jyotirmoy05/VivekanandaAI

Swami Vivekananda AI

An AI system embodying the wisdom and teachings of Swami Vivekananda, built with pure logic, no hardcoding, and fully config-driven architecture.

💻 Code & Datasets

The full implementation, scripts, and instructions to build the vector store and run inference (including RAG support) are available on GitHub: 👉 [GitHub: https://github.com/JyotirmoyDutta05/humanoid-AI] please copy the GitHub repository link and paster it on the search bar on your web browser to access the repository.

Features

• Config-Driven: Everything configurable via config.yaml
• NLP Processing: spaCy + NLTK for intelligent text processing
• RAG System: LangChain-based Retrieval-Augmented Generation
• MPS Support: Optimized for Apple Silicon (M4 Mac Mini)
• GGUF Models: Support for Mistral 7B quantized models
• Modular Design: Clean, maintainable, production-ready code

Requirements

• Python 3.9+
• Mac Mini M4 with MPS (or CUDA/CPU)
• 16GB+ RAM recommended
• 10GB+ free disk space

Quick Start

1. Clone & Setup

# Clone repository
git clone <your-repo-url>
cd vivekananda-ai

# Run automated setup
chmod +x setup.sh
./setup.sh

2. Add Your Data

# Add PDF files
cp /path/to/pdfs/* data/raw/complete_works/

# Add your JSON dataset
cp vivekananda_dataset_1.json data/processed/

# Add Mistral model
cp mistral-7b-instruct-v0.1.Q4_K_M.gguf models/base/

3. Configure

Edit config.yaml to customize:

• Model settings
• Embedding parameters
• RAG configuration
• NLP preprocessing options

4. Create Vector Database

source venv/bin/activate
python scripts/01_embed_data.py

Expected output: ```

VIVEKANANDA AI - VECTOR DATABASE CREATION

STEP 1: LOADING DOCUMENTS Found 9 PDF files Loading: SWAMI-VIVEKANANDA-VOL-1.pdf Loaded 234 pages ...

STEP 2: PROCESSING WITH NLP Processing documents with NLP... ...

