Integrate LangChain for type-safe AI orchestration

[Copilot]

Replaces manual JSON parsing and ad-hoc prompt management with LangChain's structured output and composable chains. Adds Pydantic validation to eliminate runtime parsing errors.

Core Changes

Structured Output (src/ai_utils.py, src/models.py)

• call_gemini_structured(): Returns validated Pydantic models instead of raw JSON
• Pydantic schemas define and enforce all data structures (SiteData, Product, DesignSystem, etc.)
• Automatic retry logic (max_retries=2) via ChatVertexAI

Prompt Templates (src/langchain_prompts.py)

• Centralized PromptTemplate objects replace string formatting
• Reusable chains: component_chain, page_chain, layout_chain, css_chain
• Single source of truth for all AI interactions

Website Generator (src/website_generator.py)

• Refactored to use composable chains
• Progress tracking integrated into generation flow
• Unified error handling across all generation steps

Before/After

# Before: Manual parsing, fragile
response = call_gemini(prompt)
cleaned = response.strip().removeprefix("```json")
data = json.loads(regex.match(cleaned))  # May fail

# After: Type-safe, validated
data: SiteData = call_gemini_structured(
    prompt=prompt,
    pydantic_model=SiteData
)

Testing

• test_langchain.py: 5/5 tests passing
• CodeQL: 0 vulnerabilities
• Backward compatible: call_gemini() maintained

Dependencies

langchain
langchain-google-vertexai
langchain-core
pydantic

Future Enablement

LangChain integration enables straightforward addition of:

• Memory for conversational iteration
• Agents for multi-step reasoning
• RAG for design pattern retrieval
• Multi-model fallback strategies

Original prompt

how can langchain improve this project

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.