LivingRAG: Augmenting Graph RAG with Experience

Yuzhuo Cui; Zongye Zhang; Qingjie Liu

cs.AIAug 26, 2026

LivingRAG: Augmenting Graph RAG with Experience

Authors: Yuzhuo Cui, Zongye Zhang, Qingjie Liu

Organizations: State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · Hangzhou Innovation Institute, Beihang University, Hangzhou, Zhejiang, China

Abstract

Graph-based RAG improves multi-hop question answering by organizing evidence as a knowledge graph. However, most existing RAG systems process each query in isolation and discard useful reasoning from the LLM's response after inference. As a result, later related queries need to retrieve evidence and reason from scratch. We propose LivingRAG, a Graph RAG framework with writable and reusable reasoning experience. LivingRAG adds a writable experience store to a graph-based retrieval backbone, enabling verified experiences to be reused during inference in two ways. Stored graph signals help retrieval find entities and passages that were useful in earlier related queries. Stored summaries provide a reference reasoning pattern for answer generation. We analyze online QA streams and find reusable signals from shared entities, graph neighborhoods, and question templates. Experiments on multi-hop QA benchmarks show that LivingRAG improves accuracy over strong RAG baselines and reduces completion-token use when relevant prior experience is reused.

Figures & tables

Dataset	Entity reuse	Graph-neigh. reuse	Template reuse
2Wiki	4.30	99.70	96.30
HotpotQA	26.53	99.10	21.82
MuSiQue	40.34	95.60	39.14
MuSiQue-full	54.10	96.32	54.30
WixQA	76.69	80.45	16.04

Table 1: Reuse signals in online QA streams. Only earlier queries are used as reuse sources. All values are percentages.

Figure 1: Overview of LivingRAG. The framework first constructs a static retrieval graph offline. It then performs experience-augmented retrieval, scaffolded generation, and write-back through quality gates.

Method	2Wiki		HotpotQA		MuSiQue		MuSiQue-full		WixQA
Method	Cont.-Acc.	LLM-Acc.	Cont.-Acc.	LLM-Acc.	Cont.-Acc.	LLM-Acc.	Cont.-Acc.	LLM-Acc.	LLM-Acc.
Vanilla RAG (Top-5) ( Karpukhin et al., 2020 )	58.80	59.40	65.20	77.10	34.00	43.70	27.35	41.33	–
LightRAG ( Guo et al., 2024 )	60.60	48.20	62.20	71.70	39.20	49.10	29.83	44.89	–
GFM-RAG ( Luo et al., 2026b )	75.50	77.80	72.50	86.10	42.70	52.50	37.28	47.75	–
HippoRAG2 ( Gutiérrez et al., 2025 )	76.50	68.00	71.20	84.60	45.40	52.90	40.92	44.06	–
LinearRAG ( Zhuang et al., 2025 )	79.60	84.40	72.90	86.80	48.50	60.70	41.00	52.71	66.00

Table 2: Overall accuracy (%) on QA benchmarks. Cont.-Acc. denotes Contain-Match Accuracy. WixQA is reported only for LinearRAG and LivingRAG; unavailable entries are marked with –. The best result is marked in bold , and the second-best result is underlined.

Dataset	# Queries	Prompt tokens			Completion tokens			Run cost (USD)
Dataset	# Queries	Linear	Living	Rel. $\Delta$	Linear	Living	Rel. $\Delta$	Linear	Living	Rel. $\Delta$
2WikiMultiHopQA	1,000	6,038.2	6,241.6	+3.4%	1,067.6	906.4	-15.1%	$6.22	$5.84	-6.1%
HotpotQA	1,000	5,779.8	5,985.8	+3.6%	941.9	767.0	-18.6%	$5.72	$5.29	-7.4%
MuSiQue	1,000	5,888.4	6,092.4	+3.5%	2,098.8	1,641.7	-21.8%	$9.24	$7.97	-13.7%
MuSiQue-full	2,417	8,639.3	8,861.0	+2.6%	2,223.8	1,643.7	-26.1%	$26.57	$22.63	-14.8%
WixQA	400	6,633.0	7,412.8	+11.8%	1,653.8	1,378.1	-16.7%	$3.31	$3.14	-5.3%

Table 3: Token consumption and cost of LinearRAG and LivingRAG. Prompt and completion tokens are per-query averages. Rel.

