01 / STAR
Situation
RS Dr. M. Djamil Padang needed a fast, consistent way to conduct initial mental health screenings before patients spoke further with healthcare professionals.
HEALTHCARE · AI / NLP / BACKEND
PROJECT / 01
MindHarbor
Hybrid RAG Mental Health Screening System
A mental health screening chatbot that grew from a rule-based flow into a Hybrid RAG system.
- ROLE
- Backend architecture, retrieval pipeline
- PERIOD
- Research project
- TECH STACK
- FastAPI · PostgreSQL · Qdrant · Redis · Celery · Telegram · LLM API
PROJECT OVERVIEW
From need to outcome.
A clear project story: situation, task, action, then result.
02 / STAR
Task
Create an accessible initial screening tool that could guide users consistently and support the healthcare team before further consultation.
03 / STAR
Action
We built a chatbot using a rule-based system, NLP, a knowledge base, and decision-tree conversations. It was later expanded into a Hybrid RAG architecture that combines structured data with semantic retrieval.
04 / STAR
Result
The system was deployed to a VPS and is used by more than 120 healthcare professionals at RS Dr. M. Djamil Padang to support initial mental health screenings.
03 / PRODUCT GALLERY · 02 SCREENS
Explore the product.
Use the arrows, dots, or swipe to browse the project screens.
0102
MINDHARBOR / HERO SCREENSHOT
IMPLEMENTATION DETAILS
How the solution was built.
The key decisions behind the implementation.
BACKEND
FastAPI services coordinate screening, routing, and response generation.
RETRIEVAL
Qdrant embeddings and BM25 lexical search provide complementary context.
DATA
PostgreSQL holds deterministic screening state and structured rules.
INFRASTRUCTURE
Redis and Celery support the asynchronous retrieval workflow.
TECHNOLOGIES
FastAPIPostgreSQLQdrantRedisCeleryTelegramLLM API
EVIDENCE & METRICS
Numbers behind the result.
0.7039LaBSE F1
0.9234E5 COSINE
0.9350RAGAS FAITHFULNESS
100%QUERY ROUTER
06 / WHAT I LEARNED
What remains useful.
The project made the tradeoff between deterministic structure and contextual retrieval visible. The next improvement would be continuing to test retrieval quality against the real screening requirements.