Toti AI Assistant — HUB Digital

Architected and engineered Toti, the conversational AI assistant integrated into the HUB Digital (INOV-NORTE) platform serving higher-education institutions in Portugal.

Toti acts as an intelligent, source-grounded companion helping educators, researchers, and students discover catalog resources, navigate learning pathways, and explore educational technologies through natural language conversations.

Key Capabilities & User Experience

• Catalog Search & Discovery: Processes natural language queries (e.g., "Show me courses about AI in education", "Suggest a collaboration tool for students") against the multi-institutional catalog.
• Structured Data Tables: Organizes multi-item comparisons, tool overviews, and course offerings into scannable markdown tables directly within the conversational interface.
• Direct Navigation & Rich Action Cards: Renders interactive UI action cards with single-click deep links pointing straight to verified catalog entities.
• Fully Bilingual Support: Automatically detects input query language and delivers fluent, culturally adapted responses in Portuguese and English.
• Omnipresent & Contextual Access: Accessible via a persistent floating action widget across all platform pages, contextual "Ask Toti" buttons pre-seeding queries with active entity metadata, and a dedicated full-page chat dashboard.

Routed RAG Architecture & Engineering Highlights

• Intent-Driven RAG Pipeline: Orchestrated using Langflow, user messages pass through a lightweight deterministic intent router classifying requests into Clarification, Restricted Retrieval, or Hybrid Retrieval.
• Type-Aware Qdrant Vector Filtering: Indexes seven distinct catalog entity families (course, tool, resource, best_practice, case_study, pedagogical_method, community). Retrieval injects metadata payload filters to enforce clean boundary separation.
• Strict Anti-Hallucination & Source Grounding Policies:
  • Restricted Agent: Courses and communities are strictly bounded to internal verified catalog data in Qdrant—external web augmentation is explicitly disallowed to prevent hallucinating dates, requirements, or availability.
  • Hybrid Agent: Tools, resources, and pedagogical methods query internal records first, allowing controlled external web search exclusively to enrich already-retrieved catalog entities.
• Clarification-First Fallback: If query confidence is low or intent is underspecified, Toti immediately returns targeted clarification prompts without triggering costly or misleading vector searches.
• Dual-Engine LLM Support: Designed for high-reliability production with Azure OpenAI and containerized local model inference via Ollama.
• Real-Time Directus CMS Synchronization: Automated webhook and operation hooks keep Qdrant vector collections synchronized upon catalog create, update, and delete events.
• Frontend State Management: Built on Nuxt 3 with a dedicated Pinia chat store (stores/chat.ts) and modular composables (useChat) managing streaming session lifecycles, history caching, and authenticated Directus AI endpoints.