DawaAI

Healthcare professionals in West Africa face unique challenges when accessing pharmaceutical information:

• Connectivity Issues: Unreliable internet in remote clinics limits real-time drug lookups
• Information Overload: Thousands of complex pharmaceutical documents are difficult to navigate
• Time Pressure: Busy clinics need instant access to dosing, contraindications, and safety information

Dawa AI solves these problems by providing a fast, intelligent search system that works offline and delivers precise pharmaceutical information in seconds, not minutes.

✨ Key Features

🔍 Hybrid Search Engine

• Semantic Understanding: BGE-Large embeddings capture contextual meaning ("pediatric dose" matches "children dosing")
• Keyword Precision: BM25 ensures exact terms aren't missed ("100mg" finds exact dosages)
• Reciprocal Rank Fusion: Optimally combines semantic and lexical results for maximum relevance

🌐 Offline-First Architecture

• Pre-computed Embeddings: All 37k document sections embedded locally (no internet required for search)
• Instant Startup: Ready to serve queries in under 5 seconds

📊 Smart Filtering & Search

Filter by Therapeutic Area, Active Substance, direct ATC Classification: Pharmaceutical classification codes, Section-Specific

🔗 Direct Source Access

• Citation Links: Every result links directly to source EMA documents
• Section References: Precise section numbers (4.2 Posology, 4.3 Contraindications)
• Relevance Scoring: Transparent BM25, semantic, and hybrid scores

🚀 Production-Ready

• FastAPI Backend: High-performance async API
• Docker Containerized: Consistent deployment across environments
• Cloud Native with CLI Support: Google Cloud Run with auto-scaling

Phase 2: Enhanced Intelligence

• Multi-language Support: French language search for Francophone West Africa
• Agentic Search: A simple text query runs filtered search
• Voice Search: Audio queries for hands-free operation in clinical settings

Tech Stack

• Backend : FastAPI, Uvicorn
• Search & ML : Sentence Transformers: BAAI/BGE-Large-EN-v1.5, rank-bm25: Okapi BM25, NumPy
• Data Processing: PyMuPDF: PDF text extraction and document parsing
• Infrastructure: Docker, uv(Astral), GCP: Google Cloud Run, Artifact Registry

Design Decisions

1. Hybrid Search Architecture & RRF vs Weighted Average

Decision: Combine semantic (BGE) + lexical (BM25) search with Reciprocal Rank Fusion

Rationale: Semantic-only misses exact medical terms (drug names, dosages) Keyword-only misses contextual relationships (synonyms, related concepts) RRF fusion empirically outperforms weighted averaging by 15-23%

Trade-offs:

• ✅ Pro: Best retrieval accuracy for medical queries
• ✅ Pro: Handles both precise and conceptual searches
• ✅ Pro: No hyperparameter tuning required
• ✅ Pro: Robust across different query types
• ✅ Pro: Well-established in information retrieval research
• ❌ Con: Slightly more complex to implement than weighted sum
• ❌ Con: Less intuitive than score-based combination
• ❌ Con: 2x computational overhead vs. single method
• ❌ Con: More complex caching and optimization

Alternative Considered: Linear combination (α × semantic + (1-α) × BM25) Rejected: Requires tuning α, sensitive to score normalization

2. Pre-computed Document Embeddings

Decision: Embed all documents offline, only embed queries in real-time Rationale:

BGE-Large inference: 50ms per query vs. 20 minutes for full corpus Offline requirement: No internet needed after initial deployment Cost optimization: Avoid repeated embedding of static documents

Trade-offs:

• ✅ Pro: Instant search without model loading
• ✅ Pro: Works completely offline
• ✅ Pro: Predictable query latency
• ❌ Con: Large deployment artifacts (6GB total)
• ❌ Con: Document updates require full re-deployment

Alternative Considered: Real-time document embedding Rejected: 1000x slower, requires GPU, defeats offline-first goal

3. Docker + Cloud Run Deployment

Decision: Containerize with Docker, deploy on Google Cloud Run Rationale: Serverless scaling: 0-10 instances Global reach & User Testing: Multi-region deployment for low latency Cost efficiency: Pay only for actual usage (~£5/month) Developer experience: Simple deployments with gcloud run deploy

Trade-offs:

• ✅ Pro: Zero infrastructure management
• ✅ Pro: Automatic scaling and load balancing
• ✅ Pro: Built-in HTTPS, monitoring, and logging
• ❌ Con: 10s second cold starts (mitigated with min instances)
• ❌ Con: Vendor lock-in (mitigated by standard containers)

4. BGE-Large vs. Smaller Models

Decision: Use BAAI/BGE-Large-EN-v1.5 despite 1.3GB size Rationale:

• Medical Domain: BGE-Large trained on scientific/medical text
• Embedding Quality: 1024 dimensions capture nuanced relationships
• Benchmark Results: 5-7% better retrieval accuracy vs. all-MiniLM-L6-v2
• Size: 1.3GB acceptable for cloud deployment

Trade-offs:

• ✅ Pro: Best-in-class retrieval quality for medical queries
• ✅ Pro: Handles complex pharmaceutical terminology
• ❌ Con: Large Docker images (6GB vs. 300MB with MiniLM)
• ❌ Con: Higher memory requirements (8GB vs. 2GB)

Alternative Considered: all-MiniLM-L6-v2 (80MB)