feat: 完成全量功能开发，包括前端管理后台与后端服务优化

此提交实现了完整的知识库管理系统：
1. 新增Vue3 + Antd Vue前端管理后台，包含登录、文档管理、检索、类目设置等完整页面
2. 重构后端LLM调用抽象层，支持Ollama与OpenAI兼容服务动态切换
3. 调整默认嵌入模型配置为本地bge-m3模式
4. 优化入库任务去重逻辑与缓存清理机制
5. 完善Docker镜像构建与docker-compose部署配置
6. 修复多项测试用例与兼容性问题
7. 新增运行时配置API，支持动态调整系统参数

app/core/classifier.py

@@ -1,6 +1,7 @@
 """文档分类器
 
-入库链路第二步：基于 L1 总结，用 Ollama 小模型将文档判定为 taxonomy 中的
+入库链路第二步：基于 L1 总结，用 LLM（Ollama 或 OpenAI 兼容服务，由
+runtime_settings.models.classify 决定）将文档判定为 taxonomy 中的
 主类目 + 附加标签：
 - LLM 输出解析失败 / 类目名不在 taxonomy → 归 uncategorized（confidence=0.0）
 - 置信度低于阈值 → 主类目归 uncategorized，候选类目名保留进 tags（软召回用）
@@ -13,7 +14,7 @@ import structlog
 
 from app.config import settings
 from app.models.knowledge import UNCATEGORIZED, CategoryResult, TaxonomyCategory, load_taxonomy
-from app.services.ollama import OllamaClient
+from app.services.llm import LLMClient, create_llm_client
 
 logger = structlog.get_logger()
 
@@ -47,8 +48,10 @@ def _extract_json(raw: str) -> dict | None:
 class Classifier:
     """文档分类器：将 L1 总结判定为 taxonomy 主类目 + 附加标签"""
 
-    def __init__(self, ollama: OllamaClient | None = None, taxonomy: list[TaxonomyCategory] | None = None) -> None:
-        self.ollama = ollama or OllamaClient()
+    def __init__(self, ollama: LLMClient | None = None, taxonomy: list[TaxonomyCategory] | None = None) -> None:
+        # 默认按 runtime_settings.models.classify 选择 LLM 实现；
+        # 测试可通过 ollama 参数注入替身。
+        self.ollama = ollama or create_llm_client("classify")
         self.taxonomy = taxonomy if taxonomy is not None else load_taxonomy()
         # 合法类目名集合（含 uncategorized）
         self._valid_names = {c.name for c in self.taxonomy}

