Paper deep dive

RUVA: Personalized Transparent On-Device Graph Reasoning

Gabriele Conte, Alessio Mattiace, Gianni Carmosino, Potito Aghilar, Giovanni Servedio, Francesco Musicco, Vito Walter Anelli, Tommaso Di Noia, Francesco Maria Donini

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 20

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 87%

Last extracted: 7/21/2026, 3:19:46 AM

Summary

The paper introduces RUVA, a personalized, on-device 'Glass Box' architecture for Personal AI that utilizes a Personal Knowledge Graph (PKG) instead of standard vector databases. RUVA employs a neuro-symbolic approach where a Small Vision Language Model (SVLM) extracts structured triples from multimodal data to populate the graph, enabling transparent, human-in-the-loop memory curation. Key features include precise graph reasoning for complex queries, deterministic deletion of specific facts to ensure privacy ('Right to be Forgotten'), and offline-first operation on mobile devices to maintain data sovereignty.

Entities (13)

RUVA · systemarchitecture · 98%Personal Knowledge Graph · data-structure · 95%Small Vision Language Model · model · 92%Gemma-3n-E2b · model · 90%Retrieval-Augmented Generation · methodology · 90%GraphRAG · methodology · 88%Google Pixel 8 Pro · hardware · 85%SQLite · softwaredatabase · 85%GPT · model · 80%Kimi K2 · model · 80%Leiden algorithm · algorithm · 80%LLaMA3.3-70B · model · 80%Qwen3-32b · model · 80%

Relation Signals (11)

RUVA → uses → Personal Knowledge Graph

confidence 95% · Ruva grounds Personal AI in a Personal Knowledge Graph

RUVA → employs → Small Vision Language Model

confidence 92% · The Neural Component employs a Small Vision Language Model (SVLM)

Small Vision Language Model → isinstanceof → Gemma-3n-E2b

confidence 90% · Small Vision Language Model (SVLM, i.e., Gemma-3n-E2b)

RUVA → implements → GraphRAG

confidence 88% · RUVA is a neuro-symbolic GraphRAG architecture

RUVA → deployedon → Google Pixel 8 Pro

confidence 85% · We deployed the system on a constrained environment, a Google Pixel 8 Pro.

RUVA → storesdatain → SQLite

confidence 85% · The system backbone is a hybrid storage engine: a single SQLite database

RUVA → evaluatedagainst → Kimi K2

confidence 80% · We employ a panel of four state-of-the-art LLMs: ... and Kimi K2

RUVA → evaluatedagainst → LLaMA3.3-70B

confidence 80% · We employ a panel of four state-of-the-art LLMs: Llama3.3-70B

RUVA → →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:The Personal AI landscape is currently dominated by "Black Box" Retrieval-Augmented Generation. While standard vector databases offer statistical matching, they suffer from a fundamental lack of accountability: when an AI hallucinates or retrieves sensitive data, the user cannot inspect the cause nor correct the error. Worse, "deleting" a concept from a vector space is mathematically imprecise, leaving behind probabilistic "ghosts" that violate true privacy. We propose Ruva, the first "Glass Box" architecture designed for Human-in-the-Loop Memory Curation. Ruva grounds Personal AI in a Personal Knowledge Graph, enabling users to inspect what the AI knows and to perform precise redaction of specific facts. By shifting the paradigm from Vector Matching to Graph Reasoning, Ruva ensures the "Right to be Forgotten." Users are the editors of their own lives; Ruva hands them the pen. The project and the demo video are available at this http URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

18,639 characters extracted from source content.

