playlist-admin 2.6.83 → 2.6.85

package/cli/index.js

@@ -41,8 +41,9 @@ function projectRoot() {
 
 function exePath(root) {
   const candidates = [
+    path.join(root, 'build', 'windows', 'x64', 'runner', 'Release', 'playlist-admin.exe'),
+    path.join(root, 'build', 'windows', 'x64', 'runner', 'Debug', 'playlist-admin.exe'),
     path.join(root, 'build', 'windows', 'x64', 'runner', 'Release', 'playlist_administrator.exe'),
-    path.join(root, 'build', 'windows', 'x64', 'runner', 'Debug', 'playlist_administrator.exe'),
   ];
   return candidates.find((c) => fs.existsSync(c)) || null;
 }
@@ -54,10 +55,16 @@ function python() {
 function forward(exe, args) {
   // Deliver CLI args via PA_CLI_ARGS (JSON) — Flutter Dart exposes no
   // command-line args in release on this SDK (see lib/main.dart).
+  // PA_ROOT lets the Python bridge find rag/ scripts when invoked from the CLI.
+  const root = projectRoot();
   const child = spawn(exe, [], {
     stdio: 'inherit',
     windowsHide: false,
-    env: { ...process.env, PA_CLI_ARGS: JSON.stringify(args) },
+    env: {
+      ...process.env,
+      PA_CLI_ARGS: JSON.stringify(args),
+      ...(root ? { PA_ROOT: root } : {}),
+    },
   });
   child.on('exit', (code) => process.exit(code ?? 1));
   child.on('error', (e) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "playlist-admin",
-  "version": "2.6.83",
+  "version": "2.6.85",
   "description": "Playlist Administrator CLI - thin launcher for the Flutter app engine (pipeline, podcast, spotube) + podcast RAG query",
   "type": "module",
   "bin": {

package/rag/build_db.py

@@ -17,6 +17,9 @@ import requests
 
 from _resolve import podcast_downloads_dir, chroma_db_dir
 
+if hasattr(sys.stdout, "reconfigure"):
+    sys.stdout.reconfigure(encoding="utf-8")
+
 OLLAMA_URL = "http://localhost:11434"
 
 

package/rag/query.py

@@ -3,6 +3,7 @@
   python rag/query.py "問題" --topk 8          # 文字輸出
   python rag/query.py "問題" --json            # JSON 輸出 (含 full_text)
   python rag/query.py "問題" --no-full         # 不回傳全文, 只回傳命中片段
+  python rag/query.py "問題" --answer          # 用 Ollama qwen2.5:7b 生成回答
 """
 import argparse
 import json
@@ -19,7 +20,11 @@ import requests
 from build_db import read_text_file
 from _resolve import chroma_db_dir
 
-OLLAMA_URL = "http://localhost:11434"
+if hasattr(sys.stdout, "reconfigure"):
+    sys.stdout.reconfigure(encoding="utf-8")
+
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+CHAT_MODEL = os.environ.get("OLLAMA_CHAT_MODEL", "qwen2.5:7b")
 
 
 def embed_one(model: str, text: str) -> list[float]:
@@ -43,6 +48,57 @@ def locate_chunk(full_norm: str, chunk: str, window: int = 100) -> int:
     return idx
 
 
+def pick_chat_models() -> list[str]:
+    """回傳可嘗試的 chat 模型順序：env 指定 → 本機非 embedding 模型 → cloud 模型"""
+    forced = os.environ.get("OLLAMA_CHAT_MODEL", "").strip()
+    try:
+        tags = requests.get(f"{OLLAMA_URL}/api/tags", timeout=30).json().get("models", [])
+    except Exception:
+        return [forced] if forced else []
+    local = [m["name"] for m in tags if m.get("size", 0) > 0]
+    cloud = [m["name"] for m in tags if m.get("size", 0) == 0]
+    banned = {"bge-m3", "bge-m3:latest"}
+    local = [m for m in local if m not in banned]
+    if forced:
+        return [forced, *local, *cloud]
+    return [*local, *cloud]
+
+
+def generate_answer(question: str, hits: list[dict]) -> str:
+    """用 Ollama 生成基於檢索片段的回答；全部模型都失敗就回傳錯誤訊息"""
+    context = "\n".join(
+        f"[{h['show']} | {h['date']} | {h['file']}]\n{h['chunk']}"
+        for h in hits
+    )
+    prompt = (
+        "你是 Podcast 內容助理。請只根據以下 Podcast 逐字稿片段回答使用者的問題，"
+        "回答用繁體中文，簡潔、分點陳述，並在結尾列出引用來源（節目、日期、檔名）。\n\n"
+        f"片段資料：\n{context}\n\n問題：{question}"
+    )
+    errors = []
+    for model in pick_chat_models():
+        try:
+            resp = requests.post(
+                f"{OLLAMA_URL}/api/chat",
+                json={
+                    "model": model,
+                    "messages": [{"role": "user", "content": prompt}],
+                    "stream": False,
+                },
+                timeout=600,
+            )
+            if resp.status_code in (400, 401, 404):
+                errors.append(f"{model}: HTTP {resp.status_code}")
+                continue
+            resp.raise_for_status()
+            return resp.json()["message"]["content"].strip()
+        except requests.RequestException as e:
+            errors.append(f"{model}: {e}")
+    if errors:
+        raise RuntimeError(" | ".join(errors))
+    return ""
+
+
 def main() -> None:
     ap = argparse.ArgumentParser()
     ap.add_argument("question", nargs="?", help="要查的問題")
@@ -53,6 +109,8 @@ def main() -> None:
     ap.add_argument("--no-full", action="store_true", help="不回傳全文")
     ap.add_argument("--out", default="", help="JSON 輸出檔路徑 (與 --json 併用)")
     ap.add_argument("--show", default="", help="只搜指定節目")
+    ap.add_argument("--answer", action="store_true", help="用 Ollama 生成回答 (qwen2.5:7b)")
+    ap.add_argument("--topk-answer", type=int, default=6, help="生成回答使用的片段數")
     args = ap.parse_args()
 
     if not args.question:
@@ -120,6 +178,11 @@ def main() -> None:
 
     if args.json:
         payload = {"question": args.question, "results": results}
+        if args.answer:
+            try:
+                payload["answer"] = generate_answer(args.question, raw_hits[: max(args.topk_answer, 3)])
+            except Exception as e:
+                payload["answer_error"] = f"(Ollama 生成失敗: {e})"
         out = json.dumps(payload, ensure_ascii=False, indent=2)
         if args.out:
             Path(args.out).write_text(out, encoding="utf-8")