app/core/dedup.py

@@ -0,0 +1,216 @@
+"""文本去重策略工厂
+
+支持三种策略：
+- none：关闭去重，每次都跑完整流水线
+- sha256：精确匹配（hash 完全相同才算重复）
+- simhash：近似匹配（海明距离 <= 阈值视为重复）
+
+每种策略实现 DedupStrategy 协议：lookup(text) -> dict|None、record(text, result) -> None。
+工厂 get_dedup_strategy() 根据 runtime_settings.dedup 选择实现。
+所有策略共享同一 Redis key 前缀，但 key 后缀按策略区分避免冲突。
+"""
+
+from __future__ import annotations
+
+import hashlib
+from typing import Any, Protocol, runtime_checkable
+
+import structlog
+
+from app.core.runtime_settings import get_runtime_settings
+from app.services.redis import RedisCache
+
+logger = structlog.get_logger()
+
+# Redis 去重 key 前缀，按策略分桶
+_DEDUP_KEY_PREFIX = "dedup:"
+
+
+@runtime_checkable
+class DedupStrategy(Protocol):
+    """去重策略协议"""
+
+    async def lookup(self, text: str) -> dict[str, Any] | None:
+        """查询文本是否已入库；命中返回旧 IngestionResult（dict），未命中返回 None"""
+        ...
+
+    async def record(self, text: str, result_dict: dict[str, Any]) -> None:
+        """记录文本已入库，供后续命中复用"""
+        ...
+
+
+class NoopDedupStrategy:
+    """关闭去重：永远不命中，也不记录"""
+
+    async def lookup(self, text: str) -> dict[str, Any] | None:
+        return None
+
+    async def record(self, text: str, result_dict: dict[str, Any]) -> None:
+        return None
+
+
+class Sha256DedupStrategy:
+    """SHA256 精确去重：完全相同文本才算重复"""
+
+    def __init__(self, redis: RedisCache, ttl_seconds: int) -> None:
+        self._redis = redis
+        self._ttl = ttl_seconds
+
+    def _key(self, text: str) -> str:
+        h = hashlib.sha256(text.encode("utf-8")).hexdigest()
+        return f"{_DEDUP_KEY_PREFIX}sha256:{h}"
+
+    async def lookup(self, text: str) -> dict[str, Any] | None:
+        try:
+            return await self._redis.get_json(self._key(text))
+        except Exception:
+            logger.warning("sha256 去重查询失败，降级未命中", exc_info=True)
+            return None
+
+    async def record(self, text: str, result_dict: dict[str, Any]) -> None:
+        try:
+            await self._redis.set_json(self._key(text), result_dict, ttl=self._ttl)
+        except Exception:
+            logger.warning("sha256 去重记录写入失败", exc_info=True)
+
+
+# SimHash 实现（64 位）
+_MASK_64 = (1 << 64) - 1
+
+
+def _simhash(text: str, token_size: int = 4) -> int:
+    """计算文本 64 位 simhash
+
+    简化实现：按 token_size 字符滑窗分词，每段 md5 → 128 位 → 取低 64 位作为 hash，
+    逐位加权（hash 该位为 1 则 +1，为 0 则 -1），最终符号位定 1/0。
+    """
+    if not text:
+        return 0
+    tokens = [text[i : i + token_size] for i in range(0, len(text), token_size)]
+    weights = [1] * 64
+    vec = [0] * 64
+    for token in tokens:
+        h = int(hashlib.md5(token.encode("utf-8")).hexdigest(), 16) & _MASK_64
+        for i in range(64):
+            bit = (h >> i) & 1
+            vec[i] += weights[i] if bit else -weights[i]
+    fingerprint = 0
+    for i in range(64):
+        if vec[i] > 0:
+            fingerprint |= (1 << i)
+    return fingerprint
+
+
+def _hamming_distance(a: int, b: int) -> int:
+    return bin((a ^ b) & _MASK_64).count("1")
+
+
+class SimhashDedupStrategy:
+    """SimHash 近似去重：海明距离 <= 阈值视为重复
+
+    Redis 中存所有已入库文本的 simhash（value 为 IngestionResult + simhash）。
+    lookup 时遍历所有候选 simhash 算海明距离，找最近的一个 <= 阈值则命中。
+    注：本实现为简化版，全表扫描，适合中小规模知识库；超大规模需换 LSH 索引。
+    """
+
+    INDEX_KEY = f"{_DEDUP_KEY_PREFIX}simhash:index"  # list 形式存所有 simhash+key
+
+    def __init__(self, redis: RedisCache, ttl_seconds: int, threshold: int = 3) -> None:
+        self._redis = redis
+        self._ttl = ttl_seconds
+        self._threshold = max(0, min(64, threshold))
+
+    async def lookup(self, text: str) -> dict[str, Any] | None:
