the accuracy of the final answer, measured by metrics such as
F1 score. However, current RAG methods retrieve not only text chunks but also more flexible forms of
information, making it difficult to assess retrieval quality using traditional top-k document evaluation.
Moreover, metrics like F1 score have become less effective in evaluating answers generated by modern
generative models. Therefore, we adopt the LLM-as-a-Judge approach, leveraging LLMs to assess the
final accuracy of the generated [Link] MultiHop and RAG-QA Arena dataset settings provide a
strong evaluation framework for current RAG methods. Therefore, we follow the original benchmark’s
proposed testing methodology and evaluation metrics. Further details regarding the benchmark settings
are described below.
HotpotQA is a multi-hop question-answering dataset where each question requires combining information
from multiple documents to find the correct answer. It encourages deeper reasoning by providing
supporting facts—specific sentences from the texts that lead to the solution. Questions range widely
across domains and often involve bridging or comparison to ensure more complex, multi-step reasoning.
This makes HotpotQA a critical benchmark for evaluating advanced reading comprehension models. We
sampled 200 questions from the final dataset for evaluation.
MuSiQue is also a multi-hop question-answering dataset that challenges models to combine information
across multiple documents in a structured, step-by-step manner. Each question is designed to require
several reasoning steps, ensuring that simple “shortcut” approaches do not suffice. As a result, MuSiQue
serves as a rigorous test of advanced reading comprehension, demanding that systems accurately connect
disparate pieces of evidence to arrive at correct answers. We also sample 175 questions for the evaluation
MultiHop-RAG is a multi-hop question-answering dataset that includes four distinct question types:
comparison query, null query, inference query, and temporal query. From this dataset, we curated 375
questions to evaluate our approach. Each query in MultiHop requires synthesizing information from
multiple sources, testing a model’s ability to perform bridging inferences, handle temporal relationships,
and make higher-order logical connections. This diversity in question types provides a rigorous benchmark
for assessing whether RAG methods can integrate scattered pieces of evidence.
RAG-QA Arena is a new evaluation framework designed to assess the quality of retrieval-augmented
generation (RAG) systems on long-form question answering. It builds on Long-form RobustQA (LFRQA),
a dataset of 26K queries across seven domains including writing, tech, science, recreation and lifestyle.
Each LFRQA entry features a coherent, human-written answer grounded in multiple documents. RAG-
QA Arena leverages LLMs as evaluators, directly comparing a system’s generated answer with the
’gold’ long-form answer from LFRQA. Experimental results show that these model-based comparisons
correlate highly with human judgments, making it a challenging yet reliable benchmark for testing both
cross-domain robustness and the ability to produce integrated, long-form responses.

B.2 Baselines

We compare NodeRAG against several strong and widely used RAG methods. By default, all these
RAG methods implement their indexing process using GPT-4o-mini. However, we identify a potential
unfairness in the current evaluation setup, particularly in several key areas. To ensure the correctness
and validity of the evaluation data, it is crucial to standardize both the final answer response prompt
and the model temperature settings. Using different response prompts or varying temperature settings
for answer generation introduces inconsistencies, as a higher temperature setting may yield responses
that receive a better LLM preference score compared to those generated with a lower temperature. A
critical point to consider is that, as RAG methods, the primary focus of evaluation should be the quality
of the retrieved context rather than the final generated answer. Therefore, to ensure that final accuracy
metrics accurately reflect the quality of the retrieved context, the final answer generation process and
model settings should remain consistent across all methods. Hence, we set the temperature to 0 across the
entire evaluation and standardized response prompts for every method. The unified prompt is illustrated

14
in appendix E. Our initiative to standardize these settings also benefits other methods, such as GraphRAG,
improving their performance compared to their default settings. This underscores the broader value
of establishing fair and consistent evaluation standards. Additionally, traditional evaluation methods
such as top-k retrieval comparison have become increasingly difficult to apply uniformly, as retrieval
is no longer restricted to isolated text chunks. To address this challenge, we propose a new evaluation
standard that leverages retrieval tokens as an efficiency metric. This approach ensures that retrieval
methods achieve better effectiveness while utilizing fewer retrieval tokens, promoting a more efficient and
fair comparison framework. Current methods can only control the number of retrieval tokens through
hyperparameter tuning. Although precise control over the exact number of tokens is not possible, we
consider maintaining the average number of retrieval tokens within the range of 5K to 10K to be a
reasonable and fair comparison criterion. Below, we provide a detailed introduction to each method along
with its specific settings for reference.

