[REFACTOR] 12 refactor llm backend restructuring

[tomsmoker]

Added

• Integrated Instructor library for enhanced LLM response handling
• New prompt templates for improved query decomposition and keyword extraction
• Implemented LLMService abstract base class for decoupling LLM operations
• Created OpenAIService as a concrete implementation of LLMService
• Added LLMFactory for creating LLM service instances

Changed

• Refactored dependency management in vector and query services
• Removed FastAPI Depends usage from utility functions in vector.py
• Implemented direct calls to get_milvus_client(), get_settings(), and get_embeddings()
• Updated hybrid_search, vector_search, and other related functions
• Modified LLM service to use Instructor for structured outputs
• Updated query processing to leverage new prompt templates
• Refactored vector operations to use the new LLMService abstraction
• Updated document processing pipeline to work with the decoupled LLM service
• Modified dependency injection to include LLM service creation
• Adjusted query processing to utilize the new LLM service structure

Improved

• Enhanced reliability and maintainability of vector and query services
• Optimized query processing in query.py to align with new dependency approach
• Improved structured output handling in LLM responses
• Enhanced flexibility by allowing easy switching between different LLM providers
• Improved testability of LLM-dependent components through abstraction

Fixed

• Resolved issues related to incorrect usage of FastAPI's Depends in non-route functions
• Ensured more predictable behavior in data access and query processing operations
• Resolved issues related to direct OpenAI client usage in vector operations
• Addressed errors in document upload process due to LLM service changes