rawtherapy

RawTherapy

Context

An AI psychotherapy patient simulator for therapist training (Fastify + TypeScript, Postgres, Redis, an LLM adapter). A few standalone samples from the backend, picked independently rather than as one connected walkthrough — some of them reference modules that aren't included here.

What's here

• engine/rule-engine.ts — pure state-transition logic: therapist message → applied rules → next trust/stress/openness/mood.
• llm/adapter.ts — a small LlmAdapter interface with an OpenAI implementation (streaming + non-streaming, AbortController timeout) and a MockLlmAdapter used automatically when no API key is configured.
• services/state-cache.ts — Redis-backed session state cache, a per-session lock (SET NX EX), and an idempotency-key store.
• services/conversation-service.ts — the turn orchestrator: session lock, idempotency check, safety analysis (prompt-injection / high-risk-language detection), rule engine, prompt build, LLM call, persistence.
• frontend/components/ChatWindow.tsx, MessageBubble.tsx, MessageInput.tsx — the chat UI: message list, a composer with push-to-record voice input and a keyboard shortcut for it, and status-aware message bubbles.
• frontend/hooks/usePatientSpeech.ts — plays streamed PCM text-to-speech audio via the Web Audio API as it arrives, with a prebuffer to avoid choppy playback and a browser speechSynthesis fallback if streaming fails.

Anonymization

No real API keys, patient data, or .env values are part of this folder.