Project C: Therapy-Style Conversational Agent

Project C for UW x AI Tinkerers W26.

This project provides a test environment for studying prompt robustness, safety boundaries, and failure modes of a therapy-style conversational LLM. The agent is designed to provide neutral, reflective conversation while strictly preserving clinical and safety boundaries.

Team Members

Fouzan Abdullah

Luna Nguyen

Matthew Li

Alia Cai

Ruben Ispiryan

Purpose

The agent is designed to fail safely rather than stretch capability. It prioritizes alignment and refusal correctness over user satisfaction. It explicitly avoids providing medical or psychiatric advice, diagnosing conditions, or engaging in crisis intervention.

Features

• Agent Framework: Built using LangChain, featuring a memory system to maintain conversational state.
• Evaluation Pipeline: Evaluates agent responses using LLM-as-a-judge (Stage 1 and 2), checking compliance with safety outlines and classifying failure modes based on a strict taxonomy.
• Dashboard: Generates an interactive HTML dashboard to visualize pass rates, failure categories, latency, and model regressions over time.
• Prompting Harness: Provides tools to test single prompts or batch run categorized prompts (benign, ambiguous, adversarial).

File Structure

• src/
  • main.py: The entry point for the application. Handles running prompts, evaluating, analyzing, and generating dashboards.
  • agent.py: LangChain-based agent implementation with memory tools.
  • evaluator.py: Implements FailureEvaluator for two-stage safety compliance evaluation.
  • dashboard.py: Generates the HTML visualization dashboard.
  • analyze.py: Processes evaluation results to compute metrics.
  • model_pool.py: Manages the pool of LLM models used for agency and evaluation.
• data/
  • system_prompts/: Contains versioned system prompts (system_prompt_vX.txt).
  • test_prompts_v1.json: Categorized test prompts.
  • evaluation.json: Output of the evaluation step.
  • metrics.json: Output of the analysis step.
  • responses_combined.json: Agent interaction responses.
• agent_spec.md: Detailed specification of the agent's safe behavior and failure boundaries.
• failure_taxonomy.md: Taxonomy of failure modes for the evaluator to classify.
• requirements.txt: Python dependencies.

Setup & Installation

This project utilizes uv for package management and script execution, though standard pip can also be used.

1. Clone the repository and navigate to the project directory:

   cd ProjectC

2. Install dependencies:

   pip install -r requirements.txt

   Note: Using uv pip install -r requirements.txt is recommended for faster installation.

3. Set up environment variables: Create a .env file in the root directory and add your API keys (e.g., OPENAI_API_KEY, GROQ_API_KEY, or others depending on model_pool.py config).

Usage

The primary interface is src/main.py. You can run various commands to interact with the agent or run evaluations.

1. Chat/Test a Prompt

Send a single interactive prompt to the agent:

uv run src/main.py prompt "I'm feeling really stressed today."

2. Run a Category of Prompts

Run the agent against a specific category (benign, ambiguous, adversarial) from the test prompts file:

uv run src/main.py category benign data/test_prompts_v1.json

(Use all as the category to run everything).

3. Evaluate Responses

Run the evaluator on the generated responses to check for safety compliance:

uv run src/main.py evaluate --responses data/responses_combined.json --prompts data/test_prompts_v1.json --output data/evaluation.json

4. Analyze Results

Compute metrics from the evaluations:

uv run src/main.py analyze --evaluations data/evaluation.json --output data/metrics.json

5. Generate Dashboard

Create a visual HTML dashboard of the metrics:

uv run src/main.py dashboard --metrics data/metrics.json --output data/dashboard.html

6. Live Web Chat UI

Run a beautifully rendered interactive web chat using Chainlit:

uv run chainlit run src/chainlit_app.py -w