Naive RAG This method serves as a standard baseline among all existing RAG systems. It first divided
input document into several text chunks and encoded them into a vector space utilizing text embeddings.
Then retrieve related text chunks based on similarity of query representations. The number of retrieval
tokens can be adjusted through the top-k parameter.

HyDE HyDE serves as an improved method over traditional RAG systems. It first generates "hypotheti-
cal" texts that capture the essence of a query. It then uses this generated text to retrieve relevant documents
from a large corpus, employing vector similarity in an embedding space. This method modifies the input
query at the frontend without altering the text chunks or their embeddings. Therefore, we can still use the
top-k parameter to control the number of retrieval tokens.

GraphRAG This approach starts by segmenting the input text into chunks and extracting the entities
and relationships within them, forming a graph structure. This graph is then divided into multiple com-
munities at different levels. At query time, GraphRAG identifies the relevant entities from the question
and synthesizes answers by referencing these corresponding community summaries. Compared to tradi-
tional RAG methods, GraphRAG provides a more structured and high-level understanding of the entire
document. Through our experiments, we observed that under the default settings, the number of queries
in GraphRAG’s local mode resulted in a higher retrieval token count than the naive retrieval approach.
To ensure a fair comparison, we proportionally reduced its parameters and standardized its prompt to
match our unified prompt. The ablation study in Appendix B.4 demonstrates that after these adjustments,
GraphRAG’s accuracy improved, further validating the fairness of our evaluation methodology. Addition-
ally, we analyzed both the local and global modes of GraphRAG. Our findings indicate that the global
mode introduces significant additional overhead in terms of time and computational cost while providing
only marginal improvements compared to the local mode. This result is further supported by our ablation
study, which shows that the local mode achieves better efficiency and effectiveness.

LightRAG LightRAG is an improved approach based on GraphRAG, designed to minimize compu-

tational overhead while enhancing the comprehensiveness of retrieved information through dual-level
retrieval. This leads to more efficient retrieval and a better balance between effectiveness and speed
compared to GraphRAG. Similar to GraphRAG, the default settings of LightRAG result in a higher
retrieval token count than the Naïve approach. To ensure a fair comparison, we proportionally adjusted its
hyperparameters to maintain the number of retrieval tokens within the range of 5K to 10K.

B.3 NodeRAG Graph Statistics

The table 4 presents the number of each type of node in the indexed graph for each dataset, including
entity (N ), relationship (R), semantic unit (S), attribute (A), high-level elements (H), high-level overview
(O), and text (T ). These counts are detailed in the type statistics section. Additionally, the graph statistics
provide information on the total number of nodes, the number of non-HNSW edges, HNSW edges, and

15
 Datasets Corpus Tokens Type Statistics Graph Statistics
T S N R A O H Nodes Non-HNSW Edge HNSW Edge Edge
HotpotQA 1.93M 1985 15905 88863 56578 684 4479 4479 172603 283543 487731 759812
MuSiQue 1.84M 1907 18714 99840 61964 795 5700 5700 193922 316029 583126 888966
MultiHop-RAG 1.41M 1532 10986 43184 29286 685 2289 2289 90144 171410 203199 367486
Arena-Fiqa 1.65M 1821 9027 32470 27422 508 1714 1714 74605 143916 154109 295165
Arena-Lifestyle 1.64M 1794 9400 39464 27895 518 2221 2221 83461 149225 174461 318073
Arena-Recreation 0.93M 1003 5542 26382 16938 413 1969 1969 54180 93228 117915 207449
Arena-Science 1.43M 1583 8010 32232 23092 551 2515 2515 70425 127719 149424 276963
Arena-Tech 1.72M 1910 10837 37724 29691 534 2633 2633 85888 167950 193159 354033
Arena-Writing 1.82M 1937 11008 42723 29338 705 4435 4435 94259 149552 298565 442397

