FSW RAG Demo: Production-Grade Friction Stir Welding Defect Analysis

Enterprise-grade Retrieval-Augmented Generation (RAG) system for intelligent defect analysis in Friction Stir Welding (FSW) processes.

Features

• Hybrid Retrieval: BM25 + vector search with cross-encoder reranking
• Streaming Responses: Real-time LLM output with HuggingFace API
• Conversation Memory: Context-aware multi-turn conversations (last 10 turns)
• Memory Counter: Visual indicator showing remaining context window
• Sensor Data Integration: Real-time FSW sensor analysis with defect correlation
• Source Attribution: Full citation tracking with relevance scores
• Observability: Langfuse integration for tracing and monitoring
• Modular Architecture: Clean separation of concerns with 100% test coverage
• Production Ready: Error handling, validation, and graceful degradation
• Privacy-First: In-memory sessions, no persistent user data storage

Quick Start

Installation

# Clone repository
git clone https://github.com/SkullKrak7/RAG_Demo.git
cd RAG_Demo

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt
pip install -r requirements-dev.txt

# Setup git hooks
./setup-hooks.sh

Configuration

# Copy environment template
cp .env.example .env

# Edit .env with your credentials
# Required: HF_TOKEN
# Optional: LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY

Build Vector Store

# Place FSW PDF documents in data/ directory
python build_vectorstore.py --pdf-dir ./data --output-dir ./vectorstore

Run Application

streamlit run app.py

Access at http://localhost:8501

Architecture

System Components

┌─────────────────┐
│  Streamlit UI   │
└────────┬────────┘
         │
┌────────▼────────┐
│  RAG Pipeline   │
└────────┬────────┘
         │
    ┌────┴────┐
    │         │
┌───▼──┐  ┌──▼────┐
│Hybrid│  │Reranker│
│Retriever│  └───┬───┘
└───┬──┘      │
    │         │
┌───▼─────────▼───┐
│  LLM Generator   │
└──────────────────┘

Module Structure

rag_demo/
├── core/              # Configuration, models, exceptions
├── ingestion/         # Document loading and vectorstore building
├── retrieval/         # Hybrid retrieval and reranking
├── generation/        # LLM generation and response formatting
├── pipeline/          # End-to-end RAG orchestration
└── observability/     # Langfuse tracing integration

Usage

Basic Query

from rag_demo.core.config import RAGConfig
from rag_demo.ingestion.builder import VectorStoreBuilder
from rag_demo.retrieval.retriever import HybridRetriever
from rag_demo.pipeline.pipeline import RAGPipeline

# Initialize
config = RAGConfig()
builder = VectorStoreBuilder(config)
vectorstore = builder.load_vectorstore()

# Create retriever
retriever = HybridRetriever(vectorstore, documents, config)

# Create pipeline
pipeline = RAGPipeline(retriever, config)

# Query
response = pipeline.query("What causes wormhole defects in FSW?")
print(response.answer)

for source in response.sources:
    print(f"- {source.doc_name} (Page {source.page_num})")

Streaming Query

for chunk in pipeline.stream_query("Explain FSW process parameters"):
    print(chunk, end="", flush=True)

With Observability

from rag_demo.observability.tracer import RAGTracer

config = RAGConfig(langfuse_enabled=True)
tracer = RAGTracer(config)
pipeline = RAGPipeline(retriever, config, tracer=tracer)

response = pipeline.query("What are common FSW defects?")

# Feedback
tracer.score_feedback(1.0, "user_feedback")
tracer.flush()

Configuration

Environment Variables

Variable	Description	Default
HF_TOKEN	HuggingFace API token	Required
MODEL_NAME	LLM model identifier	meta-llama/Llama-3.1-8B-Instruct
TEMPERATURE	LLM temperature	0.05
RETRIEVAL_K	Documents to retrieve	5
RERANK_TOP_K	Documents after reranking	3
CHUNK_SIZE	Document chunk size	500
CHUNK_OVERLAP	Chunk overlap	50
LANGFUSE_ENABLED	Enable tracing	false

See .env.example for complete configuration options.

Development

Running Tests

# All tests
pytest tests/ -v --cov=rag_demo

# Specific module
pytest tests/unit/test_pipeline.py -v

# With coverage report
pytest tests/ --cov=rag_demo --cov-report=html

Code Quality

# Format code
black rag_demo/ tests/

# Lint
pylint rag_demo/

# Type checking
mypy rag_demo/

Project Status

Completed Features

• Modular RAG architecture
• Hybrid retrieval (BM25 + vector)
• Cross-encoder reranking
• Streaming LLM responses
• Source attribution and citations
• Langfuse observability
• Vector store builder
• Streamlit UI
• 100% test coverage on core modules

Roadmap

• CI/CD pipeline (GitHub Actions)
• Integration tests
• RAGAS evaluation framework
• Performance monitoring
• API documentation
• Docker deployment

Technical Specifications

Dependencies

• LangChain: 1.2.8
• Streamlit: 1.53.1
• ChromaDB: 1.4.1
• Sentence Transformers: 5.2.2
• Langfuse: 3.12.1
• PyPDF: 6.6.2

Models

• Embeddings: sentence-transformers/paraphrase-MiniLM-L3-v2 (384 dims)
• LLM: meta-llama/Llama-3.1-8B-Instruct (8B params)
• Reranker: cross-encoder/ms-marco-MiniLM-L-6-v2

Performance

• Query Latency: < 2s (with reranking)
• Embedding Speed: ~50ms per query
• Vector Store Size: ~500MB (5 documents)

License

MIT License - Copyright (c) 2026 Sai Karthik Kagolanu

Contact

• Author: Sai Karthik Kagolanu
• GitHub: @SkullKrak7
• Project: RAG_Demo