Expand or collapse full text

RUVA: Personalized Transparent On-Device Graph Reasoning Gabriele Conte 1 , Alessio Mattiace 1 , Gianni Carmosino 1 , Potito Aghilar 1 , Giovanni Servedio 1 , Francesco Musicco 1 , Vito Walter Anelli 1 , Tommaso Di Noia 1 and Francesco Maria Donini 2 1 Politecnico di Bari, 2 Universit ` a degli Studi della Tuscia g.conte12, a.mattiace, g.carmosino1@studenti.poliba.it, potito.aghilar, giovanni.servedio, vitowalter.anelli, tommaso.dinoia@poliba.it f.musicco@phd.poliba.it, donini@unitus.it Abstract The Personal AI landscape is currently dominated by “Black Box” Retrieval-Augmented Generation. While standard vector databases offer statistical matching, they suffer from a fundamental lack of accountability: when an AI hallucinates or retrieves sensitive data, the user cannot inspect the cause nor correct the error. Worse, “deleting” a concept from a vector space is mathematically imprecise, leaving behind probabilistic “ghosts” that violate true pri- vacy. We propose RUVA, the first “Glass Box” ar- chitecture designed for Human-in-the-Loop Mem- ory Curation. Ruva grounds Personal AI in a Per- sonal Knowledge Graph, enabling users to inspect what the AI knows and to perform precise redaction of specific facts. By shifting the paradigm from Vector Matching to Graph Reasoning, RUVA en- sures the “Right to be Forgotten.” Users are the ed- itors of their own lives; RUVA hands them the pen. 1 Introduction Since their release, Large Language Models have quickly transitioned from novelties to commodities [ Liang et al., 2025 ] .The early industry dream of hyper-personalizing these models via fine-tuning has collapsed under the weight of computational costs and the risk of catastrophic forget- ting [ Luo et al., 2025 ] . Consequently, the field has con- verged on Retrieval-Augmented Generation (RAG) [ Lewis et al., 2020 ] as the only viable path for personalization. How- ever, the reliance on standard Vector RAG has created a cri- sis: the “Black Box” problem. Current RAG systems com- press a user’s life (e.g., emails, photos, chats, calendar en- tries) into high-dimensional embedding spaces. While effi- cient for matching semantically similar keywords, the user cannot inspect why the system fails or fix it. More criti- cally, true privacy requires the ability to delete, but one can- not surgically remove a concept from a vector space; one can only suppress it. This mathematical imprecision leaves be- hind “probabilistic ghosts”, traces of sensitive data that the AI might accidentally resurrect [ Wang et al., 2025 ] . GraphRAG methods [ Edge et al., 2024 ] address these limitations by in- corporating graph topology into retrieval. We introduce RUVA (the Etruscan word for “Brother”), a system designed to be as trustworthy and protective as a close sibling. RUVA is a neuro-symbolic GraphRAG ar- chitecture [ Kautz, 2022 ] that runs entirely on-device, rep- resenting a paradigm shift from opaque retrieval to transpar- ent reasoning. Trust begins with sovereignty. By running locally, RUVA ensures that the user’s Personal Knowledge Graph (PKG) [ Balog and Kenter, 2019; Skjæveland et al., 2024 ] never leaves their device. There is no cloud to hack; the user’s life remains physically in their pocket. Unlike vector stores, RUVA grounds its intelligence in a “Glass Box” archi- tecture. Here, every memory is an explicit node, and every association is a visible edge. This structure highlights a core consideration: Vectors allow Matching; Graphs allow Rea- soning. Consider the query: “Did Sarah call before I arrived at work?” A standard vector system simply retrieves chunks containing “Sarah” and “Work” based on similarity, often hal- lucinating the temporal relationship. In contrast, RUVA tra- verses the graph topology, comparing the timestamp of the Call node against the Arrival node to provide the answer. Finally, RUVA redefines the user’s role from a passive data source to the Editor-in-Chief. Because the memory is struc- tured, it is fully editable. If RUVA learns something incor- rect or sensitive, the user can use the “Red Pen” to perform a precise redaction, cutting that specific node (the reified in- formation object) and its edges. The knowledge is deleted instantly and mathematically, ensuring the “Right to be For- gotten.” RUVA is not another chatbot. It is an AI that users can actually trust, because RUVA allows them to see inside its head and edit its memories. The project and the demo video are available at http://sisinf00.poliba.it/ruva/. 