README.md

metadata

title: MEPLIFEOS FUGTH AI Gateway
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: false

MEPLIFEOS — FUGTH AI Gateway

A lightweight, protected GGUF inference API for FUGTH Management OS.

• Models: Qwen2.5-0.5B-Instruct and Qwen2.5-1.5B-Instruct (Q4_K_M GGUF).
• Both load into memory; only one inference is processed at a time (shared bounded FIFO queue).
• POST /v1/chat — protected generation with caller-provided RAG context.
• GET /health — public readiness, queue depth, and loaded models.
• POST /admin/default-model — change the default model until restart (admin key required).
• / — manual smoke-test page; keys are entered in the browser and are not stored.
• /docs — API documentation.

Configuration in Space Settings → Variables and secrets

Required Secret: FUGTH_API_KEY — a strong random key (32+ characters).

Optional Secret: FUGTH_ADMIN_KEY — a separate strong random key for admin configuration.

Optional variables:

Name	Default	Meaning
FUGTH_DEFAULT_MODEL	1.5b	Default on startup (1.5b or 0.5b)
MAX_QUEUE_SIZE	8	Number of waiting requests (not counting the active one)
REQUEST_TIMEOUT_SECONDS	110	HTTP request wait time before returning 504
N_CTX	2048	Context window for each model
N_THREADS	2	CPU inference threads

Models download on startup and live on the Space's ephemeral disk. Restarts can require a download. Do not upload model weights to the repo.

API example

curl -X POST 'https://fugthchat-meplifeos.hf.space/v1/chat' \
  -H 'Content-Type: application/json' \
  -H 'X-API-Key: YOUR_SECRET_VALUE' \
  -d '{
    "model": "0.5b",
    "business_name": "Example Hair Studio",
    "message": "What time do you close on Saturday?",
    "context": "Hours: Monday-Friday 9AM-6PM. Saturday 10AM-3PM. Sunday closed.",
    "max_tokens": 120
  }'

For the 1.5B model, set "model": "1.5b". Omit model to use the default.

Important: RAG retrieval, authentication of end-users, customer isolation, rate limits, and account authorization belong in the FUGTH backend. This service receives only context selected by the backend; it does not implement vector retrieval, durable conversations, or fine-tuning. Keep FUGTH_API_KEY on your backend, never in public website source code. Do not pass sensitive data unless your architecture and policies permit it.

What if it fails to build?

• Check the Space's Build logs for missing compiler or network errors.
• Check Container logs for downloading/initializing the models.
• Visit /health and wait for "ready": true.
• If Hugging Face blocks restarting older Docker Spaces, changing source code cannot bypass the account/hardware restriction.
• To limit downloads for troubleshooting, set LOAD_MODELS=0.5b as a Space variable, then return to LOAD_MODELS=0.5b,1.5b after validation.

For production-scale usage, add a persistent job queue, tenant rate limits, billing controls, observability, and another always-on inference provider. Free Spaces can sleep.