Introducing Counselor Copilot, Bridging the Gap in Crisis Counseling

At the LlamaIndex RAG-a-thon, the team's Counselor Copilot took 2nd place in the Traditional track and 1st place in the Datastax/AstraDB category. The target is the TrevorText counseling flow of The Trevor Project, which provides crisis support for LGBTQ+ youth.

Crisis counselors face high demand during peak hours while juggling multiple chats and checking complex PDF-format contact and case documents. This increases the risk of burnout and leaves less time to focus on conversations with youth.

The solution, Counselor Copilot, is a real-time assistant that runs in the background without requiring counselors to enter separate prompts. It reads contact context and chat history together to suggest appropriate responses, which counselors can edit before use, and it does not respond directly to contacts.

The core features are as follows.

• Extracts contact data in real time from complex PDFs and provides a summary at the start of a conversation
• Assesses the need for urgent intervention based on chat context and suggests escalation to a supervisor
• Generates draft responses using the organization's guidelines and resources
• Searches for location-based local resources and shares them via email
• Fills out case forms in the CRM and summarizes the conversation for the record

The implementation starts by using LlamaParse to extract and summarize relevant information from PDFs within the CRM. After that, a LlamaIndex ReAct Agent monitors the conversation, calling on tools such as escalation, response/resource recommendation, and web search and delivery as appropriate to the situation.

For response recommendations, they built a vector database containing The Trevor Project documents, used RAG to retrieve situation-appropriate guidelines, and used GPT-4 to draft responses for counselors. Finally, they used the conversation content to fill in Salesforce fields and automated the counseling summary as well.

Future expansion directions were also presented.

• Fine-tuning an open-source LLM on counseling conversation records to improve cost and quality
• Adding a conversation stage detection tool that reflects the recommended style for each stage of the conversation
• Collecting counselors' thumbs up/down feedback for use in subsequent training

This approach points toward reducing administrative work so counselors can focus on their core role, enabling more crisis response with limited staff.