💬 NLP

Tell Me: An LLM-powered Mental Well-being Assistant with RAG, Synthetic Dialogue Generation, and Agentic Planning

This paper introduces "Tell Me," a comprehensive LLM-powered mental well-being system that integrates a RAG-based conversational assistant, a synthetic dialogue generator for research and data augmentation, and an agentic planning crew to deliver personalized self-care plans and guided meditation, aiming to bridge gaps in mental health support through responsible, context-aware AI innovation.

Original authors: Trishala Jayesh Ahalpara

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Trishala Jayesh Ahalpara

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine Tell Me as a digital "wellness workshop" built by researchers at Fujitsu. It's not a doctor, and it's not a replacement for a real therapist. Instead, think of it as a high-tech, reflective journal that can talk back to you, help you practice conversations, and build you a custom self-care schedule.

The paper presents this system as a collection of three distinct tools working together, like three different specialists in a single room:

1. The "Knowledgeable Listener" (The RAG Assistant)

Imagine you are talking to a friend who has read every book on mental health but needs a quick reminder of the specific advice that fits your situation.

How it works: Instead of just guessing what to say based on its general training, this tool has a "library" of real (but anonymized) therapy conversations. When you share a problem, it quickly searches this library for similar situations and uses those examples to craft a response.
The Analogy: It's like having a conversation where your friend can instantly pull up a relevant story from a notebook to say, "Oh, someone else felt this way, and here is how they handled it," rather than giving generic advice like "Just breathe."
What they found: In their tests, people felt this "library-connected" version was clearer and more helpful than a standard AI that doesn't use the library. However, it was sometimes a tiny bit slower because it had to do the "searching" first.

2. The "Safe Rehearsal Room" (Synthetic Dialogue Generator)

Researchers often need data to study therapy, but real therapy sessions are private and cannot be shared.

How it works: This tool acts like a stage director. You give it a "character profile" (e.g., "a 24-year-old student stressed about school"), and it plays both roles: the anxious student and the supportive therapist. It generates a fake conversation based on that profile.
The Analogy: Think of it as a flight simulator for therapists. Just as pilots practice in a simulator without risking a real plane, researchers can use these fake conversations to study how therapy works without ever needing to see a real, private patient.
What they found: The tool is good at mimicking the structure of therapy (listening and reflecting), but the paper notes it sometimes tries to "fix" problems too quickly, like a mechanic jumping to repair a car before fully understanding the noise it's making.

3. The "Personal Project Manager" (The Agentic Planner)

While the first tool helps you talk about your feelings, this tool helps you do something about them.

How it works: This is a team of AI agents working together. One agent listens to your conversation and picks out the main worries. A second agent researches healthy habits. A third agent writes a 7-day plan and even records a guided meditation audio file for you to listen to.
The Analogy: Imagine you tell a personal trainer, "I'm tired and stressed." Instead of just saying "Good luck," this system acts like a coach who writes you a specific weekly workout schedule, books your yoga class, and records a voice note to guide you through a breathing exercise.
What they found: The system successfully turned a messy, emotional story into a clean, organized 7-day schedule with morning, afternoon, and evening tasks, plus a downloadable audio track.

The Big Picture

The paper emphasizes that Tell Me is a prototype for research, not a medical device.

Safety First: The system has a "guardrail" that acts like a bouncer. If you say something indicating you might hurt yourself, it stops the conversation and directs you to real-world help immediately.
The Goal: The authors see this as a way to lower the barrier to getting support. Just as we exercise to stay physically fit, they hope tools like this can become a normal part of daily mental maintenance—something you use to reflect and plan, complementing (but never replacing) professional care.

In short: It's a three-in-one tool that helps you chat with a knowledgeable AI, lets researchers practice with fake therapy sessions, and builds you a custom self-care plan with audio guides.

Technical Summary: Tell Me – An LLM-Powered Mental Well-being Assistant

Problem Statement

Despite the rapid adoption of Large Language Models (LLMs) in mental health applications, current systems face three persistent limitations:

Lack of Context: Responses often remain generic or shallow due to insufficient contextual grounding.
Data Scarcity and Confidentiality: Access to authentic therapeutic dialogues is restricted, hindering dataset availability, reproducibility, and the training of specialized models.
Static Functionality: Many existing well-being tools lack the ability to adapt to individual needs or evolving emotional states, offering little beyond static advice.

Methodology

The authors introduce Tell Me, an open-source prototype designed as a lightweight, extensible testbed for responsible LLM applications in mental well-being. The system architecture integrates three distinct modules, unified via a Streamlit front-end that supports both cloud-based (e.g., GPT-4o, Claude 3.7 Sonnet) and local (Ollama) inference.

1. RAG-Based Assistant (Context-Aware Dialogue)

This core module provides reflective, knowledge-grounded dialogue.