\Delta

columns report

(\mathrm{LivingRAG}-\mathrm{LinearRAG})/\mathrm{LinearRAG}

. Negative values indicate reductions. Run cost is computed with Qwen3.6 Plus pricing:

0.50 per million prompt tokens and

3.00 per million completion tokens.

Figure 2: Completion-token trends over online QA streams. Each panel corresponds to one dataset. The blue and orange curves show the average completion tokens of LinearRAG and LivingRAG in each stream segment. Green shading between the two curves marks segments where LivingRAG uses fewer completion tokens, while red shading marks segments where it uses more. The first 10% of each stream is expanded to show the startup stage.

Dataset	Graph signal	Template signal
Dataset	Graph trans.	Opp.	Reuse
2Wiki	0.00	96.30	84.15
HotpotQA	0.00	21.82	9.04
MuSiQue	91.20	39.14	40.69
MuSiQue-full	82.42	54.30	50.44
WixQA	91.50	16.04	7.03

Table 4: Reuse opportunities and realized reuse in LivingRAG. “Opp.” denotes the template-reuse opportunity from Section 3.1 . “Reuse” denotes realized scaffold reuse during inference. Graph transfer is measured from activation-map fusion. The opportunity and realized values use different measurements, so the table should be read as a trend comparison rather than a direct numerical comparison. All values are percentages.

Variant	2W	MQ-F	Wix	Red.
LivingRAG	87.00	58.42	70.25	23.6%
w/o Act. fusion	86.70	56.32	67.45	15.1%
w/o Scaffold	83.80	56.92	69.55	8.3%
w/o Quality Gate	84.60	55.52	68.25	18.0%

Table 5: Ablation results. 2W, MQ-F, and Wix report LLM accuracy (%) on 2WikiMultiHopQA, MuSiQue-full, and WixQA. Red. is average completion-token reduction against LinearRAG.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Dynamics of the NS and GS gates during online inference. NS pass and experience write-back are computed over each window. GS pass is computed only over candidates that reach the GS check. The black curve shows the cumulative number of stored experiences. The first 10% of each stream is expanded to show the startup stage.

Dataset	Questions	NS pass (%)	GS checked	GS pass (%)	GS reject	Stored experiences
2Wiki	1000	13.80	138	72.46	38	100
HotpotQA	1000	63.30	633	86.73	84	549
MuSiQue	1000	36.30	363	86.23	50	313
MuSiQue-full	2417	21.31	515	91.84	42	473
WixQA	400	39.25	157	100.00	0	157
All	5817	31.05	1806	88.15	214	1592

Appendix

Table 6: Effect of the NS and GS gates on experience write-back. All runs use

\tau_{\mathrm{novel}}=0.7

and

\tau_{\mathrm{gs}}=0.6

. NS pass is measured over all questions. GS pass is measured only on candidates that reached the GS check. GS reject counts the checked candidates that failed the grounding gate.

