RUVA: Personalized Transparent On-Device Graph Reasoning

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 forgetting [Luo et al., 2025]. Consequently, the field has converged on Retrieval-Augmented Generation (RAG) [Lewis et al., 2020] as the only viable path for personalization [Wang et al., 2023; Anelli et al., 2017]. However, the reliance on standard Vector RAG has created a crisis: the "Black Box" problem. Current RAG systems compress a user's life (e.g., emails, photos, chats, calendar entries) into high-dimensional embedding spaces. While efficient for matching semantically similar keywords, the user cannot inspect why the system fails, or fix it [Servedio et al., 2025; Di Palma et al., 2025]. More critically, true privacy requires the ability to delete, but one cannot surgically remove a concept from a vector space; one can only suppress it. This mathematical imprecision leaves behind "probabilistic ghosts", traces of sensitive data that the AI might accidentally resurrect [Wang et al., 2025].

∗Corresponding authors.

GraphRAGmethods [Edge et al., 2024] address these limitations by incorporating 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 architecture [Kautz, 2022] that runs entirely on-device, representing a paradigm shift from opaque retrieval to transparent 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" architecture. 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 Reasoning. 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 hallucinating the temporal relationship. In contrast, RUVA traverses 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 structured, it is fully editable. If RUVA learns something incorrect or sensitive, the user can use the "Red Pen" to perform a precise redaction, cutting that specific node (the reified information object) and its edges. The knowledge is deleted instantly and mathematically, ensuring the "Right to be Forgotten." 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 https://ruva-assistant.github.io/.

2 The RUVA Architecture

RUVA is engineered as an offline-first, mobile-native system designed to guarantee data sovereignty. To ensure that personal data never leaves the user's possession, all computation occurs locally on the user's mobile device; the Personal Knowledge Graph (PKG) is physically bound to the handset, eliminating cloud-based attack vectors [Aghilar et al., 2023]. Furthermore, the system is optimized for mobile constraints, utilizing quantized models and efficient storage to operate alongside the OS without draining the battery. As illustrated in Figure 1, RUVA implements a Type 3 NeuroSymbolic architecture (taxonomy from Kautz [2022]) that separates perception from memory. The Neural Component employs a Small Vision Language Model (SVLM, i.e., Gemma-3n-E2b [Kamath et al., 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 signals are fed into the Symbolic Component, a PKG used for storage and reasoning (Figure 2a). Unlike the neural component, the memory, once created, is deterministic, inspectable, and editable. The system backbone is a hybrid storage engine: a 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 "spiderweb" topology roots at a central User node, from which entities (Events, Photos, Messages) branch via typed edges, unifying the user's digital footprint.

2.1 Multimodal IngestionWorkflow

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 [Aghilar et al., 2025]. The SVLM generates a dense caption, which is then parsed to extract structured 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, clustering merges an email from "Sarah Green" and an invite from "S. Green" into a single Person node, preventing 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, executing an N-hop traversal from these anchors to identify connected events. Finally, during Answer Generation (Figure 2b), the retrieved subgraph, containing explicit timestamps and relationships, is serialized and fed into the SVLM.

2.3 The DeletionMechanism

One of the most significant advantages of the RUVA architecture is user-controlled knowledge deletion. In vectoronly 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 information object. The node, its properties, and associated vector indices are permanently excised from the memory, granting users 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; (ii) complex reasoning, and (iii) memory deletion.

The demonstration begins with Multimodal Ingestion (Scenario 1), illustrating how the system reacts when multimodal data enters the system. Let us introduce an example. 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 (Scenario 2). Again, let us show it through an example. The user asks: "How much have I spent on the trip so far?" Answering 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 the Trip node and following the temporal edge to the Receipt node, answering "You have spent 95 EUR for the ticket."

Figure 2: RUVA App Screenshots. Figure 2a shows the inferred PKG, while in Figure 2b RUVA answers leveraging multimodal data.

Finally, we showcase the Deterministic Deletion (Scenario 3), addressing the "Right to be Forgotten." In this example, 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 removing it, the system executes a deterministic DELETE cascade 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 information to answer your question." This scenario demonstrates 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 systemon a constrained environment, a Google Pixel 8 Pro. RUVA periodically scans the device's local data and compares it with existing nodes. Despite 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 prepared a benchmark comprising 71 multi-source objects (spanning calendar events, images, notes, textual documents, calls, alarms, and contacts) and 52 triplets containing questions, desired answers, and RUVA's answers. The benchmark specifically probes three distinct scenarios: Multimodal Ingestion (Figure 3a), Complex Reasoning (Figure 3b), and Deterministic Deletion (Figure 3c). The benchmark consists of 20 ingestion checks and 32 complex reasoning questions. To as-

∆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 (score0) and without (score1) its corresponding information objects, measuring ∆ = score0 - score1. 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 [Grattafiori et al., 2024], Qwen332B [Yang et al., 2025], GPT [Agarwal et al., 2025], and Kimi K2 [Bai 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 Neutral responses. (Figure 3a, Figure 3b). Reliability analysis shows strong cross-model agreement, with a Spearman's Rank Correlation (ρ) [Spearman, 1904] of 0.82 and a Krippendorff'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 and Kimi K2 achieve up to 92% agreement (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. Achieving interactive latencies and high semantic accuracy on edge devices, RUVA proves the cloud is no longer a prerequisite for intelligence, ensuring data sovereignty. Its graph topology enables deterministic deletion, operationalizing the "Right to be Forgotten". By allowing users to inspect, correct, and erase memories, RUVA restores human agency.

Acknowledgments

The authors acknowledge partial support of OVS Fashion Retail Reloaded, Lutech Digitale 4.0, Servizi Locali 2.0, Konecta New KOCS ERA, P+ARTS, PNRR-MAD-202212376656, REACH-XY and CALLIOPE. Furthermore, this work has been carried out while Giovanni Servedio was enrolled in the Italian National Doctorate on Artificial Intelligence run by Sapienza University of Rome in collaboration with Politecnico di Bari.