2 The RUVA Architecture RUVA is engineered as an offline-first, mobile-native system designed to guarantee data sovereignty. To ensure that per- sonal data never leaves the user’s possession, all computa- tion occurs locally on the user’s mobile device; the Personal Knowledge Graph (PKG) is physically bound to the hand- set, eliminating cloud-based attack vectors. Furthermore, the system is optimized for mobile constraints, utilizing quan- tized models and efficient storage to operate alongside the OS without draining the battery. As illustrated in Figure 1, RUVA implements a Type 3 Neuro-Symbolic architecture (taxon- arXiv:2602.15553v1 [cs.AI] 17 Feb 2026 APPs PKG Population PKG Utilization (I) Interaction (I) Query Graph RetrievalGeneration Triples Extraction Answer Metadata Text Image Triples Injection [EXAMPLE] The user takes a picture of his cat. A cat with thick gray fur, light eyes, and white whiskers. (Cat)-[has]->(thick gray fur) (Cat)-[has]->(light eyes) (Cat)-[has]->(white whiskers) i (photo_01)-[has_date]->(01-Sep-2025) (photo_01)-[has_time]->(10:04) dateTime: 2025-09-01T10:04:00 PKG U Captioning Text to Triples Gemma3n-E2b Community Detection ... Figure 1: The RUVA architecture. The Ingestion Workflow transforms multimodal data into semantic triples to populate the Personal Knowl- edge Graph. The Retrieval Workflow performs graph traversal to generate grounded, hallucination-free answers, entirely on-device. omy from Kautz [ 2022 ] ) that separates perception from mem- ory. The Neural Component employs a Small Vision Lan- guage Model (SVLM, i.e., Gemma-3n-E2b [ Gemma Team, 2025 ] ) as the system’s “eyes and ears.” Its role is to convert unstructured noise (pixels, raw text) into structured signals, accepting the probabilistic nature of recognition. These sig- nals are fed into the Symbolic Component, a PKG used for storage and reasoning (Figure 2a). Unlike the neural compo- nent, the memory, once created, is deterministic, inspectable, and editable. The system backbone is a hybrid storage en- gine: a single SQLite database extended with sqlite-vec. For this research, the engine is extended to accommodate GraphRAG [ Edge et al., 2024 ] operations. This design choice is critical for the “Glass Box” paradigm. The resulting graph topology resembles a “spiderweb,” where a central User node acts as the root, and all entities, Events, Photos, Messages, branch out through typed edges, creating a uni- fied representation of the user’s digital footprint. 2.1 Multimodal Ingestion Workflow RUVA operates a background “Triple Extraction” pipeline that transforms heterogeneous data streams into semantic triples. For visual data, such as a photo of a train ticket, the image is passed to the local SVLM. The SVLM gener- ates a dense caption, which is then parsed to extract struc- tured entities, effectively converting pixels into nodes like (:Receipt)-[:cost]->("95 EUR"). Textual data, such as emails or notes, is processed directly by the SVLM to extract entities and temporal metadata. To prevent graph fragmentation, RUVA employs an Entity Resolution strategy along with community detection (Leiden algorithm [ Traag et al., 2019 ] ). For instance, if the system ingests an email from “Sarah Green” and a calendar invite from “S. Green,” the clustering identifies these as the same entity, merging them into a single Person node rather than creating duplicates. 2.2 Graph-Grounded Retrieval Workflow Standard Vector RAG fails at complex reasoning because it relies solely on semantic similarity. RUVA solves this via a multi-step Graph-Grounded Retrieval mechanism. Upon receiving a query like “Did Sarah call before I arrived at work?”, the system first executes an Anchor Search using vector similarity to find the relevant “Anchor Nodes” (e.g., the Person:Sarah node and the Location:Work node). It then performs a configurable Topological Expansion, ex- ecuting an N -hop traversal from these anchors to identify connected events. Finally, during Answer Generation (Fig- ure 2b), the retrieved subgraph, containing explicit times- tamps and relationships, is serialized and fed into the SVLM. 2.3 The Deletion Mechanism One of the most significant advantages of the RUVA archi- tecture is user-controlled knowledge deletion.In vector- only systems, deleting a concept is mathematically imprecise, leaving behind probabilistic ghosts. In RUVA, since every memory is a discrete node in a relational database, deletion is exact. When a user chooses to forget a specific memory (e.g., “The project meeting on Friday”), RUVA executes a standard SQL DELETE CASCADE transaction targeting the reified in- formation