[Enhancement] Intent Classification to Route Queries Correctly

[Smilin01]

Problem

All queries go through the full search + RAG pipeline, even ones that don't need web search
(e.g., "explain recursion", "what is 2+2", "hello").

Expected Behavior

Classify intent before routing:

Intent	Action
Factual / Current Events	Full search + RAG
Conceptual / Definitional	LLM-only (skip search)
Math / Code	Structured generation
Greeting / Chitchat	Lightweight response
Ambiguous	Search with low confidence flag

Why It Matters

• Reduces unnecessary SearXNG calls (critical since we self-host)
• Faster response for non-search queries
• Lower Groq token usage

Suggested Approach

• Use llama-3.1-8b-instant as a fast intent classifier (single prompt, JSON output)
• Gate the search pipeline on intent result

Labels

enhancement performance good first issue