saef-api / README.md
m4az19's picture
Deploy SAHAB-1 backend and pretrained AI
124816b verified
|
Raw History Blame Contribute Delete
2.04 kB
metadata
title: SAHAB-1 AI Backend
emoji: 🌧️
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
base_path: /api/health
pinned: false
models:
  - giswqs/s2-water-unetplusplus-efficientnet-b4

SAHAB-1 · satforge 2026

Deployment target: m4az19/saef-api. Direct API origin: https://m4az19-saef-api.hf.space. This deployment replaces the previous SAEF BiomedCLIP application in this Space. Existing old files may remain in repository history; the new Dockerfile uses the SAHAB-1 backend entrypoint. No legacy SAEF model is used. See MODELS.md.

Backend for the Vercel dashboard. Runs satellite water inference, NDVI, seasonal rainfall analysis, device telemetry and Qwen3.5:0.8b explanations. Weights download during the Docker build; no training is required.

Before starting, add Secrets SATFORGE_DASHBOARD_KEY and SATFORGE_DEVICE_KEY in Space Settings. Generate two different random values. The API deliberately refuses to start without these secrets in a Space. Only the API port 7860 is exposed; Ollama listens on loopback.

Use Public or Protected visibility. The Vercel server gateway authenticates with X-Dashboard-Key; the ground receiver uses X-Device-Key. The /api/health endpoint is public for startup monitoring.

Default writable data is ephemeral at /home/sahab/data. Restarting/rebuilding can erase observation history and cached images; weights are baked into the container. This is a demonstration configuration, not durable telemetry hosting. Do not place the SQLite database directly on an object-storage/FUSE bucket without establishing SQLite locking support. Durable operation needs a suitable filesystem-backed volume or a separate database and asset-storage integration.

CPU Basic is a starting point, not a measured performance guarantee. Idle Spaces can sleep; use hardware with an appropriate uptime policy for continuous monitoring. Satellite acquisition times remain visible; model labels do not establish drought impacts. Missing device measurements remain unavailable.