playlist-admin 2.7.15 → 2.15.17

package/rag/_resolve.py

@@ -1,37 +1,66 @@
-"""Resolve podcast RAG paths from the app config (same as the CLI/GUI)."""
-import json
-import os
-from pathlib import Path
-
-
-def find_base() -> Path:
-    env_bp = os.environ.get("BASE_PATH")
-    if env_bp:
-        return Path(env_bp)
-    local = os.environ.get("LOCALAPPDATA")
-    pointer = None
-    if local:
-        pointer = Path(local) / "playlist-admin" / "data" / "config.json"
-        if not pointer.exists():
-            pointer = Path(local) / "Playlist Administrator" / "data" / "config.json"
-    try:
-        if pointer and pointer.exists():
-            data = json.loads(pointer.read_text(encoding="utf-8"))
-            bp = data.get("base_path") or data.get("basePath")
-            if bp:
-                return Path(bp)
-    except Exception:
-        pass
-    return Path(r"C:\Users\CPXru\Music\Spotube")
-
-
-def podcast_downloads_dir() -> Path:
-    return find_base() / "podcast_downloads"
-
-
-def rag_data_dir() -> Path:
-    return find_base() / "podcast_downloads" / "podcast_rag"
-
-
-def chroma_db_dir() -> Path:
+"""Resolve podcast RAG paths from the app config (same as the CLI/GUI)."""
+import json
+import os
+import platform
+from pathlib import Path
+
+
+def _is_wsl() -> bool:
+    """偵測是否在 WSL 中執行。"""
+    if "microsoft" in platform.release().lower():
+        return True
+    if os.path.exists("/proc/sys/fs/binfmt_misc/WSLInterop"):
+        return True
+    return False
+
+
+def _win_to_wsl(path: Path) -> Path:
+    """將 Windows 路徑轉為 WSL /mnt/c/... 路徑。"""
+    s = str(path)
+    if len(s) >= 2 and s[1] == ":":
+        drive = s[0].lower()
+        rest = s[2:].replace("\\", "/")
+        return Path(f"/mnt/{drive}{rest}")
+    return path
+
+
+def find_base() -> Path:
+    env_bp = os.environ.get("BASE_PATH")
+    if env_bp:
+        bp = Path(env_bp)
+        if _is_wsl():
+            bp = _win_to_wsl(bp)
+        return bp
+
+    # WSL: 直接用 /mnt/c/ 路徑
+    if _is_wsl():
+        return Path("/mnt/c/Users/CPXru/Music/playlist-admin")
+
+    # Windows: 讀 config.json
+    local = os.environ.get("LOCALAPPDATA")
+    pointer = None
+    if local:
+        pointer = Path(local) / "playlist-admin" / "data" / "config.json"
+        if not pointer.exists():
+            pointer = Path(local) / "Playlist Administrator" / "data" / "config.json"
+    try:
+        if pointer and pointer.exists():
+            data = json.loads(pointer.read_text(encoding="utf-8"))
+            bp = data.get("base_path") or data.get("basePath")
+            if bp:
+                return Path(bp)
+    except Exception:
+        pass
+    return Path(r"C:\Users\CPXru\Music\playlist-admin")
+
+
+def podcast_downloads_dir() -> Path:
+    return find_base() / "podcasts"
+
+
+def rag_data_dir() -> Path:
+    return find_base() / "cache" / "podcast"
+
+
+def chroma_db_dir() -> Path:
     return rag_data_dir() / "chroma_db"

package/rag/build_db.py

@@ -1,216 +1,302 @@
-"""把 podcast 逐字稿 (.txt) 建立成 ChromaDB 向量資料庫
-用法: python rag/build_db.py [--data DIR] [--db DIR] [--model MODEL] [--reset] [--workers N]
-"""
-import argparse
-import os
-import re
-import sys
-import hashlib
-import time
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from pathlib import Path
-
-sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
-
-import chromadb
-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"
-
-