object. The node, its properties, and all associated vector indices are instantly and permanently excised from the memory. This provides the user with deterministic control over their digital past. 3 Demonstration Scenarios We demonstrate the “Glass Box” RUVA architecture with three use cases: (i) ingestion of a multimodal memory; (i) complex reasoning, and (i) memory deletion. The demonstration begins with Multimodal Ingestion (Scenario 1), illustrating how the system reacts when mul- timodal data enters the system. Let us introduce an ex- ample. The user syncs a calendar containing a “Weekend Trip” and uploads a photo of a train ticket. The Multimodal Ingestion Workflow triggers: the local VLM converts the ticket pixels into a structured node (:Receipt :amount "95 EUR"), while the system identifies the temporal overlap with the calendar event, forging a relationship edge. Next, we demonstrate the Complex Reasoning (Sce- nario 2). Again, let us show it through an example. The user asks: “How much have I spent on the trip so far?” Answer- ing this requires bridging the concept “The Trip” from the calendar with granular data “95 EUR” from the image. The GraphRAG runtime executes a multi-hop traversal, locating (a) Extracted PKG in RUVA(b) Answer to multimodal query Figure 2: RUVA App Screenshots. Figure 2a shows the inferred PKG, while in Figure 2b RUVA answers leveraging multimodal data. the Trip node and following the temporal edge to the Receipt node, answering “You have spent 95 EUR for the ticket.” Finally, we showcase the Deterministic Deletion (Sce- nario 3), addressing the “Right to be Forgotten.” In this ex- ample, the user decides that a financial piece of information (i.e., the receipt in Scenario 2) is sensitive and should not be retained. By selecting the Receipt in the gallery app and re- moving it, the system executes a deterministic DELETE cas- cade transaction. The node and all associated vector indices are excised. To verify this, the user asks the same question again, and the system responds: “I couldn’t find relevant in- formation to answer your question.” This scenario demon- strates true data sovereignty: the user acts as the editor of their own memory, enforcing privacy with mathematical certainty in a way that opaque vector-only systems cannot guarantee. 4 Evaluation This section assesses the RUVA’s operational feasibility of the architecture and the semantic reasoning quality. Performance. We deployed the system on a constrained en- vironment, a Google Pixel 8 Pro. RUVA periodically scans the device’s local data and compares it with existing nodes. De- spite the constrained environment, RUVA ensures interactive latencies, with an average ingestion time of 2.4s (including VLM processing and triple extraction) and a graph retrieval latency of only 38ms for single-hop queries. Model Accuracy. To comprehensively test RUVA, we pre- pared a benchmark comprising 71 multi-source objects (span- ning calendar events, images, notes, textual documents, calls, alarms, and contacts) and 52 triplets containing questions, de- sired answers, and RUVA’s answers. The benchmark specif- ically probes three distinct scenarios: Multimodal Ingestion (Figure 3a), Complex Reasoning (Figure 3b), and Determin- istic Deletion (Figure 3c). The benchmark consists of 20 in- gestion checks and 32 complex reasoning questions. To as- 0%20%40%60%80%100% GPT Kimi Llama Qwen (a) Scenario 1: Multimodal Ingestion 0%20%40%60%80%100% GPT Kimi Llama Qwen Score 5Score 4Score 3Score 2Score 1 (b) Scenario 2: Complex Reasoning 0%20%40%60%80%100% GPT Kimi Llama Qwen ∆4 ∆3∆2 (c) Scenario 3: Deterministic Deletion (Delta Scores) Figure 3: Judicial scores for Scenario 1 and 2, and∆ scores for 3. sess Deterministic Deletion, each question is evaluated with (score 0 ) and without (score 1 ) its corresponding information objects, measuring ∆ = score 0 − score 1 . We considered answers whose judges’ score is positive (3-5). Experiments. We employ a panel of four state-of-the-art LLMs: Llama3.3-70B [ Meta, 2024 ] , Qwen3-32B [ Qwen Team, 2025 ] , GPT [ OpenAI, 2025 ] , and Kimi K2 [ Kimi Team et al., 2025 ] Instruct. Following the Prometheus [ Kim et al., 2024 ] strategy, judges evaluate benchmark triplets via Chain- of-Thought reasoning. To mitigate inter-model bias [ Lee et al., 2025 ] , we map the resulting 5-point scores to three categories: Positive (4–5), Neutral (3), and Negative (1–2). RUVA demonstrates robust reasoning, with 61% of responses rated as Positive, increasing to 71% when including Neu- tral responses. (Figure 3a, Figure 3b). Reliability analy- sis shows strong cross-model agreement, with a Spearman’s Rank Correlation (ρ) [ Spearman, 1904 ] of 0.82 and a Krip- pendorff’s Alpha [ Krippendorff, 2011 ] of 0.81, indicating consistency in ordinal evaluations. The 83% Percentage Agreement [ McHugh, 2012 ] and Cohen’s κ [ Cohen, 1968 ] of 0.81 confirm this finding. Notably, gpt-oss-120b and kimi-k2-instruct-0905 achieve up to 92% agree- ment (κ = 0.92), highlighting highly aligned assessments. 