+        try:
+            fingerprint = _simhash(text)
+            # 简化：扫描所有 simhash 记录找最近的
+            # 用 set:dedup:simhash:fps 存所有 fingerprint，每个 fp 对应一个 record key
+            # 这里用 list key 简化（适合小规模）
+            index = await self._redis.get_json(self.INDEX_KEY) or {"entries": []}
+            entries = index.get("entries", [])
+            best_dist = self._threshold + 1
+            best_key: str | None = None
+            for entry in entries:
+                fp = entry.get("fingerprint")
+                key = entry.get("key")
+                if fp is None or key is None:
+                    continue
+                dist = _hamming_distance(fingerprint, int(fp))
+                if dist <= self._threshold and dist < best_dist:
+                    best_dist = dist
+                    best_key = key
+            if best_key is None:
+                return None
+            return await self._redis.get_json(best_key)
+        except Exception:
+            logger.warning("simhash 去重查询失败，降级未命中", exc_info=True)
+            return None
+
+    async def record(self, text: str, result_dict: dict[str, Any]) -> None:
+        try:
+            fingerprint = _simhash(text)
+            key = f"{_DEDUP_KEY_PREFIX}simhash:{fingerprint:016x}"
+            # 1. 写记录
+            await self._redis.set_json(key, result_dict, ttl=self._ttl)
+            # 2. 更新索引
+            index = await self._redis.get_json(self.INDEX_KEY) or {"entries": []}
+            entries = index.get("entries", [])
+            entries.append({"fingerprint": fingerprint, "key": key})
+            await self._redis.set_json(self.INDEX_KEY, {"entries": entries}, ttl=self._ttl)
+        except Exception:
+            logger.warning("simhash 去重记录写入失败", exc_info=True)
+
+
+# ---------------------------------------------------------------------------- #
+# 工厂
+# ---------------------------------------------------------------------------- #
+
+_strategy_cache: DedupStrategy | None = None
+_strategy_signature: tuple[str, int, int] | None = None  # (strategy, ttl, threshold)
+
+
+def get_dedup_strategy(redis: RedisCache | None) -> DedupStrategy:
+    """按 runtime_settings.dedup 构造去重策略
+
+    Redis 不可用时降级为 NoopDedupStrategy（关闭去重），不影响主流程。
+    策略配置变更时自动重建单例。
+    """
+    global _strategy_cache, _strategy_signature
+
+    rt = get_runtime_settings()
+    sig = (rt.dedup.strategy, rt.dedup.ttl_seconds, rt.dedup.simhash_threshold)
+
+    if _strategy_cache is not None and _strategy_signature == sig:
+        return _strategy_cache
+
+    if redis is None or rt.dedup.strategy == "none":
+        strategy: DedupStrategy = NoopDedupStrategy()
+    elif rt.dedup.strategy == "sha256":
+        strategy = Sha256DedupStrategy(redis=redis, ttl_seconds=rt.dedup.ttl_seconds)
+    elif rt.dedup.strategy == "simhash":
+        strategy = SimhashDedupStrategy(
+            redis=redis, ttl_seconds=rt.dedup.ttl_seconds, threshold=rt.dedup.simhash_threshold
+        )
+    else:
+        logger.warning("未知去重策略，降级为 noop", requested=rt.dedup.strategy)
+        strategy = NoopDedupStrategy()
+
+    _strategy_cache = strategy
+    _strategy_signature = sig
+    return strategy
+
+
+def invalidate_dedup_strategy_cache() -> None:
+    """清除策略缓存（runtime_settings 更新后调用）"""
+    global _strategy_cache, _strategy_signature
+    _strategy_cache = None
+    _strategy_signature = None
+
+
+def compute_text_hash(text: str) -> str:
+    """计算文本哈希（兼容旧接口，sha256 策略下与 dedup key 一致）"""
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+# 兼容旧测试与外部引用：保留 DEDUP_KEY_PREFIX 导出
+DEDUP_KEY_PREFIX = _DEDUP_KEY_PREFIX

