feat installer and download model on first launch

chatbot.py

@@ -1,15 +1,22 @@
 import json
 import os
 import re
+import ssl
 import time
+import urllib.request
 import uuid
 from collections.abc import Callable
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from pathlib import Path
 
+import certifi
+
 # CUDA 13.x moved its runtime DLLs to bin/x64, but llama_cpp's loader only
 # checks CUDA_PATH/bin, so point it at the right folder before importing.
+# Only relevant when running from source against a system CUDA toolkit - a
+# packaged build ships the runtime DLLs next to llama_cpp's own, so this is a
+# no-op there (CUDA_PATH won't be set on a machine with just an NVIDIA driver).
 _cuda_path = os.environ.get("CUDA_PATH")
 if _cuda_path:
     _cuda_bin_x64 = Path(_cuda_path) / "bin" / "x64"
@@ -19,7 +26,16 @@ if _cuda_path:
 from ddgs import DDGS  # noqa: E402
 from llama_cpp import Llama  # noqa: E402
 
-MODEL_PATH = Path(__file__).parent / "Qwen3.5-27B-IQ4_XS.gguf"
+# Model lives outside the install/source directory - a packaged build's own
+# folder may be read-only (Program Files) or discarded on update, and the
+# model is far too large to embed in an installer, so it's fetched into the
+# user's profile on first run instead.
+MODEL_FILENAME = "Qwen3.5-27B-IQ4_XS.gguf"
+MODELS_DIR = Path.home() / ".dev_assist_chatbot" / "models"
+MODEL_PATH = MODELS_DIR / MODEL_FILENAME
+MODEL_DOWNLOAD_URL = (
+    "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/main/Qwen3.5-27B-IQ4_XS.gguf"
+)
 
 SYSTEM_PROMPT = (
     "You are a knowledgeable developer assistant. Answer questions about "
@@ -192,6 +208,40 @@ def build_messages(history: list[tuple[str, str]], question: str) -> list[dict[s
     return messages
 
 
+def download_model(on_progress: Callable[[int, int], None] | None = None) -> None:
+    """Fetch MODEL_PATH from MODEL_DOWNLOAD_URL if it isn't already present.
+    Downloads to a .part file alongside the target and renames on success, so
+    a crash or interrupted download can't leave a truncated file that
+    load_model() would then try (and fail) to use. If given, on_progress is
+    invoked with (bytes_downloaded, total_bytes) as each chunk arrives -
+    total_bytes is 0 if the server doesn't report a Content-Length."""
+    if MODEL_PATH.exists():
+        return
+    MODELS_DIR.mkdir(parents=True, exist_ok=True)
+    part_path = MODEL_PATH.with_suffix(MODEL_PATH.suffix + ".part")
+
+    request = urllib.request.Request(
+        MODEL_DOWNLOAD_URL, headers={"User-Agent": "dev-assist-chatbot"}
+    )
+    # A fresh/sandboxed Windows install may not have fetched the intermediate
+    # CA cert for huggingface.co's chain yet (normally done transparently via
+    # Windows Update on first need, which a locked-down or offline sandbox
+    # can't reach) - use certifi's bundled CA list instead of the OS trust
+    # store so this doesn't depend on that happening.
+    ssl_context = ssl.create_default_context(cafile=certifi.where())
+    with urllib.request.urlopen(request, context=ssl_context) as response:
+        total = int(response.headers.get("Content-Length", 0))
+        downloaded = 0
+        with open(part_path, "wb") as f:
+            while chunk := response.read(1024 * 1024):
+                f.write(chunk)
+                downloaded += len(chunk)
+                if on_progress is not None:
+                    on_progress(downloaded, total)
+
+    part_path.rename(MODEL_PATH)
+
+
 def load_model() -> Llama:
     return Llama(
         model_path=str(MODEL_PATH),