Yes, this is a common challenge when combining a knowledge base with agent decision flows. Passing large KB context directly into prompts can quickly lead to token limits and reduced reliability. A better approach is to separate concerns: Use a retrieval-based system (RAG) for the KB, where only the most relevant chunks are fetched per query instead of injecting the entire dataset. This keeps token usage low and responses focused. For decision making, keep node prompts lightweight and structured. The agent should call the KB only when needed, rather than mixing all logic and data in a single prompt. I usually design the flow by starting with intent detection or classification, then retrieving the most relevant knowledge base context dynamically, injecting only the top matches into the prompt, and using structured prompts to ensure consistent decision making. To improve reliability, I add guardrails such as response validation and fallback logic, apply confidence scoring before replying, and use strict prompt templates for each node. This way you reduce token usage, improve response accuracy, and keep the agent behavior consistent.