Improve Context Management in Message Generation

[avishek0769]

Summary

Replace naive “last N messages” context assembly with a more robust context strategy for better quality and lower token waste.

Problem

sendMessage currently appends recent messages directly, which can cause context bloat, stale history inclusion, and weaker grounding in long chats.

Expected Solution

Implement improved context window management (for example: rolling summary + recent turns + retrieval context). Ensure token-aware truncation and predictable prompt structure.

Scope

• backend/controllers/chatMessage.controller.js
• Optional helper modules under backend/utils/
• Prompt and retrieval composition logic

Acceptance Criteria

• Context builder enforces token/length budget before LLM call.
• Long chats remain coherent without unbounded prompt growth.
• Regression tests or reproducible checks verify better handling of long sessions.
• Source grounding behavior remains intact.

Notes

Potential approaches: conversational summarization, semantic memory retrieval, recency weighting, or hybrid strategy.