Problem what We wanted to Solve : Tech Stack

Frontend: React + TypeScript (Vite), TailwindCSS, shadcn/ui (Radix), TanStack React Query for API data fetching/caching. Backend: Node.js + TypeScript + Express. Single dev server runs API + Vite (frontend) together in development. Database: PostgreSQL with Drizzle ORM (drizzle-orm, drizzle-kit) for schema + migrations/push. Maps: Leaflet / react-leaflet for rendering provider locations on the map. Optional AI: Google Gemini embeddings via @google/generative-ai.

How AI Is Used (Search)

Purpose: turn a natural-language query (e.g., “tutor in State College for python”) into an embedding vector, then find the most semantically similar provider profiles. Embeddings model: Gemini text-embedding-004. When a profile is created, we embed: " - ". When searching, we embed the user query and compute cosine similarity between the query vector and each stored profile vector. Resilience / fallback: If GEMINI_API_KEY is missing or embedding fails, the API falls back to keyword/token matching across name, serviceType, description, and location. API always returns a small Top-K list (Top 3) and avoids returning an empty list where possible.

Database Structure (Postgres + Drizzle) Single main table: profiles Columns: id (serial, primary key) name (text, required) service_type (text, required) — exposed in app code as serviceType description (text, required) location (text, required) lat (text, optional) lng (text, optional) embedding (json, optional) — stored as number[] (vector) for semantic search Insert behavior: Client sends an InsertProfile without id and without embedding. Server computes embedding (if AI enabled) and stores it along with the profile row.