Implement Context-Based Response Validation Logic

[nuwangeek]

Description

Implement logic to handle queries that can be answered from the current user query and conversation history. This includes greeting detection and context-based response generation.

Objective

• Detect greetings (hello, hi, goodbye, thanks) and provide friendly responses
• Analyze conversation history to answer follow-up questions
• Use LLM to intelligently determine if query references conversation context

Technical Requirements

3.1 Greeting Detection

Supported Languages: Estonian (et), English (en)

Greeting Types:

• hello: Tere, Hello, Hi, Good morning
• goodbye: Nägemist, Bye, Goodbye, See you
• thanks: Tänan, Thank you, Thanks
• casual: Hei, Hey, Moi

Response Generation: Use LLM to generate contextually appropriate responses in the same language

3.2 Context Analysis

Input: User query + conversation history (last 10 turns)

LLM Task:

1. Check if query is a greeting → respond appropriately
2. Check if query references conversation history → extract answer
3. Otherwise → fallback to RAG workflow

Output Format:

{
  "is_greeting": true/false,
  "can_answer_from_context": true/false,
  "answer": "response text or null",
  "reasoning": "brief explanation"
}

Acceptance Criteria

• Implement ContextAnalyzer with LLM-based greeting detection
• Implement conversation history formatting (last 10 turns)
• Implement ContextWorkflowExecutor with streaming support
• Add greeting response constants (ET/EN)
• Add fallback to RAG when context cannot answer
• Add cost tracking for context check LLM calls
• Unit tests for greeting detection
• Unit tests for context-based responses
• Integration tests for context workflow