feat installer and download model on first launch

packaging/build.spec

@@ -0,0 +1,92 @@
+# PyInstaller spec for Dev Assist Chatbot.
+#
+# Build (from the repo root, in an environment with the CUDA 13.x toolkit
+# installed - the toolkit is only needed at BUILD time, to copy its runtime
+# DLLs into the bundle; end users only need an NVIDIA driver):
+#
+#   uv run pyinstaller packaging/build.spec --noconfirm
+#
+# Output lands in dist/DevAssistChatbot/. That whole folder is the app -
+# onedir rather than onefile, since unpacking a multi-GB bundle to a temp
+# dir on every launch would make startup unusable.
+import os
+from pathlib import Path
+
+from PyInstaller.utils.hooks import collect_submodules
+
+PROJECT_ROOT = Path(SPECPATH).parent  # noqa: F821 - injected by PyInstaller
+VENV_SITE_PACKAGES = PROJECT_ROOT / ".venv" / "Lib" / "site-packages"
+LLAMA_CPP_LIB = VENV_SITE_PACKAGES / "llama_cpp" / "lib"
+
+if not LLAMA_CPP_LIB.is_dir():
+    raise SystemExit(
+        f"Expected llama_cpp's compiled libs at {LLAMA_CPP_LIB} - run this via "
+        "'uv run pyinstaller ...' from the repo root so the project venv is used."
+    )
+
+CUDA_PATH = os.environ.get("CUDA_PATH")
+if not CUDA_PATH:
+    raise SystemExit(
+        "CUDA_PATH is not set. Building requires the CUDA toolkit that "
+        "llama-cpp-python was compiled against, so its runtime DLLs can be "
+        "copied into the bundle (end users won't need the toolkit installed)."
+    )
+CUDA_BIN = Path(CUDA_PATH) / "bin" / "x64"
+
+# llama.cpp's ggml backend (ggml-cuda.dll etc.) is loaded dynamically at
+# runtime, not linked at import time, so PyInstaller's static analysis can't
+# discover it - every DLL in llama_cpp/lib/ has to be added explicitly. They
+# all go into the same output folder so Windows' DLL search (which checks a
+# loaded DLL's own directory for its dependencies) can resolve cublas from
+# next to ggml-cuda.dll.
+CUDA_RUNTIME_DLLS = [
+    "cudart64_13.dll",
+    "cublas64_13.dll",
+    "cublasLt64_13.dll",
+    "nvrtc64_130_0.dll",
+    "nvrtc-builtins64_134.dll",
+]
+
+binaries = [(str(dll), "llama_cpp/lib") for dll in LLAMA_CPP_LIB.glob("*.dll")]
+for name in CUDA_RUNTIME_DLLS:
+    dll_path = CUDA_BIN / name
+    if not dll_path.is_file():
+        raise SystemExit(f"Expected CUDA runtime DLL not found: {dll_path}")
+    binaries.append((str(dll_path), "llama_cpp/lib"))
+
+# ddgs picks its search backend (duckduckgo, bing, ...) via importlib at
+# runtime rather than a static import, so PyInstaller won't find those
+# engine modules on its own either.
+hidden_imports = collect_submodules("ddgs") + ["pynvml"]
+
+a = Analysis(  # noqa: F821
+    [str(PROJECT_ROOT / "main.py")],
+    pathex=[str(PROJECT_ROOT)],
+    binaries=binaries,
+    datas=[],
+    hiddenimports=hidden_imports,
+    hookspath=[],
+    excludes=[],
+    noarchive=False,
+)
+pyz = PYZ(a.pure)  # noqa: F821
+
+exe = EXE(  # noqa: F821
+    pyz,
+    a.scripts,
+    [],
+    exclude_binaries=True,
+    name="DevAssistChatbot",
+    console=False,
+)
+
+coll = COLLECT(  # noqa: F821
+    exe,
+    a.binaries,
+    a.zipfiles,
+    a.datas,
+    # UPX would try (slowly, and sometimes badly) to compress the ~650MB of
+    # CUDA runtime DLLs; skip it.
+    upx=False,
+    name="DevAssistChatbot",
+)