app/core/file_parser.py

@@ -1,14 +1,19 @@
-"""多格式文件文本提取：按扩展名分发到对应解析器
+"""多格式文件文本提取：按扩展名分发到对应解析器（插件化）
 
 支持的扩展名：
 - .txt / .md：UTF-8 解码（errors="replace" 兜底）
 - .html / .htm：标准库 html.parser 剥离标签提取可见文本
-- .pdf：pypdf 逐页 extract_text 拼接；文本层为空（扫描件/图片型）时
-        自动降级为 OCR（pypdfium2 渲染 + rapidocr-onnxruntime 识别）
-- .docx：python-docx 段落文本拼接（不含表格/页眉页脚）
+- .pdf：PDF 文本层提取插件（默认 pypdf）；文本层为空（扫描件）时降级到 OCR 插件
+- .docx：DOCX 解析插件（默认 python-docx）
+
+插件化设计：
+- PdfTextExtractor / DocxParser / OcrEngine 三个 Protocol
+- 每种插件类型有注册表 + 默认实现 + 工厂
+- 工厂根据 runtime_settings.parsers 选择具体插件
+- 未注册或依赖缺失时降级到默认插件并告警
 
 未识别扩展名抛 ValueError("不支持的文件类型: {ext}")；
-解析异常统一包装为 ValueError("文件解析失败: {detail}")，原异常链式保留。
+解析异常统一包装为 ValueError("文件解析失败: {detail}")。
 """
 
 from __future__ import annotations
@@ -17,17 +22,19 @@ import io
 from collections.abc import Callable
 from html.parser import HTMLParser
 from pathlib import Path
-from typing import Any
+from typing import Any, Protocol, runtime_checkable
 
 import structlog
 
 from app.config import settings
+from app.core.runtime_settings import get_runtime_settings
 
 logger = structlog.get_logger()
 
