Technical Summary: GenAI-Driven Approach to RISC-V Supply Chain Exploration

1. Problem Statement

The rapid adoption of RISC-V in the automotive sector introduces significant complexity in supply chain management. Unlike vertically integrated proprietary ecosystems, the RISC-V landscape is a decentralized network of stakeholders, including IP core designers, EDA tool providers, foundries, and system integrators. This fragmentation results in critical supply chain data being dispersed across heterogeneous, unstructured, and semi-structured sources, such as technical whitepapers, standards specifications, and visual artifacts (diagrams, block schematics, and scanned documents).

Traditional analytical methods, which rely on structured databases and manual curation, are insufficient for capturing the scale and richness of this multi-modal information. Furthermore, the automotive domain imposes stringent regulatory requirements (e.g., ISO 26262) for traceability and validation, necessitating robust mechanisms to map interdependencies and monitor risks. The core challenge addressed by this paper is the lack of automated tools capable of synthesizing textual and visual data to generate formal, machine-interpretable models of RISC-V supply chain dependencies.

2. Methodology

The paper proposes a GenAI-driven workflow that integrates Large Language Models (LLMs), Vision-Language Models (VLMs), and Model-Driven Engineering (MDE) to transform raw data into structured knowledge and formal process models. The workflow operates through the following stages:

2.1 Multi-Modal Data Ingestion and Extraction

The system ingests heterogeneous inputs, including technical documents, vendor reports, and visual artifacts.

• Textual Processing: LLMs are employed to extract entities (stakeholders, components, IP blocks) and relationships from unstructured text.
• Visual Processing: VLMs interpret diagrams, flowcharts, and tables to extract structural information that text-only models might miss.
• Prompting Strategy: The extraction is guided by specific prompts. For instance, to generate knowledge graphs, the system uses prompts instructing the model to output Neo4j-compatible node and relationship definitions without extraneous commentary. Similarly, for process extraction, prompts direct the generation of PlantUML activity diagrams with multiple swimlanes representing actor interactions.

2.2 Knowledge Graph Construction

Extracted entities and relationships are organized into a Knowledge Graph (KG) using Neo4j.

• Schema: The graph captures participants (e.g., semiconductor companies), products (processors, IP cores), and interdependencies (licensing, supply relationships).
• Querying: Analysts can utilize the Cypher query language to explore multi-hop dependencies and identify bottlenecks. The framework includes a reasoning layer where LLMs assist in formulating Cypher queries based on natural language guidelines provided by human analysts.

2.3 Process Extraction and Formal Modeling

To address workflow analysis, the system generates formal process models using PlantUML activity diagrams.

• Intermediate Representation: The extracted process semantics are converted into PlantUML text, which is then transformed into a JSON format for machine processing.
• Rule-Based Validation: A rule-based framework is introduced to validate the ordering and coordination of activities. Constraints are defined using operators such as before, after, after-true, and after-false to model conditional execution paths (e.g., validation checks). Role-annotated rules (e.g., Role: Action) are used to assign responsibilities to specific actors (designers, vendors, foundries).
• Human-in-the-Loop: The workflow incorporates a human analyst who can refine extracted models, adjust rules, and validate assumptions, ensuring domain expertise is integrated with automated intelligence.

2.4 Implementation

The workflow is implemented using the open-source orchestration tool n8n, which connects input handling, GenAI model prompting, and MDE analysis.

• Models: The system utilizes locally deployed models to ensure data privacy, particularly relevant for sensitive automotive domains. The evaluation uses openai/gpt-oss-120b for text and Qwen/Qwen3.5-122B-A10B for visual tasks, running on a cluster with three H200 NVL GPUs.
• Interaction: Users can interact via an interactive chat interface or through webhooks that accept file uploads (diagrams, documents) to trigger workflow execution.

3. Key Contributions

• Multi-Modal Integration: The paper presents a unified framework that combines LLMs for textual semantic understanding and VLMs for visual structural extraction, addressing the limitation of text-centric supply chain analysis tools.
• Formal Modeling via MDE: By integrating Model-Driven Engineering, the approach transforms unstructured data into formal representations (Knowledge Graphs and PlantUML diagrams) that support constraint-based analysis and risk detection.
• Rule-Based Process Validation: The introduction of a rule-based layer for validating activity ordering and role assignments allows for the systematic detection of unsafe or inconsistent supply chain behaviors at design time.
• Local Deployment Strategy: The demonstration of a locally deployable pipeline using open-source tools and locally hosted models offers a privacy-preserving alternative to cloud-based GenAI services, which is critical for the automotive industry.

4. Experimental Results

The approach was evaluated using a representative RISC-V supply chain description and corresponding diagrams. Ground truth data was established with the assistance of GPT-5.3.

• Knowledge Graph Extraction:

  • LLM Performance: Achieved 88% accuracy in concept extraction (22/25), 83% in relationships (19/23), and 75% in attributes (6/8).
  • VLM Performance: Achieved 72% in concepts, 65% in relationships, and 63% in attributes.
  • Query Generation: The system generated Cypher queries with 75% end-to-end correctness (15/20 queries fully correct).

• Process Analysis:

  • LLM Performance: Identified 91% of activities (10/11), 100% of participants, 82% of relationships, and 78% of artifacts. It successfully generated 80% of the validation rules.
  • VLM Performance: Identified 73% of activities, 75% of participants, and 64% of relationships. VLMs were not used for rule generation due to limitations in abstract reasoning.

• Performance Metrics: The average response time was approximately 11.5 seconds for LLMs and 14 seconds for VLMs.

The results indicate that while LLMs excel in semantic reasoning and rule generation, VLMs provide complementary value in extracting structural elements from visual inputs. The combination of both modalities yields a more comprehensive analysis than either modality alone.

5. Significance and Claims

The paper claims that this GenAI-driven approach effectively addresses the challenges of analyzing decentralized and heterogeneous RISC-V supply chains. By synthesizing multi-modal data into formal models, the framework enhances transparency, supports decision-making, and improves resilience in complex semiconductor supply chains.

The authors emphasize that the use of locally deployable models, combined with iterative validation and intermediate representations, brings performance close to state-of-the-art commercial models while maintaining data sovereignty. This is presented as a significant advantage for sensitive domains like automotive, where data privacy and regulatory compliance are paramount. The work concludes that this methodology provides a scalable foundation for monitoring supply chain developments, identifying emerging actors, and assessing systemic risks in the RISC-V ecosystem.