Use agent for context injection instead of similarity

[suraj-phanindra]

Replace pure semantic similarity threshold with intelligent two-stage approach:

Stage 1: Semantic search using embeddings to get top candidates (fast filter)
Stage 2: LLM-based evaluation to select truly relevant facts (smart filter)

Key Changes:

• Add ContextInjectionAgent using LlamaIndex workflow
• Agent analyzes query context and candidate facts
• Makes nuanced decisions about relevance vs. noise
• Understands contextual connections beyond vector similarity
• More selective: only injects facts that genuinely improve conversation

Updated retrieval_agent.py:

• Integrate ContextInjectionAgent into ContextRetrievalAgent
• Lower semantic threshold (0.3) to give agent more candidates
• Pass top 10-15 candidates to injection agent for evaluation
• Maintain backward compatibility with use_agent=False flag

Benefits:

• Reduces irrelevant context injection
• Better understands nuanced relevance
• Improves conversation quality across platforms
• More intelligent than fixed similarity thresholds