Table 4: Comprehensive dataset statistics, detailing corpus size, type statistics (T, S, N, R, A, O, H), and graph
statistics. The graph statistics include the number of document compilation nodes, HNSW semantic edges, and total
edges. Each value represents a key metric relevant to graph-based document processing and retrieval.

the total number of edges. The data indicate that the number of HNSW edges is comparable to that of non-
HNSW edges, highlighting the integration of semantic connections within the graph. Notably, overlapping
edges are removed when merging non-HNSW and HNSW edges. For instance, in the MultiHop-RAG
benchmark, there are 171,410 non-HNSW edges and 203,199 HNSW edges. However, the total number of
edges after merging is 367,486, which is only 7,123 fewer than the sum of both edge types. This indicates
the uniqueness of these two types of edges and highlights the effectiveness of the HNSW algorithm.

B.4 Graph RAG Ablation

Method Accuracy Avg. Processing Time Avg. Tokens

GraphRAG (default) 37.14% 4.82s 10.4k
Graph-L 41.71% 2.94s 6.6k
Graph-G 33.14% 22.65s 1.11M

Table 5: Performance Comparison of GraphRAG Variants. Default is the default setting. Local and global represent
the local and global modes under unified prompt and hyperparameter settings.

The default setting of GraphRAG, along with its own prompting mechanism, is not standardized for
evaluation, as both the number of retrieval tokens and the choice of prompts significantly impact perfor-
mance. Hence, we introduce a unified prompt and adjust the hyperparameters of GraphRAG to ensure a
fair comparison within a specific range. As shown in the table B.4, GraphRAG with our unified prompt
achieves higher performance, demonstrating that the original prompting strategy is not optimal for this
task. This further ensures fairness in comparison, as performance is influenced solely by the quality of the
retrieved context. Moreover, the global mode of GraphRAG requires significantly longer processing time
and incurs higher computational costs due to the LLM analyzing all community summaries, leading to
increased complexity and resource consumption. Additionally, for multi-hop question answering, this
approach results in degraded performance. Therefore, we conducted an exploratory ablation study only
on the MuSiQue dataset, while for other datasets, we estimated query time and retrieval token statistics
based on sampled queries.

16
C Algorithm details

C.1 Terminology

Abbr. Full Name Description Function Example

"Hinton was awarded the No-
bel Prize in 2023 for his ground-
breaking contributions to artifi-
cial intelligence, particularly in
deep learning. His pioneering
Full-text chunks from the
work on backpropagation laid
original source. It contains
Retrievable; the foundation for modern neu-
rich detailed information, al-
T Text Entry points from vec- ral networks, influencing both
though it integrates a large
tor similarity academia and industry. The
amount of unrelated semantic
recognition came amid increas-
information.
ing discussions on the ethical
implications of AI, with Hinton
himself advocating for responsi-
ble AI development and regula-
tion."
Local summaries that are in-
dependent and meaningful
events summarized from text
Retrievable; "Hinton was awarded the Nobel
chunks. They serve as a mid-
S Semantic Unit Entry points from vec- Prize for inventing backpropa-
dle layer between text chunks
tor similarity. gation."
and entities, acting as the ba-
sic units for graph augmenta-
tion and semantic analysis.
"Geoffrey Hinton, often referred
to as the "Godfather of Deep
Attributes of key entities, de-
Retrievable; Learning," is a pioneer in the
rived from relationships and
A Attribute Entry points from vec- field of artificial intelligence. In
semantic units around impor-
tor similarity. 2024, he was awarded the No-
tant entities.
bel Prize for his contributions
to AI and deep learning. "
Insights summarizing graph "Due to the increasing impor-
communities. Encapsulates Retrievable; tance of AI, the Nobel Prize is
H High-Level Element core information or any high Entry points from vec- awarded to scholars who have
level ideas from a community. tor similarity. made tremendous contributions
to the field of AI."
Titles or keywords summariz- Non-Retrievable;
O High-Level Overview ing Entry points from accu- "AI significance"
high-level elements. rate search.
Connections between entities
represented as nodes. Acts Retrievable; "Hinton received the Nobel
R Relationship
as connector nodes and sec- Non-Entry points Prize."
ondary retrievable node.
Non-Retrievable;
Named entities such as peo-
N Entity Entry points from accu- "Hinton," "Nobel Prize"
ple, places, or concepts.
rate search..

