Engineering IDRIS: A Multimodal Agentic AI Assistant

IDRIS is BIGRIT’s conversational AI system for interacting with documents, images, voice messages, and live web information directly through WhatsApp and Telegram.

Under the hood, IDRIS is more than a conventional chatbot. Its backend combines Retrieval-Augmented Generation, dynamic tool selection, intent routing, multimodal processing, and autonomous workflows to determine how each request should be handled.

From chatbot to agentic system

A conventional LLM application often follows a fixed flow:

User → Prompt → LLM → Answer

IDRIS was designed around a more flexible architecture:

User Request
      ↓
Intent / Task Analysis
      ↓
Tool Selection
      ↓
┌─────┼────────┬──────────┐
↓     ↓        ↓          ↓
RAG  Web     Media     Document
     Search  Tools     Processing
└─────┴────────┴──────────┘
      ↓
Result Validation
      ↓
Response Generation

This architecture allows different requests to take different execution paths rather than forcing every request through the same pipeline.

The engineering challenge is therefore not simply generating a good response. It is making the entire workflow reliable when models, tools, APIs, external data, and user inputs behave imperfectly.

More Contents:

Learn with IDRIS AI: your own hyper-personalised AI tutor