# Fractafrag — Project Decisions

## D001 — Backend Language & Framework
- **Choice:** Python + FastAPI
- **Rationale:** AI/ML integrations (pgvector, LLM clients, embeddings) are Python-native. FastAPI gives async performance with Pydantic auto-generated OpenAPI docs. Celery + Redis is mature for job queues.
- **Made by:** Collaborative
- **Revisable:** No

## D002 — Frontend Stack
- **Choice:** React 18 + Vite + Three.js + TanStack Query + Zustand + Tailwind CSS
- **Rationale:** Three.js for 3D shader rendering, raw WebGL for feed thumbnails. React UI, TanStack Query for server state, Zustand for client state.
- **Made by:** Collaborative
- **Revisable:** No

## D003 — Database & Cache
- **Choice:** PostgreSQL 16 + pgvector + Redis 7
- **Rationale:** pgvector for taste/style/desire embeddings (ANN). Redis for sessions, feed cache, rate limiting, Celery broker.
- **Made by:** Collaborative
- **Revisable:** No

## D004 — Container Orchestration
- **Choice:** Single Docker Compose stack, self-hosted, no cloud dependencies
- **Rationale:** Self-contained with nginx reverse proxy. .env-driven config.
- **Made by:** Collaborative
- **Revisable:** No

## D005 — Media Storage (Q1)
- **Choice:** Docker volume initially, S3-compatible config flag for later migration
- **Rationale:** Volume is simplest for single-server. Add Minio/S3 when storage grows large.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D006 — Style Embedding Model (Q2)
- **Choice:** Heuristic classifier + LLM structured output for M1, fine-tune later
- **Rationale:** No training data yet for fine-tuning. Heuristic is fast/cheap, LLM fills accuracy gaps.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D007 — Renderer Approach (Q3)
- **Choice:** Puppeteer + Headless Chromium
- **Rationale:** Accurate browser-equivalent rendering. Profile at M2 and optimize if needed.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D008 — Generation Status UX (Q4)
- **Choice:** Polling for M5, SSE upgrade later
- **Rationale:** Simpler to implement. Generation takes 5-30s, 2s polling is acceptable UX.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D009 — Comments Scope (Q6)
- **Choice:** Defer to post-M5 polish sprint
- **Rationale:** Schema is in place. Feature is not on critical path for core product loop.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D010 — Moderation Approach (Q7)
- **Choice:** Admin API endpoints only (/api/v1/admin/queue). No admin UI for M4.
- **Rationale:** Simple approve/reject actions via API. Admin panel deferred until scale demands it.
- **Made by:** Agent (per spec recommendation)
- **Revisable:** Yes

## D011 — Creator Economy
- **Choice:** Deferred until organic traction (500 DAU, 1000 shaders, 20 active creators)
- **Rationale:** Build the hooks (schema stubs, engagement tracking), not the features. Monetization on a platform nobody uses is worthless.
- **Made by:** Collaborative (per spec Section 11)
- **Revisable:** Yes