Table 6: Node Types in the heterograph

C.2 K-core & Betweenness centrality

In this subsection, we present the methodology for identifying important entities and generating their
attribute summaries, ensuring alignment with the mathematical framework established in the main text.
The selection of important entities, denoted as N ∗ , is based on two fundamental structural graph metrics:
K-core decomposition and betweenness centrality. These metrics collectively ensure that the selected

17
nodes are not only structurally integral but also play a pivotal role in facilitating information flow.
The K-core decomposition, denoted as K(G 1 ), identifies nodes within densely connected subgraphs,
ensuring that selected entities contribute significantly to the structural cohesion of the graph. Meanwhile,
betweenness centrality, denoted as B(G 1 ), highlights nodes that serve as critical intermediaries between
different regions of the graph, capturing entities essential for information dissemination.
The process of identifying important entities follows the steps outlined in Algorithm 1.

Algorithm 1 Identification of Important Entities

Input: Graph G 1 = (V, E)
Output: Important entity set N ∗
Step 1: Compute K-core decomposition
Compute the core threshold:
P 1/2
v∈Vdeg(v)
kdefault = ⌊log(|V|) × ⌋
|V|

Extract the K-core subgraph:

K(G 1 ) = {v ∈ V | degG 1 (v) ≥ kdefault }

Step 2: Compute betweenness centrality

for each v ∈ V do
Approximate betweenness centrality using shortest-path sampling:

b(v) = betweenness_centrality(G 1 , k = 10)

end for
Compute the average betweenness centrality:
P
v∈V b(v)
b̄ =
|V|

Compute the scale factor:

scale = ⌊log10 (|V|)⌋
Step 3: Select important nodes
for each v ∈ V do
if b(v) > b̄ × scale then
Add v to B(G 1 )
end if
end for
Compute the final set of important entities:

N ∗ = K(G 1 ) ∪ B(G 1 )

Return N ∗

C.3 Semantic Matching within Community

To establish meaningful semantic relationships among high-level element nodes, we propose the Seman-
tic Matching within Community algorithm. This algorithm ensures that entities with strong semantic
similarities are connected within their respective communities. The motivation behind this approach is

18
Algorithm 2 Semantic Matching within Community
Input: Graph G = (V, E), node embeddings Φ(V), community partition {Cn }
Output: Semantic edges Eh
Step 1: Select high-level element nodes
Extract nodes with labels S, A, or H:

V{S,A,H} = {v ∈ V | ψ(v) ∈ {S, A, H}}

Step 2: Apply K-means clustering to node embeddings

Set number of clusters: q
K = |V{S,A,H} |

Perform K-means clustering on V{S,A,H} ), obtaining clusters {Sk }

Step 3: Establish semantic edges within communities
for each community Cn do
for each cluster Sk do
Identify nodes within the community and cluster:

VCn ,Sk = V{S,A,H} ∩ Cn ∩ Sk

for each pair (v, v ′ ) where v ∈ {S, A}, v ′ ∈ H do

Add semantic edge:
eh (v, v ′ ) ∈ Eh
end for
end for
end for
Return Eh

to organically integrate H nodes into the graph structure by establishing connections with semantically
related nodes within the same community. Formally, the process is summarized in Algorithm 2.
The algorithm begins by identifying nodes that belong to three specific categories: structure nodes (S),
attribute nodes (A), and high-level nodes (H). These nodes are collectively defined as:

V{S,A,H} = {v ∈ V | ψ(v) ∈ {S, A, H}}

Since these nodes exhibit inherent semantic relationships, we cluster them based on their embeddings,
which capture their contextual meaning. To partition the nodes into semantically similar groups, we apply
the K-means clustering algorithm (MacQueen et al., 1967) to the embedding representations of V{S,A,H} .
which balances computational efficiency and granularity. This clustering process results in a partitioning
of nodes into K semantic clusters, denoted as Sk , where each cluster contains nodes with closely related
semantic representations.
After clustering, the algorithm establishes edges between semantically related nodes within the same
community. Communities are predefined structural subgroups in the graph, denoted as Cn , ensuring
that local relationships are preserved. For each community-cluster pair, semantic edges are introduced
between nodes in V{S,A} and nodes in VH . Specifically, for any node pair (v, v ′ ), where v ∈ V{S,A} and
v ′ ∈ VH , an edge eh (v, v ′ ) is established if both nodes belong to the same community and the same
semantic cluster.

