# requirements.txt — HuggingFace Spaces (CPU, 12 GB RAM, 16 GB disk)
# OPTIMIZED: pre-built wheels only — zero C++ source compilation.
# PyTorch CPU wheel is installed separately in Dockerfile BEFORE this file.
# ─────────────────────────────────────────────────────────────────────────────
#
# Build-stall root causes fixed:
#   1. llama-cpp-python: pinned to ==0.3.30 (latest wheel on abetlen's CPU index).
#      Using >=0.3.0 resolves to 0.3.31 on PyPI which has NO wheel — triggers
#      a full CMake/C++ build that stalls for 8+ hours on HF Spaces CPU hardware.
#   2. unstructured[pdf,image]: removed — pulled numba→llvmlite (LLVM compilation).
#      OCR fallback in document_loader.py is wrapped in try/except ImportError and
#      is non-critical; PyMuPDF handles the vast majority of PDFs natively.
#
# CRITICAL: --extra-index-url MUST be declared first (before the package line)
# so pip resolves llama-cpp-python from the wheel index, not PyPI source.
# ─────────────────────────────────────────────────────────────────────────────

--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

# LLM Runtime — pinned to latest pre-built CPU wheel (py3-none-manylinux2014_x86_64)
# DO NOT use >=; that resolves to the PyPI source tarball and stalls the build.
llama-cpp-python==0.3.30

# ── Web Framework ─────────────────────────────────────────────────────────────
flask>=3.0.0
flask-cors>=4.0.0
flask-limiter>=3.5.0

# ── LangChain — text splitting only (no community extras) ─────────────────────
# langchain-community removed: not imported anywhere; adds aiohttp + SQLAlchemy.
# langchain base removed: not directly imported; only sub-packages are used.
langchain-core>=0.2.0,<2.0.0
langchain-text-splitters>=0.3.0,<2.0.0

# ── CrewAI (agent orchestration — ingest crew + LLM wrapper) ──────────────────
crewai>=0.36.0,<2.0.0

# ── Embeddings — sentence-transformers (bge-small ~130 MB, pure Python) ───────
# Also provides CrossEncoder used for reranking in crew.py.
sentence-transformers>=3.0.0

# ── Transformers (transitive dep of sentence-transformers, pure Python) ────────
transformers>=4.40.0

# ── HuggingFace Hub (model download in gen_llm.py + HfApi in app.py) ─────────
huggingface-hub>=0.23.0

# ── Vector Database (embedded, pre-built manylinux wheel) ────────────────────
chromadb>=0.5.0

# ── Graph Database (embedded, pre-built cp312 wheel) ─────────────────────────
kuzu>=0.11.3

# ── Hybrid Search (pure Python) ───────────────────────────────────────────────
rank-bm25>=0.2.2

# ── Document Loaders (all pre-built wheels or pure Python) ────────────────────
PyMuPDF>=1.24.0
python-docx>=1.1.0
openpyxl>=3.1.0
pandas>=2.0.0
pytesseract>=0.3.10
Pillow>=10.0.0

# ── Utilities ─────────────────────────────────────────────────────────────────
requests>=2.31.0
python-dotenv>=1.0.0
tqdm>=4.66.0
numpy>=1.24.0
pydantic>=2.0.0

# ── Encryption (session key security via cryptography.fernet) ─────────────────
cryptography>=41.0.0

# ── System monitoring (CPU/RAM banner in UI via psutil) ───────────────────────
psutil>=5.9.0

# ── NLP entity extraction (pre-built cp312 wheel + py3-none-any model) ────────
spacy>=3.7.0
en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl