Medical-AI-Chatbot - Visual Studio Code
Built with production-grade technologies
Pinecone
Groq
AWS
Technical Architecture
Built with production-grade RAG pipeline featuring hybrid retrieval, cross-encoder reranking, and cloud-native deployment on AWS infrastructure.
RAG-Powered Accuracy.
92% answer relevancy with source-grounded medical responses.
Hybrid Retrieval Pipeline
MMR + BM25 ensemble with cross-encoder reranking for precision.
AWS Cloud Deployment
Docker containers on EC2 with GitHub Actions CI/CD pipeline.
Vector Search with Pinecone
384-dim embeddings for semantic medical document retrieval.
Llama 3.3 70B Generation
Context-only LLM responses with 88.7% faithfulness score.
Comprehensive Testing Suite
153 tests covering unit, integration, security, and performance.
Intelligent Medical Assistant
Beautiful, responsive interface with light and dark modes. Get evidence-based medical information with source citations.
Light Mode
Dark Mode
<5sResponse Time
92%Accuracy
YesSource Citations
2Themes
Open Source
Hybrid Retrieval Pipeline
Production-grade RAG architecture combining multiple retrieval strategies for maximum accuracy and relevance.
View on GitHub01
Query Rewriting
Llama 3.1 8B resolves pronouns and expands medical terminology for precise retrieval.
# Pronoun Resolution "What are its side effects?" → "What are the side effects of metformin?"
02
Hybrid Retrieval
MMR (60%) + BM25 (40%) ensemble combines semantic understanding with keyword matching.
ensemble_retriever = EnsembleRetriever( retrievers=[mmr_retriever, bm25_retriever], weights=[0.6, 0.4] )
03
Cross-Encoder Reranking
ms-marco-MiniLM-L-6-v2 scores query-document pairs for final relevance ranking.
reranker = CrossEncoder( "cross-encoder/ms-marco-MiniLM-L-6-v2" ) top_docs = reranker.rerank(query, docs, top_n=8)
04
Context-Only Generation
Llama 3.3 70B generates responses strictly grounded in retrieved context.
system_prompt = """Answer ONLY from the provided context. If information is not in the context, say 'I don't have enough information to answer that.'"""
Context-Aware Chunking
Optimized document segmentation preserves semantic coherence for medical content.
Chunk 1
Chunk 2
Chunk 3
100-token overlap ensures context continuity
Chunk Size
800 tokens
Optimal for medical paragraphsChunk Overlap
100 tokens
Prevents information loss at boundariesEmbedding Model
all-MiniLM-L6-v2
384-dimensional vectorsVector Index
Pinecone
Serverless, high-performanceSession Memory
Maintains last 10 exchanges (20 messages) for pronoun resolution only.
max_session_messages: 12
Temperature Control
Set to 0 for deterministic, factual medical responses without creativity.
temperature: 0
Hallucination Prevention
Strict context-only mode reduces hallucinations from 35% to 5%.
hallucination_rate: 5%
System Architecture
Production-grade RAG pipeline with hybrid retrieval, cross-encoder reranking, and cloud-native deployment
CI/CD Pipeline Architecture
GitHub Actions with Docker, ECR, and EC2 deployment
RAGAS Evaluation Metrics
92.33%Answer RelevancyRAGAS Score
88.78%FaithfulnessContext Adherence
90.00%Context RecallDocument Coverage
4.2sResponse TimeUnder 5s SLA
5%Hallucination RateDown from 35%
100%Test Pass Rate153 Tests
RAG Pipeline Flow
0.8s
1
Query Rewrite
Llama 3.1 8B resolves pronouns and expands queries
1.2s
2
Hybrid Retrieval
MMR + BM25 ensemble (0.6/0.4 weights)
0.3s
3
Cross-Encoder Rerank
ms-marco-MiniLM-L-6-v2 selects top-8 docs
2.1s
4
Response Generation
Llama 3.3 70B with context-only mode
Technology Stack
Flask/PythonBackend
LangChainOrchestration
PineconeVector DB
Groq APILLM Inference
HuggingFaceEmbeddings
AWS EC2Deployment
DockerContainers
GitHub ActionsCI/CD
Key Configuration Parameters
chunk_size800
chunk_overlap100
embedding_dim384
mmr_k10
mmr_fetch_k30
lambda_mult0.5
reranker_top_n8
max_session_msgs12
temperature0
Technical Documentation
Comprehensive documentation covering architecture, testing, deployment, and evaluation metrics
Software Requirements Specification
IEEE 830-1998 compliant SRS document covering functional and non-functional requirements, system architecture, and use cases.
27 Functional Requirements12 Non-Functional RequirementsIEEE Standard Format
RAG Architecture
Complete documentation of the Retrieval-Augmented Generation pipeline including document indexing, query processing, and response generation.
Document Indexing PipelineHybrid Retrieval SystemCross-Encoder Reranking
CI/CD Pipeline
GitHub Actions workflow with Docker containerization, Amazon ECR registry, and automated EC2 deployment.
Automated TestingDocker ContainerizationAWS Deployment
RAGAS Evaluation Report
Comprehensive evaluation using the RAGAS framework measuring faithfulness, relevancy, precision, and recall metrics.
92.33% Answer Relevancy88.78% Faithfulness90% Context Recall
Testing Documentation
Complete test suite documentation covering unit, integration, security, performance, and system workflow tests.
153 Total Tests100% Pass Rate9 Test Categories
Security Testing
Security vulnerability testing including XSS prevention, SQL injection protection, and API key security validation.
XSS ProtectionSQLi PreventionAPI Security
Project Highlights
Problem Solved
- •Information overload in medical queries reduced through curated retrieval
- •Hallucination rate reduced from 35% to 5% using RAG architecture
- •Context loss in conversations handled via pronoun resolution
- •Source attribution enables response verification
Key Achievements
- •92% accuracy vs 65% for non-RAG systems (+27% improvement)
- •100% SLA compliance with all queries under 5 seconds
- •Production-ready with Gunicorn, health endpoint, and Docker
- •Automated CI/CD with GitHub Actions and AWS deployment
Frequently Asked Questions
Everything you need to know about Medi-Query's RAG architecture and technical implementation
What is Medi-Query and how does it work?
How accurate is the RAG pipeline?
What technology stack powers Medi-Query?
How is the system deployed?
What testing coverage does Medi-Query have?
How does the conversation memory work?