-def embed_one_batch(model: str, texts: list[str]) -> tuple[list[str], list[list[float]]]:
-    """單批 embedding；400 時逐個重試，回傳 (成功文字, 對應向量)"""
-    resp = requests.post(
-        f"{OLLAMA_URL}/api/embed",
-        json={"model": model, "input": texts},
-        timeout=600,
-    )
-    if resp.status_code == 400:
-        kept: list[str] = []
-        vectors: list[list[float]] = []
-        for t in texts:
-            r2 = requests.post(
-                f"{OLLAMA_URL}/api/embed",
-                json={"model": model, "input": [t]},
-                timeout=600,
-            )
-            if r2.status_code == 400:
-                print(f"!! 跳過問題 chunk (len={len(t)}): {t[:100]!r}")
-                continue
-            r2.raise_for_status()
-            kept.append(t)
-            vectors.append(r2.json()["embeddings"][0])
-        return kept, vectors
-    resp.raise_for_status()
-    return texts, resp.json()["embeddings"]
-
-
-def embed_texts(model: str, texts: list[str], batch: int = 32, workers: int = 6) -> tuple[list[str], list[list[float]]]:
-    """併發呼叫 Ollama /api/embed；回傳 (成功文字, 對應向量)"""
-    batches = [texts[i:i + batch] for i in range(0, len(texts), batch)]
-    if len(batches) <= 1:
-        if not batches:
-            return [], []
-        return embed_one_batch(model, batches[0])
-
-    kept_all: list[str] = []
-    vectors: list[list[float]] = []
-    results: list[tuple[list[str], list[list[float]]]] = [None] * len(batches)
-    with ThreadPoolExecutor(max_workers=workers) as ex:
-        fut = {ex.submit(embed_one_batch, model, b): i for i, b in enumerate(batches)}
-        for f in as_completed(fut):
-            results[fut[f]] = f.result()
-    for k, v in results:
-        if k is None:
-            continue
-        kept_all.extend(k)
-        vectors.extend(v)
-    return kept_all, vectors
-
-
-def read_text_file(path: Path) -> str:
-    """讀檔，自動嘗試 UTF-8 / Big5 / GBK"""
-    raw = path.read_bytes()
-    for enc in ("utf-8-sig", "utf-8", "big5", "gb18030"):
-        try:
-            return raw.decode(enc)
-        except UnicodeDecodeError:
-            continue
-    return raw.decode("utf-8", errors="replace")
-
-
-def split_sentences(text: str) -> list[str]:
-    """依中文句號斷句，過濾空白行"""
-    text = re.sub(r"\s+", "", text)
-    parts = re.split(r"(?<=[。！？!?；;])", text)
-    return [p for p in parts if len(p) >= 4]
-
-
-def make_chunks(sentences: list[str], max_len: int = 600, overlap: int = 60) -> list[str]:
-    """句子接成 chunk，每塊約 max_len 字元，前後重疊 overlap；超長句強制硬切"""
-    chunks: list[str] = []
-    cur = ""
-    for s in sentences:
-        while len(s) > max_len:
-            if cur:
-                chunks.append(cur)
-                cur = cur[-overlap:] if len(cur) > overlap else cur
-            chunks.append(s[:max_len])
-            s = s[max_len:]
-        if len(cur) + len(s) > max_len and cur:
-            chunks.append(cur)
-            tail = cur[-overlap:] if len(cur) > overlap else cur
-            cur = tail + s
-        else:
-            cur += s
-    if cur:
-        chunks.append(cur)
-    return chunks
-
-
-def detect_meta(path: Path) -> dict:
-    """從檔名/路徑猜節目與日期"""
-    parts = path.parts
-    show = parts[-2] if len(parts) >= 2 else "unknown"
-    name = path.stem
-    m = re.search(r"(20\d{2})[_\-_](\d{1,2})[_\-_](\d{1,2})", name)
-    date = f"{m.group(1)}-{m.group(2).zfill(2)}-{m.group(3).zfill(2)}" if m else ""
-    return {"show": show, "file": name, "date": date}
-
-
-def main() -> None:
