Home screen
chat interface
follow up questions & step thread
follow up answer
followup answer sources

ProCheck: AI-Powered Medical Protocol Assistant

Inspiration

Imagine a doctor in an emergency room searching for "fever from bug bites" — but the official protocol is titled "Malaria Treatment Guidelines." Traditional keyword search fails, wasting precious minutes.

This gap between how doctors think and how protocols are stored inspired ProCheck — an AI-driven system that understands medical intent, not just keywords. Using Elasticsearch's hybrid search and Google Gemini AI, it makes clinical knowledge instantly accessible through natural language.

What We Built

ProCheck is a semantic medical protocol search and checklist generator that helps healthcare professionals find and apply the right information in seconds.

Core Technologies

Elasticsearch (Hybrid Search) – Combines BM25 keyword matching with vector similarity using Reciprocal Rank Fusion (RRF):

RRF Score = ∑(r∈R) 1/(k + rank_r(d))

Ensures documents relevant by meaning or keyword both rise to the top.

Google Gemini AI – Generates embeddings, expands queries with medical terminology, and powers conversational answers with citations.

FastAPI – Async Python backend handling search orchestration.

React + TypeScript – Real-time, type-safe frontend with streaming AI chat.

Firebase – Authentication and user data management (saved protocols, chat history).

How It Works

Query Expansion: Gemini enhances queries with medical terms.
Semantic + Keyword Search: BM25 and vector search run in parallel.
RRF Fusion: Combines results for optimal precision and recall.
Conversational Layer: Users can ask follow-ups or generate checklists — every response includes verified citations.

How We Built It

Architecture

A modern three-tier setup:

React frontend → FastAPI backend → Elasticsearch, Gemini AI, Firebase

Development Journey

Week 1: Setup core stack and fixed Elasticsearch's new client issues.
Week 2: Implemented hybrid search with Gemini embeddings. Added caching (LRU + Firestore) cutting query latency from 3–5s to under 500ms.
Week 3: Built conversational AI with context summarization to stay within token limits.
Week 4: Optimized UI with react-window for smooth scrolling and optimistic updates for instant feedback.
Week 5: Added document upload with chunked async embedding for large PDFs.

Key Learnings

Hybrid Search Requires Tuning: BM25 excels for exact terms, vector search for conceptual ones. Proper RRF weighting was crucial.
Cache Everything Expensive: Embedding generation added 150ms per query — caching cut latency and API costs by 80%.
Balance Context in Chat: Keeping the last three messages plus summaries maintained quality without exceeding limits.
Speed Beats Features: Healthcare users prioritized instant, reliable search over UI flair.

Challenges

Index Mapping Errors: Missing dense_vector field forced zero-downtime reindexing.
Chat Race Conditions: Fixed with optimistic UI updates and reconciliation.
Auth Failures: Solved by adding Firebase token interceptors to API requests.
Rate Limits: Mitigated Gemini's constraints with exponential backoff and jitter.

Impact

In testing with healthcare professionals, ProCheck cut protocol lookup time from 3–5 minutes to under 10 seconds, improving first-try accuracy from 60% to 85%. Citations boosted trust — users could verify every claim instantly.

Where It Helps Most

Emergency Medicine: Fast retrieval during critical care.
Medical Education: Interactive learning via chat.
Research & Telemedicine: Cross-comparison of guidelines and patient-specific checklists.

Beyond Healthcare

The same hybrid-search architecture can power:

Legal research
Engineering documentation
Scientific literature
Regulatory compliance

Any domain needing concept-based retrieval benefits from this approach.

Performance Metrics

Metric	Value
Search latency	450 ms (cached)
Cache hit rate	85%
Indexed protocols	130+
Concurrent users	5+
Uptime	99.8%
Handles	100+ chat turns smoothly

Dataset Summary

Total protocols: 134 (infectious, chronic, emergency, pediatric)
Sources: WHO, CDC, NHS, MoHFW, AHA, Mayo Clinic
Schema:

Key	Data Type	Description (Implied)
disease	`string`	The specific disease or condition covered (e.g., "Malaria", "COVID-19").
region	`string`	The geographical region or country the protocol applies to.
year	`number`	The publication or effective year of the protocol.
organization	`string`	The issuing authority (e.g., "WHO", "Ministry of Health").
title	`string`	The official title of the protocol or guideline.
section	`string`	A specific section within the larger protocol (e.g., "Diagnosis", "Treatment").
body	`string`	The main content or text of the protocol section.
source_url	`string`	A URL link to the original source document.
last_reviewed	`string`	The date the record was last checked or updated.

Conditions Covered in the Dataset

The dataset encompasses 134 protocol records and over 35 unique diseases/groups, organized into the following categories:

Category	Conditions
Vector-Borne Infectious	Malaria, Dengue, Zika, Chikungunya, Yellow Fever
Respiratory Infectious	COVID-19, Influenza, Tuberculosis, Pneumonia, Measles, Meningitis, Bronchitis
Gastrointestinal/Infectious	Gastroenteritis, Hepatitis, Sepsis, Urinary Tract Infection
Chronic Diseases	Diabetes, Hypertension, Asthma, COPD, Obesity, Hypothyroidism
Cardiovascular/Emergencies	Heart Attack, Stroke, Cardiac Arrest, Angina, Atrial Fibrillation
Pediatric Conditions	Fever, Dehydration, Asthma, Ear Infection, Food Allergy (children)
Other Infectious/General	Choking, Burns, Poisoning, Severe Bleeding, Anaphylaxis, Concussion, Seizure
Preventive	Vaccination Schedule