SUCCESS! VECTOR DATABASE READY Total chunks embedded: 8,469

### 5. Test RAG

```bash
python scripts/02_query_rag.py

6. Run Streamlit App

streamlit run app.py

📁 Project Structure

vivekananda-ai/
│
├── config.yaml                 # Central configuration
├── requirements.txt            # Python dependencies
├── setup.sh                    # Automated setup script
│
├── utils.py                    # Core utilities
├── nlp_processor.py            # NLP processing (spaCy + NLTK)
│
├── scripts/
│   ├── 01_embed_data.py       # Create vector database
│   ├── 02_query_rag.py        # Test RAG retrieval
│   ├── 03_test_mistral.py     # Test Mistral model
│   └── 04_prepare_finetune.py # Prepare fine-tuning data
│
├── data/
│   ├── raw/                   # Original PDFs
│   ├── processed/             # JSON datasets
│   └── extracted_text/        # Extracted text files
│
├── vectorstore/
│   └── vivekananda_db/        # FAISS vector database
│
├── models/
│   ├── base/                  # Base models (GGUF)
│   └── fine_tuned/            # Fine-tuned adapters
│
├── outputs/
│   ├── logs/                  # Application logs
│   └── results/               # Results & analysis
│
└── app.py                     # Streamlit app

Configuration

All settings in config.yaml:

Model Settings

model:
  generation:
    max_tokens: 512
    temperature: 0.7
    top_p: 0.9

Embeddings

embeddings:
  model_name: "sentence-transformers/all-mpnet-base-v2"
  chunk:
    size: 500
    overlap: 50

NLP Processing

nlp:
  spacy:
    model: "en_core_web_sm"
  nltk:
    tokenizer: "punkt"
  preprocessing:
    lemmatize: true
    remove_stopwords: false

RAG

rag:
  retrieval:
    top_k: 5
    similarity_threshold: 0.5

How It Works

1. Document Loading

• Loads PDFs, text files, and JSON Q&A pairs
• Uses LangChain loaders for consistency
• Extracts Vivekananda-specific metadata

2. NLP Processing

• spaCy: Sentence segmentation, lemmatization, NER
• NLTK: Tokenization, stopword removal, stemming
• Config-driven preprocessing pipeline

3. Chunking

• Intelligent sentence-boundary-aware chunking
• Configurable size and overlap
• Preserves context

4. Embedding

• HuggingFace sentence transformers
• MPS-accelerated on Apple Silicon
• Normalized embeddings for semantic search

5. Vector Store

• FAISS for fast similarity search
• Persistent storage
• Metadata filtering support

6. RAG Retrieval

• Query → Embed → Retrieve top-k chunks
• Re-ranking (optional)
• Context assembly for LLM

7. Generation

• Mistral 7B GGUF (quantized)
• Context-aware prompts
• Vivekananda-style responses

Dataset

Your JSON dataset structure:

{
  "instruction": "What is Karma Yoga?",
  "response": "Karma Yoga is the path of selfless action...",
  "source": "Karma-Yoga.pdf",
  "work_type": "Karma-Yoga",
  "topic": "philosophy"
}

• 52 Q&A pairs (current)
• Expandable to 500-2000 for fine-tuning
• Rich metadata for filtering

🔧 Advanced Usage

Modify Embedding Model

Edit config.yaml:

embeddings:
  model_name: "BAAI/bge-small-en-v1.5"  # Faster alternative

Change Chunk Size

embeddings:
  chunk:
    size: 300  # Smaller chunks
    overlap: 30

Adjust NLP Processing

nlp:
  preprocessing:
    lowercase: true
    remove_stopwords: true
    lemmatize: true

Tune RAG Retrieval

rag:
  retrieval:
    top_k: 10  # More context
    similarity_threshold: 0.3  # Lower threshold

Testing

Test Embeddings

python scripts/01_embed_data.py

Test RAG Retrieval

python scripts/02_query_rag.py "What is Karma Yoga?"

Test Mistral Model

python scripts/03_test_mistral.py

Interactive Testing

streamlit run app.py

Performance

On Mac Mini M4:

Task	Time	Memory
Embedding 9 volumes	10-15 min	4-6 GB
Query retrieval	50-200 ms	2-3 GB
Model inference	1-3 sec	6-8 GB

Troubleshooting

spaCy Model Not Found

python -m spacy download en_core_web_sm

NLTK Data Missing

import nltk
nltk.download('punkt')
nltk.download('stopwords')

MPS Not Available

Check config.yaml:

hardware:
  device: "cpu"  # Fallback to CPU

Out of Memory

Reduce batch size in config.yaml:

embeddings:
  batch_size: 16  # Smaller batches

Next Steps

1. Fine-Tuning: Train on your 52 Q&A pairs
2. Expand Dataset: Create 500+ more examples
3. Evaluation: Test response quality
4. Deployment: FastAPI → AWS
5. Frontend: Build user interface

Resources

• LangChain Documentation
• spaCy Documentation
• NLTK Documentation
• FAISS Documentation
• Mistral Documentation

Contributing

1. Fork the repository
2. Create feature branch
3. Make changes
4. Test thoroughly
5. Submit pull request

License

MIT License - See LICENSE file

Acknowledgments

• Swami Vivekananda's teachings
• Ramakrishna Mission for digitized texts
• LangChain community
• HuggingFace team

---

Built with for spreading Vivekananda's wisdom

For questions: Create an issue

Vivekananda AI – High-Fidelity Persona Fine-Tuning Pipeline

This repository includes a production-ready pipeline to build a high-fidelity, persona-aligned AI system in the voice of Swami Vivekananda using a Mistral architecture.

Features

• Supervised fine-tuning (SFT) with LoRA/QLoRA (Transformers + Accelerate + PEFT)
• Preference optimization via DPO (TRL)
• Optional RAG interface (FAISS + sentence-transformers)
• Strict persona system prompt enforcing tone and guardrails
• Inference with probabilistic decoding and token probability tracing
• Dockerfile for GPU training environment

Repository Additions

• training/data_preprocess.py: Convert processed dataset to JSONL instruction format
• training/sft_train.py: SFT training with LoRA/QLoRA
• training/dpo_train.py: DPO preference optimization
• training/export_adapter.py: (optional) merge/unload adapters – to be added if needed
• inference/generate.py: Persona inference with sampling controls and token probabilities
• rag/build_index.py, rag/retrieve.py: Optional retrieval augmentation
• configs/train_sft.yaml, configs/train_dpo.yaml: Training configs
• prompts/system_vivekananda.txt: Persona system prompt
• docker/Dockerfile: GPU training environment

Quick Start

1. Preprocess dataset to JSONL

   • python training/data_preprocess.py --source data/processed/vivekananda_dataset_1.json --out data/datasets/sft_train.jsonl --val data/datasets/sft_val.jsonl --val-ratio 0.05

2. SFT training (LoRA/QLoRA)

   • Ensure GPU environment (Docker provided below)
   • accelerate launch training/sft_train.py --config configs/train_sft.yaml --train data/datasets/sft_train.jsonl --val data/datasets/sft_val.jsonl --output runs/sft_vivekananda

3. DPO preference optimization

   • Prepare preference dataset JSONL with {prompt, chosen, rejected}
   • accelerate launch training/dpo_train.py --config configs/train_dpo.yaml --prefs data/datasets/dpo_prefs.jsonl --sft-checkpoint runs/sft_vivekananda/step_XXXX --output runs/dpo_vivekananda

4. Optional RAG

   • Build index: python rag/build_index.py --source data/processed/vivekananda_dataset_1.json --out vectorstore/faiss_index
   • Retrieve for inference: python rag/retrieve.py --index vectorstore/faiss_index --query "Your question"
   • Pass concatenated context into inference via --rag-context

5. Inference with persona and probabilistic decoding

   • python inference/generate.py --model mistralai/Mistral-7B-Instruct-v0.2 --adapter runs/sft_vivekananda/step_XXXX --prompt "Question" --temperature 0.3 --top_p 0.85 --top_k 40 --max_tokens 512 --rag-context "<CONTEXT>..."

Docker (GPU Training)

• Build: docker build -t vivekananda-ai:train -f docker/Dockerfile .
• Run: docker run --gpus all -it -v $(pwd):/workspace vivekananda-ai:train bash

Notes

• Ensure bitsandbytes detects GPU; use CUDA-compatible base.
• For strict persona enforcement, customize prompts/system_vivekananda.txt.
• Avoid fabrication; when facts are missing, the prompt instructs the model to say so plainly.