This project leverages Falcon LLMs for compliance, risk management, and real-time monitoring of LLM chatbots. Implementation Azure CosmosDB: Stores user info and violations. Multithreading in Flask: Enhances scalability. LRU Cache: Reduces latency for similar queries. Falcon LLM-powered Keyword Extractor: Creates a Rule Knowledge Graph (KG) from admin-uploaded PDFs, serving as an alternative to fine-tuning. Azure Blob Storage & Neo4j: Stores admin’s PDFs and the Knowledge Graph. Semantic Router: Retrieves relevant subgraphs and identifies violations. User Authentication: Clerk integrated with CosmosDB. Deployment: Frontend on Vercel, backend on PythonAnywhere. Objective Periodically updated Rules Knowledge Graph. Display rule violations, reasons, and priority levels. Support for multi-modal (image+text) query checks. Assign Risk scores based on violations. Analyze user behavior patterns. Methodology Flask Server: Multithreaded for handling requests and database queries. AI Pipeline: Rule Registration: Extracts and decomposes rules from PDFs, forms a knowledge graph, integrates external sources, and allows admin modifications. Query Processing: Degenerates user queries, retrieves relevant subgraphs, checks for rule violations, and reasons for violations. Risk Scoring Risk scores based on EU AI Act categories: Critical, High, Medium, Minimal. Analytics Admin panel displays rule violation frequency, custom timeframe violations, and user-wise analytics using Chart.js. Scope Compliance Assurance: Adherence to policies and regulations. Risk Management: Identifies and mitigates risks. Real-time Monitoring: Ensures transparency and accountability.
Category tags: