How srinidevar/semantic-image-search works
Project type and purpose
• This is a containerized, full-stack semantic image-search application.
• It ingests images, computes CLIP embeddings, stores them in Milvus (a vector database), and provides a VueJS UI for text-based searches.
Main components
a. UI Container (VueJS frontend)
b. API Container (FastAPI service)
c. Image Server Container (uploads, stores, and serves raw images)
d. Vector DB (Milvus)
e. Metadata Store (etcd) and Object Store (MinIO) – Milvus dependencies
f. CLIP Model (loaded inside the API service)
Relationships and data flows
1) User → Browser/UI: loads static assets (HTML/CSS/JS).
2) UI → API over HTTP:
– Index Image requests (multipart/form-data uploads)
– Search queries (text)
3) API → Image Server:
– POST new image for storage
– GET image for display
4) API (upon image upload) → CLIP Model → generates embedding → Milvus (vector insert)
5) API (upon search) → CLIP Model → text embedding → Milvus (vector search) → returns matching image IDs
6) API → MinIO/etcd: Milvus’s object store and metadata operations
7) API → Image Server: to retrieve actual image bytes for display in results
Architectural patterns & design principles
• Microservices: each major domain (UI, API, image server, vector DB) runs in its own container.
• Separation of concerns:
• UI handles presentation and HTTP calls
• API handles business logic, CLIP model, DB orchestration
• Image Server isolates raw image storage
• Milvus abstracts vector indexing/search
• Docker-Compose for orchestration
• Asynchronous processing model for indexing (uploads → background embedding → DB insert)
Technologies & frameworks
• VueJS + Vite → UI
• FastAPI + Python → API
• OpenCLIP model (Python library) → Embedding
• Milvus → Vector database
• etcd, MinIO → Milvus dependencies
• Docker + Docker-Compose → Deployment
Detail without over-splitting
• Six top-level containers.
• Inter-container calls and data stores.