AIQuest

Secured 3rd place at IIT Bombay Techfest's AIQuest Hackathon judged by Marsh McLennan panel

AIQuest is an advanced internal knowledge-sharing platform, inspired by Stack Overflow, designed to revolutionize collaboration and productivity within organizations. Leveraging Retrieval-Augmented Generation (RAG) with the Llama AI model, AIQuest facilitates intelligent, contextual responses to user queries by combining the power of AI with human expertise.

---

Core Capabilities

Knowledge Sharing

• Employees can ask questions, provide answers, and utilize AI-generated responses.

AI-Generated Replies

• Contextual responses are derived from:
  • Company documents stored in a hybrid RAG.
  • Previously asked questions and official answers stored in a simple RAG.
  • Wikipedia, when internal sources lack relevant information.

Role-Based Access Control (RBAC)

• Provides secure and hierarchical access for employees, moderators, and admins.

Voting Mechanism

• Users can like questions and upvote or downvote answers, influencing search ranking and visibility.

Search Facility

• Allows users to search questions based on keywords or context.

Filtering

• Filters questions by ranking (likes) or most recent.

Official Answers

• Moderators can:
  • Mark an answer as official.
  • Provide official answers stored in the simple RAG for AI use.

Organizational Features

• Organizations can:
  • Register and invite employees via email, including invite codes.
  • Gain insights through a dashboard:
    • Metrics: Total users, questions, answers, registered users, and invited users.
    • Recent questions and uploaded documents.
    • Activity graph (questions/answers vs. date).
    • Visualization of top 5 tags.
  • Upload company documents to the hybrid RAG.
  • Review and delete inappropriate questions.

Moderation of Responses

• Moderation API checks for abusive or toxic content in both user-generated and AI responses.

Moderator Features

• Moderators can:
  • Post questions, give official answers, or mark answers as official.
  • View metrics such as total users, questions, and answers.
  • Moderate questions (e.g., delete inappropriate ones, mark answers as official).
  • Visualize moderation status through a pie chart.

User Features

• Users can:
  • Post questions (moderated for toxicity) and receive AI-generated answers.
  • Answer questions, like questions, and upvote or downvote answers.

---

Tech Stack

• Backend Framework: Flask
• Database Systems: SQLite for structured data, PostgreSQL for dashboard data.
• RAG Database: Elasticsearch for hybrid and simple RAG.
• AI Integration: Llama3.2 (locally using Ollama).
• APIs: Flask-powered APIs for seamless interactions.
• AI Orchestration: LangChain and PyTorch.
• Libraries: PyPDF2 for document parsing and chunking.
• Search Ranking: BM25 (sparse vectors) and all-MiniLM-L6-v2 embeddings (dense vectors).

---

Architecture and Workflow

Hybrid RAG

• Documents are parsed, chunked, and stored using:
  • Sparse vectors for syntactic search (BM25 ranking).
  • Dense vectors for semantic search (all-MiniLM-L6-v2 embeddings).
• Combined results from syntactic and semantic searches are ranked using Reciprocal Rank Fusion (RRF) for optimal output.

Simple RAG

• Stores embeddings of questions and their official answers.

Organizational Workflow

1. Document Uploads:
   • Parsed and read via PyPDF2.
   • Stored as dense vectors in the hybrid RAG (BM25 supports syntactic search natively in Elasticsearch).
2. Dashboard Insights:
   • Metrics, graphs, and visualizations (e.g., activity trends and top tags).

Moderator Workflow

• Marking or providing official answers:
  • Embedded and stored in the simple RAG.

User Workflow

1. Posting Questions:
   • Questions are checked for toxicity using the Moderation API.
   • Stored in SQLite.
   • Processed by the hybrid RAG to retrieve relevant documents (syntactic and semantic search).
   • If no results are found, Wikipedia is queried for context.
2. Answer Generation:
   • Llama3.2 generates contextual responses based on retrieved documents or Wikipedia context.
   • Responses are moderated for toxicity and stored in the database.

AI Training Pipeline

1. Aggregates data from the RAG.
2. Filters high-quality content based on user votes and admin validations.
3. Retrains the Llama3.2 model monthly and updates the RAG database.

---

Key Features of RAG

Hybrid RAG (Documents)

• Supports both syntactic (BM25) and semantic (dense vector) search.
• Efficiently retrieves and ranks documents for AI response generation.

Simple RAG (Questions & Official Answers)

• Stores embeddings of user questions and moderator-verified official answers.
• Enables accurate contextual retrieval during AI response generation.

---

Conclusion

AIQuest combines state-of-the-art AI technologies and robust role-based features to foster a collaborative, secure, and efficient knowledge-sharing environment. By integrating advanced RAG-based workflows and an intuitive interface, the platform addresses diverse organizational needs, empowering employees and enhancing productivity.