Case	Dataset	Idx.	Query	Answer	NS	GS	Decision	Interpretation
Stored multi-hop candidate	HotpotQA	33	What subsidiary of the largest airline of the Republic of China (Taiwan) has a main hub at Taichung Airport?	Mandarin Airlines	0.793	1.000	Stored	Both gates pass. The retrieved airline passage directly supports the answer, so the candidate becomes a verified experience.
Direct WixQA grounding	WixQA	57	I would like to know how many sites I can have.	There is no limit to the number of sites you can create under a single Wix account; you can have as many as you want.	0.829	1.000	Stored	The answer contains directly supported procedural information from the retrieved help articles. This illustrates why WixQA GS scores are often high under the permissive grounding threshold.
Novelty rejection	WixQA	391	can I get a refund on the domain	No. Wix domain purchases are non-refundable and do not include a 14-day money-back guarantee or trial period. If you no longer…	0.213	–	Rejected by NS	The query repeats a stored domain-refund experience. NS assigns a low score, skips GS, and avoids storing another copy of the same support pattern.
Grounding rejection	HotpotQA	858	Spawn, a fictional character who Mustafa Shakir has expressed interest in playing, was created by?	Todd McFarlane	0.701	0.368	Rejected by GS	NS passes, but the retrieved passages do not mention the key entities. The answer matches the gold label, yet GS blocks write-back because the candidate is not grounded in retrieved evidence.

Appendix

Table 7: Representative NS and GS case studies. The cases are selected from the final LivingRAG runs without re-running retrieval or validation.

Dataset	#Q	#Chunks	#Clusters	Coverage	Avg. Size	Max Size
MuSiQue	1,000	1,354	178	79.2	4.4	21
MuSiQue-full	2,417	2,379	300	89.9	7.2	58

Appendix

Table 8: Reuse structure in the MuSiQue streams. Experience clusters are connected components of questions that share at least one decomposed single-hop question.

Dataset	Question	spaCy NER	LLM NER or useful seeds
2Wiki	When did Lothair II’s mother die?	Lothair II	Not needed
2Wiki	Which film was released first, Aas Ka Panchhi or Phoolwari?	Aas Ka Panchhi ; Phoolwari	Not needed
HotpotQA	In which county is the town in which Raymond Robertsen was born?	Raymond Robertsen	Not needed
MuSiQue	When was the person who Messi’s goals in Copa del Rey compared to get signed by Barcelona?	Messi ; Copa del Rey ; Barcelona	Not needed
WixQA	I’m want to know how much it would cost to upgrade my email plan.	[]	Email Marketing Plan ; upgrade Email Marketing Plan ; Email Marketing Plan pricing
WixQA	How do I sync the hotel app with my calendars ical link to allow visitors to book available dates on my site?	[]	Wix Bookings ; iCal URL ; syncing calendars ; booking available dates

Appendix

Table 9: Examples of query-side entity extraction. spaCy works well for standard multi-hop QA because the key retrieval cues are ordinary named entities. On WixQA, the key cues are product-domain terms, so spaCy often fails or extracts only broad entities.

Method	Query Set	0 Ent.	Avg.	Med.	Range
spaCy NER	2Wiki, first 400	1.2	1.68	2	0–4
spaCy NER	WixQA, all 400	52.5	0.61	0	0–4
LLM NER + mapping	WixQA, all 400	0.0	3.77	4	1–5

Appendix

Table 10: Query-side entity extraction statistics. LLM NER addresses the low-recall problem of standard NER on WixQA while keeping the number of query seeds small.

Statistic	Value
Chunks	7,132
Chunks with no entity	1
Avg. entities per chunk	11.23
Median entities per chunk	10
Sentences with entities	68,798
Avg. entities per entity-bearing sentence	1.79

Appendix

Table 11: Index-side entity extraction statistics for WixQA with LLM NER.

Dataset	#Chk.	Median	$<1000$	$<2000$
WixQA	7,132	1,321	44.4	59.6
2Wiki	658	4,449	0.0	0.0
HotpotQA	1,311	4,651	0.0	0.0
MuSiQue	1,354	4,701	0.0	0.0

Appendix

Table 12: Passage length distribution by character count. WixQA contains many short chunks, while the Wikipedia-style datasets are close to fixed-length chunks.