-# 模块级懒加载 OCR 引擎单例（首次调用时初始化，避免无扫描件场景白白下载模型）
-_ocr_engine: Any | None = None
-_ocr_unavailable: bool = False  # 标记 OCR 依赖不可用，后续直接跳过避免重复尝试
+
+# ---------------------------------------------------------------------------- #
+# HTML / Text 解析（无插件化需求，保留原实现）
+# ---------------------------------------------------------------------------- #
 
 
 class _VisibleTextExtractor(HTMLParser):
@@ -51,13 +58,11 @@ class _VisibleTextExtractor(HTMLParser):
             self._parts.append(data)
 
     def get_text(self) -> str:
-        # 块级标签间用空格连接，再折叠多余空白
         text = " ".join(self._parts)
         return " ".join(text.split())
 
 
 def _decode_html(content: bytes) -> str:
-    """HTML 内容解码：优先 utf-8（带 BOM），失败回退 latin-1"""
     try:
         return content.decode("utf-8-sig")
     except UnicodeDecodeError:
@@ -65,108 +70,362 @@ def _decode_html(content: bytes) -> str:
 
 
 def _parse_text(content: bytes) -> str:
-    """UTF-8 解码（errors=replace 兜底），保留原字符"""
     return content.decode("utf-8", errors="replace")
 
 
 def _parse_html(content: bytes) -> str:
-    """HTML 剥离标签，保留可见文本"""
     parser = _VisibleTextExtractor()
     parser.feed(_decode_html(content))
     parser.close()
     return parser.get_text()
 
 
-def _parse_pdf(content: bytes) -> str:
-    """PDF 解析：优先 pypdf extract_text；文本层为空（扫描件）时降级 OCR
+# ---------------------------------------------------------------------------- #
+# 插件协议
+# ---------------------------------------------------------------------------- #
 
-    OCR 流程：pypdfium2 渲染每页为 PIL Image → rapidocr-onnxruntime 识别 →
-    拼接每页识别出的文本。受 settings.pdf_ocr_* 控制：开关、最大页数、DPI。
-    OCR 依赖未安装或运行异常时降级返回空字符串（由上游 upload 端点拒绝入库）。
+
+@runtime_checkable
+class PdfTextExtractor(Protocol):
+    """PDF 文本层提取插件协议"""
+
+    def extract_text(self, content: bytes) -> str:
+        """从 PDF 二进制内容提取文本层；无文本层返回空字符串"""
+        ...
+
+
+@runtime_checkable
+class DocxParser(Protocol):
+    """DOCX 解析插件协议"""
+
+    def extract_text(self, content: bytes) -> str:
+        """从 DOCX 二进制内容提取段落文本"""
+        ...
+
+
+@runtime_checkable
+class OcrEngine(Protocol):
+    """OCR 引擎插件协议"""
+
+    def ocr_pdf(self, content: bytes, *, max_pages: int, dpi: int) -> str:
+        """对 PDF 跑 OCR，返回识别文本；失败返回空字符串"""
+        ...
+
+
+# ---------------------------------------------------------------------------- #
+# 默认 PDF 插件：pypdf
+# ---------------------------------------------------------------------------- #
+
+
+class PypdfTextExtractor:
+    """pypdf 文本层提取（默认）"""
+
+    def extract_text(self, content: bytes) -> str:
+        from pypdf import PdfReader
+
+        reader = PdfReader(io.BytesIO(content))
+        parts: list[str] = []
+        for page in reader.pages:
+            text = page.extract_text() or ""
+            if text:
+                parts.append(text)
+        return "\n".join(parts).strip()
+
+
+# ---------------------------------------------------------------------------- #
+# 备选 PDF 插件：pdfplumber（lazy import，依赖缺失时不可用）
+# ---------------------------------------------------------------------------- #
+
+
+class PdfplumberTextExtractor:
+    """pdfplumber 文本层提取（备选，对复杂排版更友好）"""
+
+    def extract_text(self, content: bytes) -> str:
+        import pdfplumber  # type: ignore[import-untyped]
+
+        parts: list[str] = []
+        with pdfplumber.open(io.BytesIO(content)) as pdf:
+            for page in pdf.pages:
+                text = page.extract_text() or ""
+                if text:
+                    parts.append(text)
+        return "\n".join(parts).strip()
+
+
+# ---------------------------------------------------------------------------- #
+# 默认 DOCX 插件：python-docx
+# ---------------------------------------------------------------------------- #
+
+
+class PythonDocxParser:
+    """python-docx 段落提取（默认）"""
+
+    def extract_text(self, content: bytes) -> str:
+        from docx import Document  # type: ignore[import-untyped]
+
+        document = Document(io.BytesIO(content))
+        parts = [p.text for p in document.paragraphs if p.text and p.text.strip()]
+        return "\n".join(parts).strip()
+
+
+# ---------------------------------------------------------------------------- #
+# 默认 OCR 插件：rapidocr-onnxruntime + pypdfium2
+# ---------------------------------------------------------------------------- #
+
+
+class RapidocrOcrEngine:
+    """rapidocr-onnxruntime OCR 引擎（默认）
+
+    流程：pypdfium2 渲染每页为 PIL Image → rapidocr 识别 → 拼接。
+    类级懒加载单例（_engine/_unavailable 为类属性，跨实例共享），
+    依赖缺失时降级返回空文本。
     """
-    from pypdf import PdfReader
 
