I’ve worked on similar LLM-based chatbot systems that rely heavily on SQL for structured data retrieval, and I’ve experienced the same issue you're facing—long response times that make the chatbot feel sluggish. CrewAI is a powerful framework for managing multi-agent workflows, but in use cases where low latency and real-time interaction are critical (like a sports chatbot), it often introduces too much orchestration overhead.

In my experience, switching from CrewAI to a combination of LangChain and LangGraph significantly improves performance. LangChain offers out-of-the-box support for querying SQL databases via its SQL agents, while LangGraph allows you to build more efficient conversational flows using a graph-based execution model. This design allows for conditional logic, state management, and parallelism—all of which help reduce response times drastically.

LangGraph agents are especially effective in chat scenarios because they avoid unnecessary layers of reasoning and can be designed to respond quickly with pre-defined flows. Compared to CrewAI, they’re more lightweight and purpose-built for conversational applications. If you structure your agent to only invoke the language model when needed—and optimize your SQL queries and database indices—you’ll see much better latency overall.

Additionally, I’d recommend caching common queries, especially for frequently requested statistics like recent match results, player scores, or leaderboards. LangChain also integrates easily with retrieval or memory components if you later decide to reintroduce a RAG-like pattern, but in a more controlled way.

To summarize, I suggest moving away from CrewAI and toward LangChain + LangGraph for your use case. This combination is more performant, more flexible for conversational flows, and generally better suited for building responsive chatbots over structured data.