5 Conclusion We introduce RUVA, redefining Personal AI from an opaque service into a transparent framework. By adopting a neuro- symbolic GraphRAG paradigm, we addressed the ”Black Box” limitations that plague current retrieval systems. RUVA is viable on edge devices, achieving interactive latencies and high semantic accuracy. This proves that the cloud is no longer a prerequisite for intelligence, and data sovereignty is possible. Thanks to the graph-based topology, RUVA en- ables deterministic deletion and operationalizes the ”Right to be Forgotten”. RUVA returns agency to the human, enabling the user to inspect, correct, and erase the memories held by their AI. RUVA demonstrates that the most powerful AI is the one that the user can trust the most. References [ Balog and Kenter, 2019 ] Krisztian Balog and Tom Kenter. Personal knowledge graphs: A research agenda. In Pro- ceedings of the 2019 ACM SIGIR International Confer- ence on Theory of Information Retrieval, ICTIR 2019, Santa Clara, CA, USA, October 2-5, 2019, pages 217–220. ACM, 2019. [ Cohen, 1968 ] Jacob Cohen.Weighted kappa: Nominal scale agreement provision for scaled disagreement or par- tial credit. Psychological Bulletin, 70(4):213–220, 1968. [ Edge et al., 2024 ] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Tru- itt, and Jonathan Larson. From local to global: A graph RAG approach to query-focused summarization. CoRR, abs/2404.16130, 2024. [ Gemma Team, 2025 ] Gemma Team. Gemma 3 technical re- port, 2025. [ Kautz, 2022 ] Henry A. Kautz. The third AI summer: AAAI robert s. engelmore memorial lecture. AI Mag., 43(1):93– 104, 2022. [ Kim et al., 2024 ] Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Min- joon Seo. Prometheus: Inducing fine-grained evaluation capability in language models, 2024. [ Kimi Team et al., 2025 ] Kimi Team, Yifan Bai, et al. Kimi k2: Open agentic intelligence, 2025. [ Krippendorff, 2011 ] Klaus Krippendorff. Computing krip- pendorff’s alpha-reliability. 2011. [ Lee et al., 2025 ] Yukyung Lee, JoongHoon Kim, Jaehee Kim, Hyowon Cho, Jaewook Kang, Pilsung Kang, and Na- joung Kim. CheckEval: A reliable LLM-as-a-judge frame- work for evaluating text generation using checklists. In Proceedings of the 2025 Conference on Empirical Meth- ods in Natural Language Processing, pages 15771–15798, Suzhou, China, November 2025. [ Lewis et al., 2020 ] Patrick Lewis, Ethan Perez, Aleksan- dra Piktus, Fabio Petroni, Vladimir Karpukhin, Na- man Goyal, Heinrich K ̈ uttler, Mike Lewis, Wen-tau Yih, Tim Rockt ̈ aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In NeurIPS, 2020. [ Liang et al., 2025 ] Weixin Liang, Yaohui Zhang, Mihai Co- dreanu, Jiayu Wang, Hancheng Cao, and James Zou. The widespread adoption of large language model-assisted writing across society. Patterns, 6(12):101366, 2025. [ Luo et al., 2025 ] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models dur- ing continual fine-tuning. IEEE Transactions on Audio, Speech and Language Processing, 2025. [ McHugh, 2012 ] Mary L. McHugh. Interrater reliability: the kappa statistic. Biochemia Medica, 22(3):276–282, 2012. [ Meta, 2024 ] Meta. The llama 3 herd of models, 2024. [ OpenAI, 2025 ] OpenAI. gpt-oss-120b & gpt-oss-20b model card, 2025. [ Qwen Team, 2025 ] Qwen Team. Qwen3 technical report. CoRR, abs/2505.09388, 2025. [ Skjæveland et al., 2024 ] Martin G. Skjæveland, Krisztian Balog, Nolwenn Bernard, Weronika Lajewska, and Trond Linjordet. An ecosystem for personal knowledge graphs: A survey and research roadmap. AI Open, 5:55–69, 2024. [ Spearman, 1904 ] Charles Spearman. The proof and mea- surement of association between two things. The Ameri- can Journal of Psychology, 15(1):72–101, January 1904. [ Traag et al., 2019 ] V. A. Traag, L. Waltman, and N. J. van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific Reports, 9(1), March 2019. [ Wang et al., 2025 ] Shang Wang, Tianqing Zhu, Dayong Ye, and Wanlei Zhou.When machine unlearning meets retrieval-augmented generation (rag): Keep secret or for- get knowledge? IEEE Transactions on Dependable and Secure Computing, page 1–16, 2025.

Instant research discovery