CBT Assistant — Local AI Mental Health Chatbot & CBT Journal with Ollama

A local-first application featuring structured thought records, mood and sleep tracking, self-assessments, and SOS exercises. Includes a conversational AI grounded in a built-in CBT knowledge base via RAG, with persistent session memory.

# macOS 14+
git clone https://github.com/KazKozDev/cbt-assistant.git && cd cbt-assistant && ./start_cbt_assistant.command

# Windows 10 22H2+ (PowerShell or cmd)
git clone https://github.com/KazKozDev/cbt-assistant.git
cd cbt-assistant
start_cbt_assistant.bat

# Linux
git clone https://github.com/KazKozDev/cbt-assistant.git && cd cbt-assistant
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
ollama pull qwen3.5:9b
python backend/server.py

Quick start: run the local AI CBT chatbot

1. Use the platform-specific commands above. The macOS and Windows launchers automatically set up the Python environment, prepare Ollama, download qwen3.5:9b, start the server on port 8000, and open your browser. Subsequent launches will reuse this setup. On Linux, ensure ollama serve is running, execute the commands, and open http://localhost:8000.
2. Begin by chatting or launching an SOS practice. During a conversation, the assistant can proactively suggest and configure a specific SOS practice, scene, or a 1–10 minute timer if you agree.
3. Switch between English and Russian in Settings.

Important

CBT Assistant is a self-help and journaling tool, not a therapist, medical device, crisis service, or substitute for professional care. The entire local knowledge base is built directly on clinical research and evidence-based CBT protocols for depression, depressive episodes, anxiety, and low-mood states, but AI-generated responses and self-assessments can still be imperfect. If you may be in immediate danger or at risk of harming yourself or someone else, contact local emergency services or a qualified crisis service now.

CBT journal app, mood tracker, self-assessments, and SOS exercises

These records provide future conversation context instead of leaving each chat isolated. Messages and Thought Diary entries are stored in SQLite; interface logs and screening state also use browser localStorage.

• Progress Dashboard: Visual charts and analytics tracking your mood dynamics, sleep patterns, and emotional wellbeing trends over time.
• Records Calendar: An interactive month-view calendar unifying daily entries across thoughts, mood, sleep, planned activities, and test scores.
• Clinical Self-Assessments: Standardized psychometric screening tools (PHQ-9 for depression, GAD-7 for anxiety, Rosenberg Self-Esteem, and Burnout assessment) with score interpretations.
• Thought Diary: Log situations, automatic thoughts, emotions, intensities, cognitive distortions, and rational responses.
• Mood & Sleep Logs: Track daily scores, notes, sleep duration, interruptions, and sleep quality.
• SOS Portals: Fullscreen interactive guides for paced breathing, 5-4-3-2-1 grounding, progressive muscle relaxation, and STOP techniques, featuring visual scenes, sound, and countdowns.
• Voice Mode & Speech Input: Hands-free voice conversation with microphone speech recognition and realistic neural text-to-speech voice responses (English and Russian).
• Reports: Generate text summaries or printable PDF reports directly in the browser.

Private AI mental health chatbot with local memory

The default chat model is qwen3.5:9b. The assistant maintains durable long-term memory across sessions by persisting a structured personal profile and rolling conversation summaries in SQLite.

• Empathetic, Open Dialogue: A completely private, safe space to speak your mind, vent about tough days, talk heart-to-heart, and receive warm, non-judgmental support.
• Strict Clinical Grounding (No Hallucinations): Prompt-engineered and gated by FastEmbed RAG to avoid inventing psychological concepts or hallucinating clinical protocols. It strictly bases therapeutic guidance on verified, evidence-based CBT knowledge ([KB:chunk_id]), abstaining from fabricated advice when context is absent.
• AI Tools & Direct Journaling: During chat, you can ask the assistant to directly record structured entries into your Thought Diary (analyzing automatic thoughts, cognitive distortions, and rational responses) or log your Sleep Diary (bedtime, wake time, duration, and sleep quality).
• Proactive Interventions: The AI can launch agreed fullscreen SOS portals, schedule planned activities, or pull recent sleep, mood, and screening assessments for contextual guidance.
• Cross-Session Memory: A persistent SESSION_ID ensures your conversation context, personal profile, and rolling summary survive browser reloads and app restarts.

Manage your derived profile and summary via:

GET    /api/memory/{session_id}
DELETE /api/memory/{session_id}

Local CBT knowledge base with RAG and FastEmbed

At startup, src/rag/knowledge_base.py splits the bundled Markdown CBT library by heading hierarchy and computes embeddings locally using FastEmbed (paraphrase-multilingual-mpnet-base-v2). The model cannot bypass this retrieval path.

GET /api/knowledge/search?q=sleep&top_k=3
GET /api/knowledge/status

Search results include stable chunk and document IDs, source paths, section hierarchy, similarity scores, index version, and a local trace ID. The AI is instructed to cite clinical claims as [KB:chunk_id]. The default similarity threshold is 0.46. When no passage clears it, the assistant provides general support but does not invent specific CBT protocols.

How the local AI assistant works

Browser UI
   ↓
FastAPI REST + WebSocket API
   ↓
Prompt assembly ← CBT knowledge search
   ↓                     ↓
Ollama chat          FastEmbed (CPU / ONNX)
   ↓                     ↓
Response             Markdown knowledge base
   └──────── SQLite + browser localStorage

The backend persists messages and structured records in SQLite. RAG restores a matching cached NumPy index or rebuilds it if knowledge content changes. Failed first builds remain unavailable; failed rebuilds fall back to the previous complete in-memory index in degraded mode.

Limitations

• Network: While local-first, features like text-to-speech, browser speech recognition, Google Fonts, CDN scripts, and YouTube media require internet access.
• Security: The server binds to 0.0.0.0:8000 with permissive CORS and lacks authentication or encryption. Do not expose it to the public internet or untrusted networks.
• Performance: Chat quality, response time, and memory usage depend on your hardware and chosen Ollama model. Chat remains unavailable until the embedding index is built.
• Platform: Docker and an automatic Linux launcher are not included; Linux requires manual setup.

  

Issues · ARCHITECTURE · Privacy & Safety · LICENSE · LinkedIn