🏡 Vaasthu Vision AI - AI-Powered Vaasthu Consultant

🚀 Project Overview

Vaasthu Vision AI is an intelligent GenAI system designed to provide authentic guidance, directional insights, and remedial suggestions based on Vaasthu Shastra. Unlike generic chatbots, it uses a Retrieval-Augmented Generation (RAG) architecture to ensure that every response is accurate, reliable, and context-aware.

The system carefully routes queries using similarity scores and critical keyword checks, preventing wrong or hallucinated answers while always prioritizing trustworthy knowledge from its Vaasthu knowledge base.

🎥 Project Demo

I’ve shared my project demo and explanation on LinkedIn — check them out below:

• 🎬 Demo Video (LinkedIn)

• 📖 Project Explanation (LinkedIn Post)

👉 Click to watch the full explanation or directly view the video output.

🎯 Objective

To build an AI assistant that:

• Understands Vaasthu rules deeply
• Provides reliable answers
• Avoids hallucinations
• Runs fast on web (connected to a slick frontend)

📊 RAGAS Evaluation Metrics

• Faithfulness: 0.9–1.0
• Answer Relevancy: ~0.7–0.85
• Latency: Measured and optimized for real-time responses
• Hallucination Rate: Reduced via retrieval grounding and fallback strategies

🛠️ Tech Stack

• Frontend: Built the website using bolt ai and customized as desired.
• Backend: Python (FastAPI / Streamlit for local)
• LLM: LLaMA3-8B-8192 via Groq API
• Vector DB: Qdrant with all-MiniLM-L6-v2 embeddings
• RAG: LangChain-powered pipeline

🧩 Data & Design Decisions

🔧 Data Transformation

• Started with 40 structured Vaasthu elements in JSON
• Converted to 350+ high-quality natural-language rules
• Added metadata: zone, rule_id, category
• Stored using RULE_START and RULE_END delimiters

🧠 Prompt Engineering

• Final prompt: minimalistic, 4 - 6 lines of answer.
• Designed for clarity, consistency, and production use
• Used temperature=0 and top_p=1 for deterministic output

⚙️ RAG Flow

1. Query → Embedding
2. Qdrant → Top 3 relevant rules
3. Custom prompt → Groq LLM (LLaMA3)
4. Final response → Displayed in UI

System Architecture:

graph TD
    A[User Question] --> B[Critical Keyword Check]
    B -->|Yes| C[RAG QA Chain]
    B -->|No| D[Vectorstore Retrieval]
    D --> E[Qdrant Similarity Score]
    E --> F{Confidence Thresholds}
    F -->|High| C
    F -->|Medium| G[`I don't know Response`]
    F -->|Low| H[Fallback Chat Chain]
    C --> I[Final Vaasthu Answer]
    G --> I
    H --> I

Loading

Query Router

This project implements a smart query routing system that decides whether a user’s question should be answered via RAG pipeline (vector database retrieval) or by an LLM fallback, based on similarity scores and critical keywords.

⚡ Workflow Overview

1. User enters a query
2. The system runs a similarity search on the vector database.
3. A similarity score is calculated for the top retrieved chunks.
4. Based on this score and rules, the query is routed:

🔎 Query Routing Logic

• Case 1: High Confidence (≥ HIGH_THRESHOLD)

  • ✅ Strong match found in vector DB
  • 👉 Response generated by RAG pipeline (qa_chain)

• Case 2: Low Confidence (< LOW_THRESHOLD)

  • ⚠️ Retrieved chunks are unreliable
  • 👉 Routed to LLM fallback, which replies:
  • "Sorry, I don’t have an idea about this query."

• Case 3: Critical Keywords Override

  • 🔑 Even if similarity score is below LOW_THRESHOLD,
  • If the query contains critical keywords (e.g., Kitchen, Bathroom, Hall),
  • 👉 Still answered through RAG pipeline (domain relevance guaranteed)

• Case 4: Nonsense / Out-of-Domain Queries

  • 🚫 No relevant match + no critical keywords
  • 👉 Routed to LLM fallback for casual/nonsense handling

🎯 Key Features

• Confidence-based Routing → prevents misleading answers from weak retrievals.
• Domain Awareness → critical keywords always prioritize vector DB results.
• LLM Fallback → handles nonsense or completely unrelated queries.
• Accuracy & Safety First → ensures only reliable information is returned.
• Contribute Feature → Allows users to submit data or upload files for review. After admin verification, contributions can be incorporated into the project.

🧠 Learnings

• Prompt structure impacts hallucination significantly
• Simpler is better: hardcoded rules > complex intent classifier
• Fast, reliable LLMs like Groq drastically improve UX
• Semantic granularity in rules increases RAG accuracy

🌐 Website of frontend only without backend integration:

🔗 Click here : Visit Site

or

website : https://jazzy-entremet-70cc2a.netlify.app/

💼 Ready for Production

The system is ready for real-world integration and can be expanded to:

• Multiple languages
• Room-by-room suggestions
• Vaastu-based house plan checker

🐳 Docker Setup

We provide a Docker image for Vaasthu Vision AI so you can run the app anywhere without installing dependencies.

Prerequisites

• Docker Desktop installed: https://www.docker.com/products/docker-desktop

• A .env file with:

  - QDRANT_URL=<your-qdrant-cloud-url>
  
  - QDRANT_API_KEY=<your-qdrant-api-key>
  
  - GROQ_API_KEY=<your-groq-api-key>
  

Run via Docker Hub Image:

Pull the latest image from Docker Hub:

• docker pull docker.io/shivaprasadnaroju/vaasthu-vision-ai:latest

Run the container with your .env:

• docker run -p 8000:8000 --env-file .env docker.io/shivaprasadnaroju/vaasthu-vision-ai:latest

Visit: http://localhost:8000/docs to access FastAPI Swagger UI.

Notes:

• The image is preconfigured to connect to Qdrant Cloud via .env.

• For a local Qdrant setup, a separate Docker Compose file can be used

🙌 Special Thanks

Inspired by traditional Indian architecture wisdom and empowered by modern AI.

📬 Connect with Me

I’m open to collaborations, feedback, or AI-based consulting.
📧 Email: shivanaroju26@gmail.com