Dataset	Question	Gold answer	Words
2Wiki	When did Lothair II’s mother die?	20 March 851	3
2Wiki	Which film was released first, Aas Ka Panchhi or Phoolwari?	Phoolwari	1
HotpotQA	In which county is the town in which Raymond Robertsen was born?	Finnmark county	2
MuSiQue	When was the person who Messi’s goals in Copa del Rey compared to get signed by Barcelona?	June 1982	2
WixQA	Can I start accepting payments on my site while my Wix Payments account is still under verification?	You can start accepting payments almost immediately, but Wix must verify your identity before the account is fully activated.	27
WixQA	I want to know if the Wix store function work for selling services instead of just physical goods.	The answer distinguishes Wix Stores from Wix Bookings . For selling services, it recommends Wix Bookings because it supports service booking and online payment.	66

Appendix

Table 13: Gold answer format differs sharply between standard multi-hop QA and WixQA. Standard QA usually expects a short span. WixQA expects a support solution.

Dataset	Median words	$\leq$ 5 words	$>$ 30 words
2WikiMultiHopQA	2	95.5	0.0
HotpotQA	2	96.1	0.0
MuSiQue	2	91.2	0.0
WixQA	59	0.2	85.0

Appendix

Table 14: Answer length statistics. WixQA answers are long support solutions rather than short answer spans.

Dataset	Method	LLM s/q	Retrieval s/q	Retrieval / LLM	Validation s/q
2Wiki	LinearRAG	30.07	0.50	1.7%	N/A
2Wiki	LivingRAG	24.04	0.54	2.2%	2.13
HotpotQA	LinearRAG	15.42	0.72	4.7%	N/A
HotpotQA	LivingRAG	14.45	0.74	5.1%	2.35
MuSiQue	LinearRAG	37.23	1.49	4.0%	N/A
MuSiQue	LivingRAG	33.88	1.50	4.4%	6.64

Appendix

Table 15: LLM, retrieval, and validation times are module-level averages. Indexing, model loading, and evaluation-only LLM calls are excluded. The Retrieval / LLM column reports the ratio between retrieval time and LLM time. Validation refers to the post-answer novelty and grounding checks used for experience write-back.

Figure 8: Prompt token trends over online QA streams. Each panel corresponds to one dataset. The blue curve shows LinearRAG and the orange curve shows LivingRAG. The first 10% of each stream is expanded to show the startup stage.

Figure 9: Online use of stored experiences in LivingRAG. The curves show activation-map retrieval, reasoning scaffold use, experience write-back, and the cumulative number of stored experiences. The first 10% of each stream is expanded to show the startup stage.

Dataset	Stored experiences	Activation-map retrieval		Reasoning scaffold reuse
Dataset	Stored experiences	Top- $K$ selections	Coverage (%)	Uses	Coverage (%)	Max uses
2Wiki	100	4820	97.00	984	85.00	103
HotpotQA	549	2814	79.60	984	58.11	20
MuSiQue	313	2799	86.90	983	70.29	30
MuSiQue-full	473	6828	93.02	2401	81.61	68
WixQA	157	1910	88.54	384	67.52	20

Appendix

Table 16: Coverage of stored experiences during LivingRAG inference. Activation-map retrieval counts how often stored experiences are selected in the top-

K

experience match set used for retrieval guidance. Reasoning scaffold reuse counts how often stored experiences are selected as generation scaffolds. Coverage is the percentage of stored experiences selected at least once.

Dataset	Graph	Template	$\Delta$ LLM- Acc.	$\Delta$ Completion tokens
2Wiki	–	–	+0.59	-12.8
	–	$\checkmark$	+2.97	-189.1
MuSiQue-full	–	–	+0.56	+23.6
	$\checkmark$	–	+2.12	-297.2
	–	$\checkmark$	+4.73	-1352.9
	$\checkmark$	$\checkmark$	+9.81	-964.7

Appendix

Table 17: Performance changes conditioned on realized graph and template reuse.

