IDEA

Intelligent Digital Educational Agent (IDEA)

An AI-powered cryptocurrency advisory system that leverages Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) to provide accurate, accessible, and educational insights into cryptocurrency concepts.

---

Project Overview

The cryptocurrency domain is complex, filled with rapidly evolving technologies and terminology that can be difficult for both beginners and experienced users to navigate.
This project develops an Intelligent Digital Educational Agent (IDEA) that:

• Processes and validates cryptocurrency-related documents
• Uses semantic search (FAISS + Sentence Transformers) for query-document matching
• Integrates LLMs (via LangChain) to generate contextually accurate responses
• Provides domain-specific advising and explanations tailored for educational purposes

---

Key Features

• Document Processing Pipeline – Extracts, validates, and embeds cryptocurrency-related PDFs
• Vector Search with FAISS – Retrieves semantically similar chunks for user queries
• Hybrid RAG + LLM Architecture – Enhances accuracy and contextual grounding
• Model and Methodology Comparison – Evaluates Llama 3.3, Llama 4 Scout, Qwen QwQ-32B, and Flan-T5 across Prompt Engineering, RAG, and Fine-Tuning
• Educational Focus – Provides level-appropriate explanations and supports progressive learning

---

System Architecture

1. PDF Upload and Validation – Filters for cryptocurrency-related documents
2. Text Extraction and Embedding – Creates FAISS index with semantic embeddings (all-MiniLM-L6-v2)
3. Query Handling – Retrieves top-k relevant chunks
4. LLM Response Generation – Augments prompts with retrieved context and generates accurate answers
5. Evaluation Framework – Compares models and approaches on performance, speed, and accuracy

---

Results and Insights

• Best Model: Llama 3.3 (70B) with Prompt Engineering (combined score: 0.451)
• Fastest Response: Llama 3.3 (70B) at approximately 0.85 seconds
• Strongest Domain Knowledge: Qwen QwQ-32B
• Key Finding: Well-crafted prompts often outperform RAG and fine-tuning in cryptocurrency advising

---

Future Directions

• Fine-tuning on cryptocurrency-specific datasets
• Optimizing RAG with larger context windows and improved retrieval mechanisms
• Expanding to specialized domains such as DeFi and NFTs
• Incorporating real-time market data
• Multi-language support and student-level adaptation
• Interactive learning features (quiz mode and progress tracking)

---

Technology Stack

• LLMs: Llama 3.3 (70B), Llama 4 Scout, Qwen QwQ-32B, Flan-T5
• Frameworks: LangChain, Groq API
• Retrieval: FAISS (Facebook AI Similarity Search)
• Embeddings: Sentence Transformers (all-MiniLM-L6-v2)
• Data Processing: Python (PDF handling, text chunking, evaluation framework)