Feature Request: Add Structured Output Support for Gemini API

Problem Statement

Currently, the hiring-agent relies on manual JSON parsing from LLM responses, which can be error-prone and inconsistent. The system extracts JSON from markdown code blocks and performs text cleanup, leading to:

• Parsing failures when LLMs return malformed JSON or unexpected formatting
• Inconsistent responses due to variations in LLM output formatting
• Additional processing overhead from text cleanup and boundary detection
• Reduced reliability in production environments

Proposed Solution

Implement structured output support using Google Gemini's native structured output capabilities. This feature would:

✅ Benefits

• 100% JSON parsing reliability - Gemini API guarantees valid JSON structure
• Improved response consistency - Schema validation ensures exact format compliance
• Better performance - Direct JSON parsing without cleanup overhead
• Enhanced accuracy - Type safety and required field enforcement
• Reduced token usage - No need for verbose JSON formatting instructions

🔧 Technical Approach

• Integrate with Gemini's response_schema parameter using existing Pydantic models
• Maintain backward compatibility with Ollama's traditional JSON parsing
• Add automatic provider capability detection
• Preserve existing prompt templates and evaluation logic

📊 Expected Impact

• Eliminate JSON parsing errors for Gemini users
• Faster processing due to reduced text cleanup
• More reliable resume evaluation pipeline
• Better developer experience with consistent outputs

Implementation Details

The feature would:

1. Detect provider capabilities automatically
2. Pass Pydantic schemas to Gemini API as response_schema
3. Handle dual parsing modes (structured for Gemini, traditional for Ollama)
4. Maintain full backward compatibility with existing configurations

Use Cases

This enhancement benefits:

• Production deployments requiring high reliability
• High-volume processing where consistency matters
• Gemini API users wanting optimal performance
• Developers seeking predictable LLM responses

Acceptance Criteria

• Resume evaluation uses structured output for scoring
• Automatic fallback to traditional parsing for unsupported providers
• Structured output works for all resume sections (basics, work, education, skills, projects, awards)
• No breaking changes to existing functionality

Additional Context

This aligns with modern LLM best practices and leverages Gemini's advanced capabilities while maintaining the system's provider-agnostic design philosophy.