-    reader = PdfReader(io.BytesIO(content))
-    parts: list[str] = []
-    for page in reader.pages:
-        text = page.extract_text() or ""
-        if text:
-            parts.append(text)
-    text_layer = "\n".join(parts).strip()
+    _engine: Any | None = None
+    _unavailable: bool = False
+
+    def ocr_pdf(self, content: bytes, *, max_pages: int, dpi: int) -> str:
+        if self._unavailable:
+            return ""
+
+        # 1. 懒加载 OCR 引擎（类级单例，跨实例共享）
+        if self._engine is None:
+            try:
+                from rapidocr_onnxruntime import RapidOCR
+
+                RapidocrOcrEngine._engine = RapidOCR()
+                logger.info("PDF OCR 引擎已初始化", dpi=dpi)
+            except Exception:
+                RapidocrOcrEngine._unavailable = True
+                logger.warning(
+                    "OCR 依赖不可用，扫描件 PDF 将无法提取文本", exc_info=True
+                )
+                return ""
+
+        # 2. 渲染并识别
+        try:
+            import pypdfium2 as pdfium
+
+            scale = max(1.0, dpi / 72.0)
+            pdf = pdfium.PdfDocument(io.BytesIO(content))
+            total = min(len(pdf), max(1, max_pages))
+            page_texts: list[str] = []
+            for i in range(total):
+                page = pdf[i]
+                pil_image = page.render(scale=scale).to_pil()
+                result, _ = self._engine(pil_image)
+                if result:
+                    lines = [
+                        item[1]
+                        for item in result
+                        if item and len(item) >= 2 and item[1]
+                    ]
+                    if lines:
+                        page_texts.append("\n".join(lines))
+            pdf.close()
+            return "\n".join(page_texts).strip()
+        except Exception as exc:
+            logger.warning(
+                "PDF OCR 失败，降级返回空文本", error=str(exc), exc_info=True
+            )
+            return ""
+
+
+# ---------------------------------------------------------------------------- #
+# 备选 OCR 插件：tesseract（lazy import，依赖缺失时不可用）
+# ---------------------------------------------------------------------------- #
+
+
+class TesseractOcrEngine:
+    """pytesseract + pdfium OCR（备选）
+
+    需要系统安装 tesseract 二进制与语言包。
+    类级 _unavailable 单例，跨实例共享。
+    """
+
+    _unavailable: bool = False
+
+    def ocr_pdf(self, content: bytes, *, max_pages: int, dpi: int) -> str:
+        if self._unavailable:
+            return ""
+        try:
+            import pypdfium2 as pdfium
+            import pytesseract  # type: ignore[import-untyped]
+            from PIL import Image  # type: ignore[import-untyped]
+        except Exception:
+            TesseractOcrEngine._unavailable = True
+            logger.warning("tesseract 依赖不可用，降级返回空文本", exc_info=True)
+            return ""
+
+        try:
+            scale = max(1.0, dpi / 72.0)
+            pdf = pdfium.PdfDocument(io.BytesIO(content))
+            total = min(len(pdf), max(1, max_pages))
+            page_texts: list[str] = []
+            for i in range(total):
+                page = pdf[i]
+                pil_image = page.render(scale=scale).to_pil()
+                text = pytesseract.image_to_string(pil_image, lang="chi_sim+eng")
+                if text:
+                    page_texts.append(text)
+            pdf.close()
+            return "\n".join(page_texts).strip()
+        except Exception as exc:
+            logger.warning(
+                "tesseract OCR 失败，降级返回空文本", error=str(exc), exc_info=True
+            )
+            return ""
+
+
+class NoopOcrEngine:
+    """关闭 OCR 的占位插件（strategy=none 时使用）"""
+
+    def ocr_pdf(self, content: bytes, *, max_pages: int, dpi: int) -> str:
+        return ""
+
+
+# ---------------------------------------------------------------------------- #
+# 插件注册表
+# ---------------------------------------------------------------------------- #
+
+_PDF_PLUGINS: dict[str, type[PdfTextExtractor]] = {
+    "pypdf": PypdfTextExtractor,
+    "pdfplumber": PdfplumberTextExtractor,
+}
+
+_DOCX_PLUGINS: dict[str, type[DocxParser]] = {
+    "python_docx": PythonDocxParser,
+}
+
+_OCR_PLUGINS: dict[str, type[OcrEngine]] = {
+    "rapidocr": RapidocrOcrEngine,
+    "tesseract": TesseractOcrEngine,
+    "none": NoopOcrEngine,
+}
+
+
+def list_pdf_plugins() -> list[str]:
+    return list(_PDF_PLUGINS.keys())
+
+
+def list_docx_plugins() -> list[str]:
+    return list(_DOCX_PLUGINS.keys())
+
+
+def list_ocr_plugins() -> list[str]:
+    return list(_OCR_PLUGINS.keys())
+
+
+# ---------------------------------------------------------------------------- #
+# 工厂
+# ---------------------------------------------------------------------------- #
+
+# 进程级插件单例缓存
+_pdf_plugin_cache: dict[str, PdfTextExtractor] = {}
+_docx_plugin_cache: dict[str, DocxParser] = {}
+_ocr_plugin_cache: dict[str, OcrEngine] = {}
+
+
+def _get_pdf_plugin() -> PdfTextExtractor:
+    rt = get_runtime_settings()
+    name = rt.parsers.pdf.plugin or "pypdf"
+    if name in _pdf_plugin_cache:
+        return _pdf_plugin_cache[name]
+    cls = _PDF_PLUGINS.get(name)
+    if cls is None:
+        logger.warning("未知 PDF 插件，降级到 pypdf", requested=name)
+        cls = PypdfTextExtractor
+        name = "pypdf"
+    instance = cls()
+    _pdf_plugin_cache[name] = instance
+    return instance
+
+
+def _get_docx_plugin() -> DocxParser:
+    rt = get_runtime_settings()
+    name = rt.parsers.docx.plugin or "python_docx"
+    if name in _docx_plugin_cache:
+        return _docx_plugin_cache[name]
+    cls = _DOCX_PLUGINS.get(name)
+    if cls is None:
+        logger.warning("未知 DOCX 插件，降级到 python_docx", requested=name)
+        cls = PythonDocxParser
+        name = "python_docx"
+    instance = cls()
+    _docx_plugin_cache[name] = instance
+    return instance
+
+
+def _get_ocr_plugin() -> OcrEngine:
+    rt = get_runtime_settings()
+    name = rt.parsers.ocr.plugin or "rapidocr"
+    if name in _ocr_plugin_cache:
+        return _ocr_plugin_cache[name]
+    cls = _OCR_PLUGINS.get(name)
+    if cls is None:
+        logger.warning("未知 OCR 插件，降级到 rapidocr", requested=name)
+        cls = RapidocrOcrEngine
+        name = "rapidocr"
+    instance = cls()
+    _ocr_plugin_cache[name] = instance
+    return instance
+
+
+def invalidate_parser_plugin_cache(kind: str | None = None) -> None:
+    """清除插件缓存（runtime_settings 更新后调用）"""
+    if kind is None or kind == "pdf":
+        _pdf_plugin_cache.clear()
+    if kind is None or kind == "docx":
+        _docx_plugin_cache.clear()
+    if kind is None or kind == "ocr":
+        _ocr_plugin_cache.clear()
+        # 重置 RapidocrOcrEngine 类级标记，允许重新初始化
+        RapidocrOcrEngine._engine = None
+        RapidocrOcrEngine._unavailable = False
+        TesseractOcrEngine._unavailable = False
+
+
+# ---------------------------------------------------------------------------- #
+# 顶层解析函数
+# ---------------------------------------------------------------------------- #
+
+
+def _parse_pdf(content: bytes) -> str:
+    """PDF 解析：优先文本层插件；文本层为空时降级 OCR 插件"""
+    rt = get_runtime_settings()
+    pdf_plugin = _get_pdf_plugin()
+    try:
+        text_layer = pdf_plugin.extract_text(content)
+    except Exception as exc:
+        logger.warning(
+            "PDF 文本层提取失败，尝试 OCR 降级",
+            plugin=rt.parsers.pdf.plugin,
+            error=str(exc),
+            exc_info=True,
+        )
+        text_layer = ""
 