-    ap = argparse.ArgumentParser()
-    ap.add_argument("--data", default=str(podcast_downloads_dir()))
-    ap.add_argument("--db", default=str(chroma_db_dir()))
-    ap.add_argument("--model", default="bge-m3")
-    ap.add_argument("--reset", action="store_true", help="重建資料庫")
-    ap.add_argument("--limit", type=int, default=0, help="只索引前 N 篇 (測試用)")
-    ap.add_argument("--workers", type=int, default=6, help="併發 embedding 數")
-    args = ap.parse_args()
-
-    client = chromadb.PersistentClient(path=args.db)
-    if args.reset:
-        try:
-            client.delete_collection("podcasts")
-        except Exception:
-            pass
-    col = client.get_or_create_collection(
-        "podcasts",
-        metadata={"hnsw:space": "cosine"},
-    )
-
-    # 已索引的檔名 (resume 用) - 分段讀取避免 SQL 變數超限
-    done_files = set()
-    offset = 0
-    while True:
-        batch = col.get(limit=5000, offset=offset, include=["metadatas"])["metadatas"]
-        if not batch:
-            break
-        for m in batch:
-            if m:
-                done_files.add(m.get("file", ""))
-        offset += 5000
-        if offset > 1000000:
-            break
-
-    files = [f for f in sorted(Path(args.data).rglob("*.txt")) if f.stem not in done_files]
-    if args.limit:
-        files = files[: args.limit]
-    print(f"待索引 {len(files)} 篇 (已跳過 {len(done_files)} 篇)")
-
-    total_chunks = 0
-    t_start = time.time()
-    for i, f in enumerate(files, 1):
-        text = read_text_file(f)
-        chunks = make_chunks(split_sentences(text))
-        # token 保險: 單 chunk 超過 2800 字元就再切
-        final_chunks: list[str] = []
-        for ch in chunks:
-            final_chunks.extend(
-                [ch[i:i + 2800] for i in range(0, len(ch), 2800)] if len(ch) > 2800 else [ch]
-            )
-        chunks = final_chunks
-        if not chunks:
-            continue
-        meta = detect_meta(f)
-        try:
-            kept_chunks, vectors = embed_texts(args.model, chunks, workers=args.workers)
-        except requests.exceptions.HTTPError as e:
-            print(f"!! HTTP 錯誤 檔案={meta['file']} 前300字={chunks[0][:300] if chunks else ''}")
-            raise
-        except requests.exceptions.ConnectionError:
-            print("無法連線 Ollama，請先啟動 (ollama serve) 並 pull bge-m3")
-            sys.exit(1)
-        if not kept_chunks:
-            continue
-        ids = [
-            hashlib.md5(f"{meta['file']}|{j}".encode()).hexdigest()
-            for j in range(len(kept_chunks))
-        ]
-        metadatas = [
-            {**meta, "chunk": j, "chars": len(c), "path": str(f)}
-            for j, c in enumerate(kept_chunks)
-        ]
-        col.add(ids=ids, documents=kept_chunks, metadatas=metadatas, embeddings=vectors)
-        total_chunks += len(kept_chunks)
-        if i % 5 == 0 or i == len(files):
-            elapsed = time.time() - t_start
-            speed = i / elapsed
-            eta = (len(files) - i) / speed if speed > 0 else 0
-            print(
-                f"[{i}/{len(files)}] {i/len(files)*100:.1f}% | "
-                f"{speed:.2f}篇/秒 | 剩餘ETA {eta/60:.1f}分 | "
-                f"chunks累計 {total_chunks} | {meta['show'][:10]}/{meta['file'][:25]}",
-                flush=True,
-            )
-
-    print(f"完成! 共 {len(files)} 篇, {total_chunks} 個 chunk")
-
-
-if __name__ == "__main__":
+"""把 podcast 逐字稿 (.txt) 建立成 ChromaDB 向量資料庫
+用法: python rag/build_db.py [--data DIR] [--db DIR] [--model MODEL] [--reset] [--workers N]
