Terus Rag Block for Moodle

Overview

This Moodle block plugin implements Retrieval-Augmented Generation (RAG) functionality, allowing users to query course content using large language models. The plugin integrates with either Google's Gemini API, Ollama or OpenAI to provide intelligent responses based on your course data.

Features

• Implements RAG (Retrieval-Augmented Generation) architecture
• Integrates with Google's Gemini API and OpenAI
• Supports vector embeddings for semantic search
• Uses hybrid ranking with BM25 and cosine similarity
• Automatically processes and indexes course content
• Proper capability management for adding block instances

Vector Database Feature Implementation

• [x] Moodle Database
• [ ] ChromaDB
• [ ] Supabase

Ai Provider Feature Implementation

• [x] Gemini
• [x] OpenAI
• [x] Ollama
• [ ] Anthropic

How It Works

graph TB
    subgraph Content Layer
        A[Course Materials] --> B[Content Extractor]
        B --> C[Text Chunker]
    end

    subgraph Embedding Layer
        C --> D[AI Provider]
        D -->|Gemini| D1[Gemini Embeddings]
        D -->|OpenAI| D2[OpenAI Embeddings]
        D1 & D2 --> E{Vector Storage}
        E -->|Simple| F[Moodle DB]
        E -->|Scalable| G[ChromaDB]
        E -->|Cloud| H[Supabase]
    end

    subgraph Query Layer
        I[User Query] --> J[Query Processor]
        J --> K[Embedding Generation]
        K --> L[Vector Search]
        F & G & H --> L
    end

    subgraph Response Layer
        L --> M[Context Assembly]
        M --> N{AI Provider}
        N -->|Gemini| N1[Gemini API]
        N -->|OpenAI| N2[OpenAI API]
        N1 & N2 --> O[Response Formatter]
        O --> P[UI Display]
    end

    style Content Layer fill:#e1f5fe,stroke:#01579b
    style Embedding Layer fill:#f3e5f5,stroke:#4a148c
    style Query Layer fill:#e8f5e9,stroke:#1b5e20
    style Response Layer fill:#fff3e0,stroke:#e65100

AI Providers

This plugin supports two AI providers for handling RAG operations:

1. Google Gemini (Default)

   • Uses Gemini 2.0 Flash for chat completions
   • Supports text-embedding-004 for vector embeddings
   • Direct integration with Google AI Studio
   • Optimized for performance and cost-effectiveness

2. OpenAI (Alternative)

   • Supports GPT-4 and GPT-3.5 models
   • Multiple embedding model options:
     • text-embedding-3-large
     • text-embedding-3-small
     • text-embedding-ada-002
   • Flexible API endpoint configuration
   • Enterprise-grade capabilities

Installation

1. Copy the terusrag folder into your Moodle blocks directory
2. Visit the notifications page (Site administration → Notifications) to complete the installation
3. Configure the API keys in the block settings
4. Add the block to your course or Dashboard

Requirements

• Moodle 4.1.3 or later
• PHP 7.4 or later
• Access to either:
  • Gemini API (for default provider)
  • OpenAI API (for alternative provider)

Configuration

1. Go to Site Administration → Plugins → Blocks → Terus RAG
2. Choose your preferred AI provider (Gemini or OpenAI)
3. For Gemini:
   • Obtain API key from Google AI Studio
   • Configure Gemini endpoint and model settings
4. For OpenAI:
   • Obtain API key from OpenAI platform
   • Configure OpenAI endpoint and model preferences
5. Save changes and initialize the data process

Scheduled Tasks

The plugin includes scheduled tasks to maintain and update the vector embeddings of course content:

1. Content Indexing Task

   • Runs daily by default
   • Scans course content for changes
   • Updates vector embeddings for modified content

2. Running Tasks Manually

   • Via CLI:

     # List all tasks
     php admin/cli/scheduled_task.php --list
     
     # Run content indexing task
     php admin/cli/scheduled_task.php --execute=\\block_terusrag\\task\\datainitializer
     

Core Files

• provider_interface.php: Interface defining LLM provider capabilities
• gemini.php: Implementation of the Gemini API integration
• openai.php: Implementation of the OpenAI API integration
• bm25.php: BM25 ranking algorithm for text retrieval
• llm.php: Helper class with vector operations for LLM processing

Prompt Optimization

The plugin includes intelligent prompt optimization techniques to improve the quality of responses:

1. Key Term Emphasis: Automatically identifies and emphasizes important terms in user queries
2. Entity Recognition: Identifies proper nouns and technical terms to provide better context
3. Redundancy Removal: Eliminates unnecessary phrases that can dilute query effectiveness
4. Context Enhancement: Adds relevant context when queries lack sufficient specificity
5. Density Analysis: Calculates content-to-filler ratio to improve query precision

When enabled, this optimization process happens automatically when users submit queries, resulting in: - More accurate responses - Better focus on technical content - Reduced token consumption - Improved response consistency

Capabilities

• block/terusrag:addinstance - Controls who can add the block to a course
• block/terusrag:myaddinstance - Controls who can add the block to their Dashboard

Author

• Name: Khairu Aqsara
• Email: khairu@teruselearning.co.uk
• Website: https://teruselearning.co.uk

License

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.