# Optional "smart tier" dependencies. tilt runs WITHOUT any of these — browse, keyword
# search, the live view, event trees, and native resume all work on a stdlib-only clone.
# Install these to unlock semantic search, topic clustering, and mood arcs.
#
#   python -m venv py/.venv && py/.venv/Scripts/pip install -r requirements.txt   (Windows)
#   python -m venv py/.venv && py/.venv/bin/pip install -r requirements.txt        (macOS/Linux)
#
# torch is the heavy one; install the build that matches your platform/GPU from
# https://pytorch.org if the default wheel doesn't fit (e.g. CUDA vs CPU).

numpy>=1.24
torch>=2.1
transformers>=4.51          # 4.51+ required for the Qwen3 embedding model
sentence-transformers>=2.6
scikit-learn>=1.3           # concept clustering

# Titles / summaries / concept names / mood verdicts are generated by a local LLM via
# Ollama (https://ollama.com) — not a pip package. Pull a model summarize.py accepts:
#   ollama pull qwen2.5:3b-instruct
# Without Ollama, chats fall back to first-message titles and everything else still works.