+"""
+import argparse
+import os
+import re
+import sys
+import hashlib
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+
+import chromadb
+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"
+
+
+def embed_one_batch(model: str, texts: list[str]) -> tuple[list[str], list[list[float]]]:
+    """單批 embedding；400 時逐個重試，回傳 (成功文字, 對應向量)"""
+    resp = requests.post(
+        f"{OLLAMA_URL}/api/embed",
+        json={"model": model, "input": texts},
+        timeout=600,
+    )
+    if resp.status_code == 400:
+        kept: list[str] = []
+        vectors: list[list[float]] = []
+        for t in texts:
+            r2 = requests.post(
+                f"{OLLAMA_URL}/api/embed",
+                json={"model": model, "input": [t]},
+                timeout=600,
+            )
+            if r2.status_code == 400:
+                print(f"!! 跳過問題 chunk (len={len(t)}): {t[:100]!r}")
+                continue
+            r2.raise_for_status()
+            kept.append(t)
+            vectors.append(r2.json()["embeddings"][0])
+        return kept, vectors
+    resp.raise_for_status()
+    return texts, resp.json()["embeddings"]
+
+
+def embed_texts(model: str, texts: list[str], batch: int = 32, workers: int = 6) -> tuple[list[str], list[list[float]]]:
+    """併發呼叫 Ollama /api/embed；回傳 (成功文字, 對應向量)"""
+    batches = [texts[i:i + batch] for i in range(0, len(texts), batch)]
+    if len(batches) <= 1:
+        if not batches:
+            return [], []
+        return embed_one_batch(model, batches[0])
+
+    kept_all: list[str] = []
+    vectors: list[list[float]] = []
+    results: list[tuple[list[str], list[list[float]]]] = [None] * len(batches)
+    with ThreadPoolExecutor(max_workers=workers) as ex:
+        fut = {ex.submit(embed_one_batch, model, b): i for i, b in enumerate(batches)}
+        for f in as_completed(fut):
+            results[fut[f]] = f.result()
+    for k, v in results:
+        if k is None:
+            continue
+        kept_all.extend(k)
+        vectors.extend(v)
+    return kept_all, vectors
+
+
+def read_text_file(path: Path) -> str:
+    """讀檔，自動嘗試 UTF-8 / Big5 / GBK"""
+    raw = path.read_bytes()
+    for enc in ("utf-8-sig", "utf-8", "big5", "gb18030"):
+        try:
+            return raw.decode(enc)
+        except UnicodeDecodeError:
+            continue
+    return raw.decode("utf-8", errors="replace")
+
+
+def split_sentences(text: str) -> list[str]:
+    """依中文句號斷句，過濾空白行"""
+    text = re.sub(r"\s+", "", text)
+    parts = re.split(r"(?<=[。！？!?；;])", text)
+    return [p for p in parts if len(p) >= 4]
+
+
+def make_chunks(sentences: list[str], max_len: int = 600, overlap: int = 60) -> list[str]:
+    """句子接成 chunk，每塊約 max_len 字元，前後重疊 overlap；超長句強制硬切"""
+    chunks: list[str] = []
+    cur = ""
+    for s in sentences:
+        while len(s) > max_len:
+            if cur:
+                chunks.append(cur)
+                cur = cur[-overlap:] if len(cur) > overlap else cur
+            chunks.append(s[:max_len])
+            s = s[max_len:]
+        if len(cur) + len(s) > max_len and cur:
+            chunks.append(cur)
+            tail = cur[-overlap:] if len(cur) > overlap else cur
+            cur = tail + s
+        else:
+            cur += s
+    if cur:
+        chunks.append(cur)
+    return chunks
+
+
+def detect_meta(path: Path) -> dict:
+    """從檔名/路徑猜節目、日期與來源類型"""
+    parts = path.parts
+    show = parts[-2] if len(parts) >= 2 else "unknown"
