Spaces:
Running
Running
Commit History
23 days ago
-
Update parent company name to Avorquen Running
e87d5ad
28 days ago
-
fix: convert scikit-learn vocab int64 to int for json serialization in knowledge base index
cad8d5f
-
v1.1: Security audit fixes + bug fixes
- Add gateway_auth middleware requiring HF_SHARED_SECRET or valid Supabase JWT - Add per-IP rate limiting (30 req/min) with periodic memory cleanup - Fallback JWT verification using SUPABASE_JWT_SECRET - Add upload file size/type validation (max 10MB, whitelisted extensions) - Accept JSON payload on /api/knowledge/learn matching client-side extraction - Disable /docs and /redoc in production - Remove .env from git tracking and add .gitignore - Bake default non-sensitive EDUAI env vars into Dockerfilec7511f3
-
feat: add SSE streaming support to /api/chat
- New run_inference_stream() generator yields tokens one by one - Strips <think> tags on-the-fly during streaming - stream=True returns StreamingResponse, stream=False unchanged - Final SSE event includes full answer + parsed output1357c7f
-
fix: use pre-built CPU wheel to avoid build timeout
c7ca406 -
fix: pin llama-cpp-python to 0.3.35 for Gemma 4 support
af7801b -
fix: upgrade llama-cpp-python for Gemma 4 support
Old version (0.3.9) didn't recognize Gemma 4 GGUF architecture. Unpinned to install latest version with Gemma 4 support.1aebc5d
-
feat: swap LLM from Llama 3.2 1B to Gemma 4 E2B (Q4_K_M)
- Better quality model (2B effective params vs 1B) - Still fits free tier HF Space (~1.5GB vs ~0.8GB) - Strip <think> tags from output (Gemma 4 has built-in reasoning) - System prompt instructs model to skip thinking blocks - Faster, smarter responses for guest usersba743e6
-
feat: fix file processing — OCR, PDF, DOCX now work on free tier
- Added ocr_model.py (Tesseract + RapidOCR, no PyTorch) - Rewrote file_analyzer.py to work standalone without multimodal_manager - Replaced sentence-transformers with scikit-learn TF-IDF in knowledge_base.py - Removed heavy deps: sentence-transformers, soundfile (saves ~2GB) - Added lightweight deps: pdfplumber, pypdf, python-docx, Pillow, rapidocr-onnxruntime - Added tesseract-ocr to Dockerfile - Upload & learn endpoints now always use file_analyzerc402937
-
chore: trigger rebuild (previous build timed out)
4adc38e
-
fix: rebrand EduAI to Cognizya, add cognizya.vercel.app CORS, fix profanity false-positives
- System prompt: EduAI -> Cognizya with full identity rules - Identity: responds as Cognizya, mentions Velmora only when asked - CORS: added cognizya.vercel.app to allowed origins - Academic topics: whitelisted psychology, history, etc. - App title: EduAI Cloud -> Cognizya Cloudb184cb9
-
Fix: use Supabase auth.get_user() for ES256 JWT verification
99b9880 -
Add Supabase auth + per-user sessions/flashcards
38ce9bf
-
Fix: accept HEAD requests on /api/health for UptimeRobot
0f9b3f8
-
Update developer bio with accurate info
80de0ca
-
Fix: lazy models/__init__.py - remove ocr/speech/vision imports
7df158f -
Fix: lazy imports, remove heavy deps (chromadb, kokoro) for free tier
8f536ec -
Add eduai-by-srk.vercel.app to CORS whitelist
465e574 -
Fix: runtime model download, writable /tmp paths, reduced context for free tier
17af22e -
Deploy EduAI backend - LLM + TTS + RAG + 15 maths chapters
1d9dc1f -
initial commit
Verified signature86a106f