Add support for multiple LLM providers (OpenAI & Gemini)

[subhranil2605]

Summary

Adds support for Google Gemini as an alternative LLM provider alongside OpenAI. Introduces a provider abstraction layer (llm-provider.ts) that allows switching between providers via environment configuration without changing application logic.

Key Changes

New Files

• src/gemini.ts (318 lines) — Gemini API integration with structured output support

  • createFormWithGemini() — Generate form structure from descriptions
  • updateFormWithGemini() — Update forms with user feedback
  • generateSuggestionsWithGemini() — Suggest form improvements
  • judgeFormWithGemini() — Evaluate form quality

• src/llm-provider.ts (107 lines) — Provider abstraction/router layer

  • Routes all LLM calls based on LLM_PROVIDER environment variable
  • Maintains provider-agnostic interface for routes
  • Extensible for future providers (Claude, etc.)

• src/utils/nunjucks-setup.ts (72 lines) — Extracted nunjucks configuration utility

  • Consolidates template engine setup logic
  • Shared between main server and zip-download server

Modified Files

• .env.example — Added LLM_PROVIDER=openai and GEMINI_API_KEY variables
• src/validationSchemas/env.ts — Added provider and Gemini API key validation
• src/routes/prototype-routes.ts — Updated 3 function calls to use llm-provider.ts
• server.ts & data/zip-download/server.ts — Refactored to use shared nunjucks utility
• package.json — Added @google/generative-ai dependency

Technical Details

Provider Abstraction Architecture

All LLM calls route through a single llm-provider.ts layer:
User Request
↓
llm-provider.createForm(envVars, ...)
↓
├─ If 'openai' → calls openai.ts functions
└─ If 'gemini' → calls gemini.ts functions
↓
Returns JSON (same schema, any provider)

• Both providers use identical JSON schemas from data/*-schema.json
• No changes required to route handlers
• Easy to extend with new providers

Configuration

Variable	Default	Required When
LLM_PROVIDER	openai	Never (defaults)
GEMINI_API_KEY	—	When using Gemini

Supported Providers

• OpenAI — GPT-4 mini, production-ready, ~$0.15 per 1M tokens input
• Gemini — Gemini 2.0 Flash, cost-effective, ~$0.075 per 1M tokens input

Testing

• ✅ All existing unit tests pass (628 tests)
• ✅ All E2E tests pass (Playwright)
• ✅ Provider abstraction maintains existing test coverage
• ✅ New code compatible with mocked LLM providers

Test results: No breaking changes, all tests pass with provider abstraction

Migration Path

Fully backward compatible:

• Default provider is OpenAI (no configuration changes needed)
• To use Gemini:
  1. Set LLM_PROVIDER=gemini
  2. Set GEMINI_API_KEY=your-api-key

3. Can rollback anytime by changing LLM_PROVIDER back to openai

Dependencies Added

• @google/generative-ai — Official Google SDK for Gemini API