Trigger rebuild to resolve scheduling failure

Real Google Auth Upgrade: Integrated Google Identity Services (GIS). Button now triggers real Google Popup if GOOGLE_CLIENT_ID is set in secrets.

UI Fix: Replaced broken Google SVG with FontAwesome icon. Added Google Identity Script for upcoming OAuth integration.

UI Upgrade: Broadened Google Auth button and renamed models to academic research titles (Distilled Knowledge Transfer and Recursive Cognitive Reasoning). Improved Google Login logic.

Final Fix: Resolved YOLO PyTorch 2.6 security blocks, forced browser cache-busting for app.js, and included initialized vqa.db

Fix critical null-byte corruption in main.py

Heartbeat push to force Hugging Face build trigger

Fix login lock: Handled missing Florence references that crashed JS, and manually forced SQLite admin creation.

Major Project Finalization: Renamed Models, Implemented SQLite DB Auth, User Tracking, and adjusted model restraints per request.

Fix Boss API parsing: Implement Aggressive Extraction and Panic Fallback

Deploy Gemma-3-12B Boss API with 3-tier VLM fallback system

Final Fix: Use verified router endpoint and :hf-inference model suffix for Boss API

Fix HF Boss Router URLs with 3-tier fallback logic

Switch Boss API to Mistral Pixtral 12B to fix persistent 404/410 Deprecation loop

Fix StopIteration: Switch to hybrid fallback requests logic for HF Boss API

Migrate HF Boss API to Hugging Face SDK to auto-resolve router endpoints

Fix Hugging Face Serverless API deprecation route

Revamp explanation generation in YOLO model using FLAN-T5, and add Florence-2 Fusion model

Fix libxcb ImportError: Add GUI Linux dependencies for OpenCV into explicit Dockerfile layer

Add packages.txt to fix OpenCV missing libxcb.so.1 GUI dependencies

Fix ImportError and RAM overload: blip_yolo_model seamlessly reuses blip_model instances natively

Expand to 5 models safely, no secrets

Simplify down to only 3 pure models: Local, Gemini, Groq Llama 4 Scout. Removed broken 90B/Fusion models.

Remove Fusion. Provide two purely native Groq endpoints (90b and Scout)

Add native Groq vision alongside Fusion model

Switch Groq strategy to Fusion (BLIP + LLaMA 3.3 70B Text) to maximize rate limits instantly

Update Groq to Llama 4 Scout as Llama 3.2 is decommissioned

Add Groq Llama 3.2 Vision implementation with strict formatting constraints

Remove Florence and HF API models, keep only Gemini and local pipelines

Fix HF Vision API URL to proper router domain

Switch HF Vision pure API to raw requests to expose HTTP errors

Switch HF Vision pure API to Llava 1.5 to avoid timeout payload drops

Switch HF Vision API strictly to Llama Vision

Implement 5th Option: Raw HF Serverless Vision-Instruct API

Migrate text api to official hf hub and bypass gated models

Implement ultimate Florence + Llama 70B text fusion pipeline

Pin transformers to 4.45.2 for Florence-2

Fix PyTorch dependency version conflict

Fix Florence-2 Flash Attention dependency crash

Implement local Florence-2 fallback

Switch HF API Model to Qwen2.5-VL due to Llama server issues

Fix Hugging Face API URL routing

Add Hugging Face Serverless API option (Option 3)

Initialize VQA system deployment

Initial commit of GenAI VQA System