Dataset	Setting	Cont.-Acc.	LLM-Acc.	Completion tokens	Graph trans.	Template reuse
2Wiki	Paper	79.60 / 82.70	84.40 / 87.00	1067.6 / 906.4	0.00	84.15
	3 orders	79.13 $\pm$ 0.42 / 83.33 $\pm$ 2.03	83.80 $\pm$ 1.22 / 89.40 $\pm$ 5.14	1038.7 $\pm$ 25.4 / 806.4 $\pm$ 92.7	0.00 $\pm$ 0.00	81.54 $\pm$ 2.69
MuSiQue-full	Paper	41.00 / 44.39	52.71 / 58.42	2223.8 / 1643.7	82.42	50.44
	3 orders	41.36 $\pm$ 1.12 / 44.21 $\pm$ 0.90	53.10 $\pm$ 0.44 / 58.09 $\pm$ 0.88	2231.8 $\pm$ 26.5 / 1744.8 $\pm$ 94.4	81.84 $\pm$ 0.55	50.15 $\pm$ 0.64
WixQA	Paper	–	66.00 / 70.25	1653.8 / 1378.1	91.50	7.03
	3 orders	–	66.25 $\pm$ 1.39 / 70.25 $\pm$ 1.00	1689.4 $\pm$ 38.6 / 1403.6 $\pm$ 22.2	91.92 $\pm$ 0.38	7.29 $\pm$ 0.30

Appendix

Table 18: Robustness across different online query orders.

Setting	Article R@5	Hit@5	MRR@5
Native PPR	62.42	68.00	42.74
+ Experience fusion	63.42	69.00	44.82

Appendix

Table 19: Retrieval-level transfer on WixQA.

Explore similar work

RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM
Jul 13, 2026Mikhail Komarov, Ivan Bondarenko, Stanislav Shtuka +5KG Construction LLM Reasoning with Graphs
1ITMO University · 2Novosibirsk State University · 3Far Eastern Federal University +1
The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA
Mar 14, 2026Yasaman Zarrinkia, Venkatesh Srinivasan, Alex ThomoGraph Retrieval-Augmented Generation Multi-Hop QA
University of Victoria · Santa Clara University
ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback
Apr 8, 2026Matthew PenarozaRetrieval-Augmented Generation Graph Retrieval-Augmented Generation

Jul 13, 2026J/K move · Enter open · S save

RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

Mikhail Komarov, Ivan Bondarenko, Stanislav Shtuka, Oleg Sedukhin+4

1ITMO University · 2Novosibirsk State University · 3Far Eastern Federal University+1

KG Construction LLM Reasoning with Graphs Graph Retrieval-Augmented Generation Small Language Models Knowledge Representation LLM Information Extraction+1

Open paper AskPDF arXiv

Graph retrieval-augmented generation (GraphRAG) enhances large language models with structured knowledge, yet existing systems construct knowledge graphs in a single extraction pass, producing noisy entities and brittle retrieval. RAGU, an open-source modular GraphRAG engine, addresses this by separating extraction from consolidation: entities and relations pass through two-stage typed extraction, DBSCAN-backed deduplication, LLM summarization, and Leiden community detection. A key insight motivates a compact extractor: the skills an in-pipeline LLM needs - comprehension, extraction, reasoning over context - are language skills that grow only weakly with model size, unlike factual world knowledge. Accordingly, we train Meno-Lite-0.1, a 7B model optimized for language skills, which outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% relative harmonic mean) and matches it on English GraphRAG tasks. On GraphRAG-Bench (Medical), RAGU retrieves the most complete context at every factoid level (evidence recall up to 0.84 vs.

\leq

0.76) and overtakes HippoRAG2 on synthesis tasks; on multi-hop factoid QA, the apparent HippoRAG2 advantage is shown to be largely an answer-format artifact. RAGU is installable via

\texttt{pip install graph_ragu}

, runs on a single GPU, and is released under MIT. The source code is publicly available at https://github.com/RaguTeam/RAGU, and the Meno-Lite-0.1 model can be obtained from https://huggingface.co/bond005/meno-lite-0.1.

Abstract

Explore similar work

RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA

ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback