# Vendored engine + model live here at build time; never commit them.
#
# `python tools/vendor_llm_engine.py` copies the llama.cpp runtime (a couple
# of MB of shared libraries + the pure-Python ctypes binding) into
# `vendor/llama_cpp/`, and optionally a GGUF into `model/`. Both are build
# artefacts: they make the engine + weights part of the installed program,
# and neither belongs in git (the engine is ~12 MB, the model ~0.1-1 GB).
#
# The vendored package KEEPS its own name (`llama_cpp`) because its internal
# imports are absolute (`from llama_cpp._ctypes_extensions import ...`); the
# loader puts `vendor/` on sys.path so those resolve to the baked copy rather
# than to anything in site-packages.
vendor/
model/
