ponshu-room-lite/docs/architecture/archive/future_plan.md

61 lines
2.8 KiB
Markdown

# Project Roadmap & Future Tasks
## 🧭 Current Decision Point
We are at a crossroads. The current app is stable, and plans for expansion are ready.
### 🔘 Option A: Polish Current App (Low Priority)
Focus on minor unimplemented features.
* [ ] **Micro-interactions (Priority C)**:
* Tab switching animations (fade/slide)
* Dialog entrance animations
* Badge unlock celebrations
* [ ] **Coach Mark Fixes**: Verify/Fix if tutorial overlay persists incorrectly.
* [ ] **Image Compression**: Refactor to use `image` package instead of simple file copy.
### 🔘 Option B: Synology Infrastructure (High Stability)
Establish the data bunker and security.
* [ ] **Phase 2A: Container Manager Setup**:
* Setup `posimai-db` (Postgres) container.
* Setup `ai-proxy` (FastAPI) container.
* Setup `cloudflared` tunnel for secure remote access.
* [ ] **Phase 2B: Automation**:
* Implement nightly batch processing (e.g., AI Recommendations).
### 🔘 Option C: Incense App Expansion (New Feature)
Build the "Posimai Core" platform.
* [ ] **Core Refactoring**: Extract Gemini, Camera, Hive logic to `lib/core`.
* [ ] **Flavor Setup**: Configure build flavors for Sake vs Incense.
* [ ] **Incense App MVP**: Implement `ScentStats` and Zen Mode.
---
## 💡 Architecture FAQ
### Q1. Can Synology handle AI Analysis locally?
**Short Answer: Not recommended for Image/Vision tasks.**
* **Reason**: Standard Synology NAS devices (DS220+, DS923+, etc.) lack powerful GPUs (Graphics Processing Units).
* **Performance**: Running a "Vision LLM" (like Llama 3.2 Vision) on a CPU-only NAS would take **30-120 seconds per image**, compared to **1-3 seconds** with Gemini API.
* **Exception**: Unless you have a specific AI-focused device (e.g., Synology DVA series or a NAS with a PCIe GPU added), it is not practical for user experience.
### Q2. How to avoid high Gemini Token usage?
**Strategy 1: Use the Free Tier (Recommended)**
* **Gemini 1.5 Flash** offers a generous free tier:
* 15 requests per minute (RPM).
* 1,500 requests per day (RPD).
* This is sufficient for personal use and small-scale testing.
**Strategy 2: Caching (Architecture)**
* **Implementation**: Store the AI Analysis result in the local DB (Postgres on Synology).
* **Logic**: Before sending an image to Gemini, check if this exact image hash has been analyzed before. (Only works for exactly identical files).
* **Note**: For new photos, you cannot avoid the first analysis.
**Strategy 3: Local Proxy Limits**
* The current `ai-proxy` already implements a "Rate Limit" (10/day). This prevents runaway token usage/cost.
---
## 🗺️ Long-term Vision
* **Posimai Core**: A single codebase powering multiple collection apps.
* **Hybrid Cloud**: Google for "Brain" (AI), Synology for "Memory" (DB/Backup).