19
By integrating semantic matching within community constraints, this algorithm enhances the structural
integrity of the graph while maintaining computational feasibility. The choice of K-means clustering
efficiently groups nodes with similar semantic properties, while the enforcement of community constraints
ensures that edges are only formed between nodes that naturally belong to the same substructure. Conse-
quently, the proposed method balances semantic consistency and graph locality, making it well-suited for
applications requiring structured knowledge representation and retrieval.

C.4 Dual Search

To efficiently locate relevant entry points within the Hetero Graph G, we propose the Dual Search algorithm,
which integrates exact matching on structured nodes and vector similarity search on rich information
nodes. This hybrid approach ensures a balance between precision and recall by leveraging both symbolic
and dense representations. The core idea is to utilize exact string matching for well-structured nodes while
employing approximate nearest neighbor search for nodes containing rich contextual information. By
doing so, the algorithm improves both retrieval accuracy and robustness to query variations.
Given a query, a LLM extracts a set of relevant entities, denoted as N q , and embeds the query into a
vector representation q. Entry points in the graph are then determined by:

Ventry = {v ∈ V | Φ(v, N q , q)},

where the condition function Φ(v, N q , q) determines whether a node qualifies as an entry point:
(
v ∈ V{N,O} ∧ M(N q , v),
Φ(v, N q , q) =
v ∈ V{S,A,H} ∧ R(q, v, k).

Here, the exact matching function M(N q , v) returns true if node v matches one of the extracted entities
in N q . This ensures that titles or named nodes such VN,O are retrieved deterministically. Meanwhile,
the similarity-ranking function R(q, v, k) applies HNSW, selecting the top-k most similar nodes to q in
vector space. This is particularly beneficial for context-rich nodes for v ∈ VS,A,H , which may not contain
exact query terms but remain semantically relevant.
One of the major strengths of Dual Search is its ability to mitigate errors from noisy or ambiguous queries.
Since N and O nodes serve exclusively as entry points and are not directly retrieved, any errors in exact
matching do not propagate into the retrieval results. Furthermore, Personalized PageRank (PPR) is applied
to refine retrieval by selecting highly weighted cross nodes—nodes that are both structurally significant
and semantically aligned with multiple entry points. This graph-based expansion helps compensate for
potential errors in entity extraction and ensures robustness against query variations. Another advantage of
this approach is its scalability. By employing HNSW, the algorithm avoids the computational overhead of
exhaustive similarity searches, making it efficient even on large-scale graphs.

