requirements.txt

Django>=5.1,<7.0
djangorestframework==3.17.1
djangorestframework-simplejwt[crypto]==5.5.1
psycopg[binary]==3.3.3
django-cors-headers==4.9.0
drf-spectacular==0.29.0
ruff==0.15.9
pytest-django==4.12.0
django-filter==25.2
python-dotenv==1.2.2
PyYAML>=6.0,<7.0

# Module 5 checkpoint auto-import — fetch + parse course-outline pages
# (schema.org/Course JSON-LD via Coursera/DataCamp, generic HTML fallback,
# YouTube Data API, freeCodeCamp GitHub JSON). No headless browser. `requests`
# is already present transitively (transformers/huggingface); pinned direct
# since the importers import it explicitly. `lxml` is the bs4 parser backend.
requests>=2.31
beautifulsoup4>=4.12,<5.0

# Module 8 (resume parsing) — lexical layer.
pdfplumber>=0.11,<1.0
rapidfuzz>=3.0,<4.0

# Module 8 full NER fallback chain (Nucha BERT -> JobBERT -> SkillNER -> SBERT+pgvector).
# CPU-only torch wheel to avoid the ~2 GB CUDA payload. Verified 2026-04-22 as
# installable end-to-end on Python 3.13.11 on Windows.
--extra-index-url https://download.pytorch.org/whl/cpu
torch>=2.11,<3.0
transformers>=4.44,<5.0
sentence-transformers>=3.0,<4.0
skillNer>=1.0.3
# skillNer transitively imports IPython (notebook display helpers) at extractor
# construction time — without this pin every resume upload logs
# "ner.skillner: raised No module named 'IPython', skipping" and layer 3 goes
# dark. See research/06-dataset-sourcing.md §5 NER fallback chain.
ipython>=8.0,<9.0
spacy>=3.8,<4.0
pgvector>=0.3.6,<1.0

# Deployment (HF Spaces Docker + Neon Postgres). Additive only; unused locally
# (runserver/local Postgres keep working). gunicorn = WSGI server, whitenoise =
# static file serving (Django admin/DRF) without nginx, dj-database-url = parse
# Neon's single DATABASE_URL.
gunicorn>=23.0
whitenoise>=6.8
dj-database-url>=2.3