Knowledge Base: Utilizes open-source datasets (Counsel Chat and Mental Health Counseling Conversations) reformatted into a Context–Response schema.
Retrieval: Employs the BAAI/bge-small-en-v1.5 embedding model for vectorization, selected for its performance on the Massive Text Embedding Benchmark (MTEB) and inference efficiency. Documents are chunked (128 tokens) and retrieved using a VectorIndexRetriever (top-3 results via cosine similarity).
Safety: An LLM-driven guardrail (Sentiment_chain via LangChain) pre-processes inputs to detect high-risk signals (e.g., self-harm), triggering a deterministic crisis-routing mechanism rather than free-form generation.
Memory: Maintains conversational coherence using a ChatMemoryBuffer (3,000-token limit).

2. Synthetic Dialogue Generator

To address data scarcity, this module generates synthetic client–therapist transcripts conditioned on user-defined client profiles (demographics, concerns, history).

Architecture: Implemented via LangChain using two isolated LLMChain objects with Conversation Buffer Memory.
Process: A "Client Chain" role-plays the user persona, while a "Therapist Chain" utilizes reflective listening techniques. The system iteratively passes chat history between the two models to produce multi-turn transcripts.

3. Agentic Well-being Planner (CrewAI)

This module translates dialogue into actionable, long-term support using a multi-agent orchestration framework.

Agents:
- Transcript Analyzer: Extracts emotions, concerns, and triggers.
- Plan Generator: Creates a structured seven-day self-care routine.
- Meditation Generator: Scripts guided audio sessions.
Workflow: Agents operate sequentially. The Analyzer's output serves as the strict context for the Plan Generator. The final meditation script is asynchronously converted to audio via Text-to-Speech (TTS) integration (e.g., Edge Neural TTS, ElevenLabs) and delivered as an MP3.

Key Contributions

Context-Sensitive RAG Assistant: A system benchmarked across nine LLMs and ten well-being scenarios, validated through an LLM-as-a-judge framework and a blind human study comparing RAG vs. non-RAG outputs.
Synthetic Dialogue Generator: A profile-conditioned tool for safe data augmentation that generates realistic therapeutic exchanges without compromising patient confidentiality.
Agentic Planner: A CrewAI-based workflow that dynamically adapts conversations into actionable self-care plans and guided meditation, bridging the gap between static tools and dynamic support.

Evaluation Results

Retrieval Performance

In a retrieval-only evaluation using 30 synthetic query-document pairs:

Hit Rate@5: 1.0000
MRR@5: 0.9222
Note: The authors caution that the perfect Hit Rate is likely a byproduct of the constrained search space ( $N=30$ ) and synthetic queries, serving primarily as a sanity check.

LLM-as-a-Judge Benchmark

Nine models were evaluated on 10 prompts across five dimensions (Safety, Empathy, Usefulness, Clarity, Overall Quality).

Top Performer: Claude 3.7 Sonnet achieved the highest score (9.5/10) when judged by GPT-5.
Observations: High-performing models (Claude, Gemma-3, LLaMA-3) demonstrated empathy validation, invitational dialogue, and cultural sensitivity. Lower-performing models exhibited issues such as pathologizing user experiences or offering generic, minimization-prone advice.

Human Evaluation

A within-subject study with 10 participants compared the RAG-enabled assistant against a non-RAG baseline.

RAG Advantages: Participants rated the RAG system significantly higher in Clarity (4.2 vs. 3.5) and Helpfulness (4.0 vs. 3.9). Qualitative feedback described the RAG assistant as "organized," "on-point," and capable of validating user inputs.
Trade-offs: The non-RAG system was noted as faster but less engaging. The RAG system occasionally suffered from slower response times.
Overall: The RAG assistant achieved a higher mean overall score (3.8 vs. 3.6), suggesting that retrieval grounding improves value for reflective support.

Significance and Claims

The paper positions Tell Me not as a substitute for professional therapy, but as a reflective space for emotional processing and a research testbed for responsible innovation.

Research Utility: The system addresses the shortage of confidential therapeutic data by providing a mechanism for generating synthetic, profile-driven dialogues, enabling safer experimentation and data augmentation.
Bridging Gaps: It demonstrates how conversational assistants can lower barriers to support by complementing existing care, offering dynamic adaptation (via the Agentic Planner) where static tools fail.
Interdisciplinary Potential: The authors highlight the system as a platform for collaboration between NLP researchers and mental health professionals to advance human–AI interaction.
Modesty: The authors explicitly acknowledge limitations, including the small sample size of the human study, the lack of clinical validation by licensed professionals, and the fact that synthetic dialogues cannot fully capture the nuance of real therapeutic exchanges. The system is presented as a prototype for daily reflection and emotional maintenance, analogous to how physical activity complements medical care.