+    name = path.stem
+    m = re.search(r"(20\d{2})[_\-_](\d{1,2})[_\-_](\d{1,2})", name)
+    date = f"{m.group(1)}-{m.group(2).zfill(2)}-{m.group(3).zfill(2)}" if m else ""
+    source = "course" if re.search(r"夜工智|機械群|課程", str(path)) else "podcast"
+    return {"show": show, "file": name, "date": date, "source": source}
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", default=str(podcast_downloads_dir()))
+    ap.add_argument("--db", default=str(chroma_db_dir()))
+    ap.add_argument("--model", default="bge-m3")
+    ap.add_argument("--reset", action="store_true", help="重建資料庫")
+    ap.add_argument("--limit", type=int, default=0, help="只索引前 N 篇 (測試用)")
+    ap.add_argument("--workers", type=int, default=6, help="併發 embedding 數")
+    ap.add_argument("--batch", type=int, default=64, help="每批 embedding 數量")
+    args = ap.parse_args()
+
+    client = chromadb.PersistentClient(path=args.db)
+    if args.reset:
+        try:
+            client.delete_collection("podcasts")
+        except Exception:
+            pass
+    col = client.get_or_create_collection(
+        "podcasts",
+        metadata={"hnsw:space": "cosine"},
+    )
+
+    # ── Dedup: remove duplicate sig entries, keep first only ──────
+    total = col.count()
+    if total > 0:
+        sig_to_ids: dict[str, list[str]] = {}
+        offset = 0
+        while offset < total:
+            batch = col.get(limit=5000, offset=offset, include=["metadatas"])
+            ids = batch["ids"]
+            metas = batch["metadatas"]
+            for rid, m in zip(ids, metas):
+                sig = (m or {}).get("sig", "")
+                if sig:
+                    sig_to_ids.setdefault(sig, []).append(rid)
+            offset += 5000
+        to_delete = [ids[1:] for ids in sig_to_ids.values() if len(ids) > 1]
+        flat_delete = [item for sublist in to_delete for item in sublist]
+        if flat_delete:
+            print(f"Dedup: removing {len(flat_delete)} duplicate records...")
+            for i in range(0, len(flat_delete), 4500):
+                col.delete(ids=flat_delete[i:i+4500])
+            print(f"Dedup done. New count: {col.count()}")
+
+    # 已索引內容 hash 去重 - 分段讀取避免 SQL 變數超限
+    done_sigs: set[str] = set()
+    offset = 0
+    while True:
+        batch = col.get(limit=5000, offset=offset, include=["metadatas"])["metadatas"]
+        if not batch:
+            break
+        for m in batch:
+            if m and m.get("sig"):
+                done_sigs.add(m["sig"])
+        offset += 5000
+        if offset > 1000000:
+            break
+
+    # 一次性遷移: 舊版 chunks 沒有 sig — 從 metadata 的 path 讀檔補上，
+    # 避免 1662 篇全部被當新的重嵌入。
+    def file_sig(path: Path) -> str:
+        raw = path.read_bytes()
+        # Strip BOM and normalize whitespace to avoid false re-embeds
+        if raw.startswith(b'\xef\xbb\xbf'):
+            raw = raw[3:]
+        return hashlib.md5(raw).hexdigest()
+
+    migrated = 0
+    offset = 0
+    while True:
+        batch = col.get(limit=5000, offset=offset,
+                        include=["metadatas"])["metadatas"]
+        if not batch:
+            break
+        for m in batch:
+            if not m or m.get("sig"):
+                continue
+            p = m.get("path")
+            if not p or not os.path.exists(p):
+                continue
+            try:
+                sig = file_sig(Path(p))
+                done_sigs.add(sig)
+                migrated += 1