20
D Case Study

Question1 Why isn’t there public transport on Christmas day in the UK?
The absence of public transport on Christmas Day in the UK can be attributed to a combi-
nation of historical, cultural, and economic factors. Historically, public transport services
were more prevalent on Christmas Day before World War II, as many people worked on
that day and required transportation. However, over the years, societal norms have shifted,
leading to a significant reduction in demand for public transport services on Christmas Day.
In contemporary society, Christmas is widely recognized as a time for family gatherings
and celebrations, which diminishes the need for public transport. Many people choose to
NodeRAG spend the day at home or with relatives, resulting in a lower demand for transit services.
Consequently, transport companies have adapted to this change by suspending services on
Christmas Day, allowing transit workers to enjoy the holiday with their families as well.
Additionally, the operational costs associated with running public transport services on
Christmas Day can be significant. With minimal passenger demand, the financial viability
of providing such services becomes questionable. Railways and other transport providers
often use this time for maintenance and engineering work, taking advantage of the reduced
disruption to carry out necessary improvements.
Public transport in the UK is significantly limited on Christmas Day due to a combination
of historical, cultural, and operational factors. Traditionally, Christmas Day is a public
holiday in the UK, and many businesses, including public transport services, close to allow
employees to spend time with their families and celebrate the holiday. This practice is
rooted in the cultural significance of Christmas, which emphasizes family gatherings and
community celebrations.
The limited operation of public transport on Christmas Day is also influenced by the low
demand for services. Historically, fewer people travel on this day compared to other holidays
GraphRAG or weekends, as many individuals prefer to stay home or visit family. As a result, transport
providers often find it economically unfeasible to run full services when the number of
passengers is expected to be minimal.
In recent years, some transport services have begun to offer limited services on Christmas
Day, particularly in major cities like London, where there may be a demand for travel to
events or gatherings. However, these services are typically not as extensive as those available
on other public holidays or regular days. Overall, the combination of cultural practices,
low demand, and operational considerations contributes to the lack of public transport on
Christmas Day in the UK.
In most parts of the world, official or religious holidays are taken very seriously. On such
days, employers typically have to pay higher rates to staff who are willing to work, but that
effort often doesn’t pay off due to low customer turnout. Families tend to arrange their travel
plans by Christmas Eve to avoid the need for transportation on Christmas Day. Christmas
avoids unnecessary travel because it’s typically a day filled with home activities, and it
LightRAG allows workers in various roles to have a day off. The lack of public demand and the staff’s
resistance contributed to the lack of services on Christmas Day. Christmas Day is also an
opportune time to perform engineering work on railways with minimal disruption. The
railroad industry is capable of completing extensive projects during this period, thanks to
thorough planning and pre-staging.
ORDER: {“1": “GraphRAG", “2": “NodeRAG"}
Both answers provide a comprehensive explanation, but answer 2 adds historical context
about public transport before World War II and mentions the use of the holiday for mainte-
LLM nance work, offering slightly more depth.
Decision ORDER: {“1": “LightRAG", “2": “NodeRAG"}
Both answers provide a comprehensive explanation, but answer 2 adds historical context
about the change in public transport services since World War II, which enriches the
explanation. Therefore, answer 2 is slightly better.
Table 7: Sample Table with Multi-line Text

21
Question2 What are rabbit hides used for in Minecraft?
In Minecraft, rabbit hides serve a couple of useful purposes, particularly in crafting and
resource management. One of the primary uses of rabbit hides is to craft leather. Players
can combine four rabbit hides in a 2x2 crafting grid to create one piece of leather. This is
especially beneficial in Survival Multiplayer (SMP) scenarios where finding other animals
for leather can be challenging.
NodeRAG
Additionally, rabbits can drop a unique item known as the “rabbit’s foot", which is an
important ingredient for brewing potions. The rabbit’s foot is specifically used in the crafting
of the Potion of Leaping, which grants players the ability to jump higher for a limited
time. Thus, rabbit hides and their associated drops contribute significantly to crafting and
potion-making in the game, enhancing the overall gameplay experience.

In Minecraft, rabbit hides are primarily used to craft leather. Players can combine four rabbit
hides in a 2x2 crafting grid to create one piece of leather. This crafting option is particularly
useful for players who may not have access to cows, which are the traditional source of
leather in the game.
GraphRAG
Additionally, rabbit hides can be beneficial in multiplayer servers that utilize entity clears, as
finding animals can be challenging. By using rabbit hides to craft leather, players have an
alternative method to obtain this valuable resource, which is essential for crafting various
items, including armor and books.

In Minecraft, Rabbit Hides are primarily used as a crafting resource. They can be combined
with other materials to create leather, which is an essential item in the game. Leather is used
for crafting various items, including armor (like leather armor), books, item frames, and
more.
LightRAG
The process of obtaining Rabbit Hides involves hunting rabbits, which can be found in
various biomes. Once collected, players can use the hides in their crafting recipes to enhance
their gameplay experience through resource management and item creation. This highlights
the importance of Rabbit Hides within the broader crafting mechanics of Minecraft.

ORDER: {“1": “TGRAG", “2": “GraphRAG"}

Both answers are truthful, but answer 1 provides more detailed information about the uses
of rabbit hides, including the additional use of rabbit’s foot for potion-making Therefore,
LLM answer 1 is better.
Decision ORDER: {“1": “TGRAG", “2": “LightRAG"}
Both answers are truthful, but answer 1 provides more detailed information about the uses
of rabbit hides, including the crafting of leather and the rabbit’s foot for potions. Therefore,
answer 1 is better.

Table 8: Sample Table with Multi-line Text

22
E Overview Prompting in NodeRAG

23
24
25
26