+            except Exception:
+                pass
+        offset += 5000
+        if offset > 1000000:
+            break
+    if migrated:
+        print(f"migrated {migrated} legacy chunks to content-sig dedup")
+
+    # 無內容/失敗檔案的跳過清單:0 chunks 的檔案不會寫進 DB,
+    # 若沒有持久狀態會每次 pipeline 都被當「待索引」重試。
+    skip_state_file = Path(args.db) / "_skipped_files.json"
+    skipped: set[str] = set()
+    if skip_state_file.exists():
+        try:
+            import json
+            skipped = set(json.loads(skip_state_file.read_text(encoding="utf-8")))
+        except Exception:
+            skipped = set()
+
+    # 檔案內容 hash — 內容更新(來源切換)時 sig 變 → 重嵌入
+    def has_content(path: Path) -> bool:
+        return bool(make_chunks(split_sentences(read_text_file(path))))
+
+    files = [
+        f for f in sorted(Path(args.data).rglob("*.txt"))
+        if file_sig(f) not in done_sigs
+        and not (f.stem in skipped and not has_content(f))
+    ]
+    if args.limit:
+        files = files[: args.limit]
+    print(f"待索引 {len(files)} 篇 (已嵌入 {len(done_sigs)} 篇, 無內容跳過 {len(skipped)} 篇)")
+
+    total_chunks = 0
+    t_start = time.time()
+    for i, f in enumerate(files, 1):
+        text = read_text_file(f)
+        chunks = make_chunks(split_sentences(text))
+        # token 保險: 單 chunk 超過 2800 字元就再切
+        final_chunks: list[str] = []
+        for ch in chunks:
+            final_chunks.extend(
+                [ch[i:i + 2800] for i in range(0, len(ch), 2800)] if len(ch) > 2800 else [ch]
+            )
+        chunks = final_chunks
+        if not chunks:
+            # 空/太短內容: 標記為已跳過, 避免每次重試
+            skipped.add(f.stem)
+            import json
+            skip_state_file.parent.mkdir(parents=True, exist_ok=True)
+            skip_state_file.write_text(
+                json.dumps(sorted(skipped), ensure_ascii=False, indent=1),
+                encoding="utf-8",
+            )
+            continue
+        meta = detect_meta(f)
+        sig = file_sig(f)
+        try:
+            kept_chunks, vectors = embed_texts(args.model, chunks, batch=args.batch, workers=args.workers)
+        except requests.exceptions.HTTPError as e:
+            print(f"!! HTTP 錯誤 檔案={meta['file']} 前300字={chunks[0][:300] if chunks else ''}")
+            raise
+        except requests.exceptions.ConnectionError:
+            print("無法連線 Ollama，請先啟動 (ollama serve) 並 pull bge-m3")
+            sys.exit(1)
+        if not kept_chunks:
+            continue
+        ids = [
+            hashlib.md5(f"{meta['file']}|{j}".encode()).hexdigest()
+            for j in range(len(kept_chunks))
+        ]
+        metadatas = [
+            {**meta, "sig": sig, "chunk": j, "chars": len(c), "path": str(f)}
+            for j, c in enumerate(kept_chunks)
+        ]
+        col.add(ids=ids, documents=kept_chunks, metadatas=metadatas, embeddings=vectors)
+        total_chunks += len(kept_chunks)
+        if i % 5 == 0 or i == len(files):
+            elapsed = time.time() - t_start
+            speed = i / elapsed
+            eta = (len(files) - i) / speed if speed > 0 else 0
+            print(
+                f"[{i}/{len(files)}] {i/len(files)*100:.1f}% | "
+                f"{speed:.2f}篇/秒 | 剩餘ETA {eta/60:.1f}分 | "
+                f"chunks累計 {total_chunks} | {meta['show'][:10]}/{meta['file'][:25]}",
+                flush=True,
+            )
+
+    print(f"完成! 共 {len(files)} 篇, {total_chunks} 個 chunk")
+
+
+if __name__ == "__main__":
     main()