-    # 文本层非空：直接返回
     if text_layer:
         return text_layer
 
-    # 文本层为空 → 尝试 OCR 降级
+    # 文本层为空 → OCR 降级
     if not settings.pdf_ocr_enabled:
         return ""
 
-    ocr_text = _ocr_pdf(
-        content, max_pages=settings.pdf_ocr_max_pages, dpi=settings.pdf_ocr_dpi
+    ocr_plugin = _get_ocr_plugin()
+    return ocr_plugin.ocr_pdf(
+        content,
+        max_pages=rt.parsers.ocr.params.get("max_pages", settings.pdf_ocr_max_pages),
+        dpi=rt.parsers.ocr.params.get("dpi", settings.pdf_ocr_dpi),
     )
-    return ocr_text
-
-
-def _ocr_pdf(content: bytes, max_pages: int, dpi: int) -> str:
-    """对扫描件 PDF 跑 OCR：渲染每页 → 识别 → 拼接
-
-    返回空字符串的场景：依赖未安装 / 渲染或识别异常 / 无识别结果。
-    任何异常仅告警不抛出，由上游按"无法提取文本"处理。
-    """
-    global _ocr_engine, _ocr_unavailable
-
-    if _ocr_unavailable:
-        return ""
-
-    # 1. 懒加载 OCR 引擎
-    if _ocr_engine is None:
-        try:
-            from rapidocr_onnxruntime import RapidOCR
-
-            _ocr_engine = RapidOCR()
-            logger.info("PDF OCR 引擎已初始化", dpi=dpi)
-        except Exception:
-            _ocr_unavailable = True
-            logger.warning("OCR 依赖不可用，扫描件 PDF 将无法提取文本", exc_info=True)
-            return ""
-
-    # 2. 渲染并识别
-    try:
-        import pypdfium2 as pdfium
-
-        scale = max(1.0, dpi / 72.0)
-        pdf = pdfium.PdfDocument(io.BytesIO(content))
-        total = min(len(pdf), max(1, max_pages))
-        page_texts: list[str] = []
-        for i in range(total):
-            page = pdf[i]
-            pil_image = page.render(scale=scale).to_pil()
-            result, _ = _ocr_engine(pil_image)
-            if result:
-                # result: [[box, text, score], ...]，按行拼接
-                lines = [
-                    item[1] for item in result if item and len(item) >= 2 and item[1]
-                ]
-                if lines:
-                    page_texts.append("\n".join(lines))
-        pdf.close()
-        return "\n".join(page_texts).strip()
-    except Exception as exc:
-        logger.warning("PDF OCR 失败，降级返回空文本", error=str(exc), exc_info=True)
-        return ""
 
 
 def _parse_docx(content: bytes) -> str:
-    """DOCX 段落文本拼接（不含表格/页眉页脚）"""
-    from docx import Document  # type: ignore[import-untyped]
-
-    document = Document(io.BytesIO(content))
-    parts = [p.text for p in document.paragraphs if p.text and p.text.strip()]
-    return "\n".join(parts).strip()
+    """DOCX 解析"""
+    return _get_docx_plugin().extract_text(content)
 
 
-# 扩展名 → 解析函数映射（启动时构建，避免每次请求重复构造）
+# 扩展名 → 解析函数映射
 _PARSERS: dict[str, Callable[[bytes], str]] = {
     ".txt": _parse_text,
     ".md": _parse_text,

app/core/ingest_tasks.py

@@ -6,10 +6,12 @@
 内存注册表为主（记录 status/created_at/updated_at/result/error），
 Redis 为持久镜像（key: ingest_task:{task_id}），每次状态迁移同步写入；
 Redis 不可用或写入失败仅记录 warning，不影响任务执行。
+
+文本去重委托给 app.core.dedup.get_dedup_strategy()，按 runtime_settings.dedup
+选择策略（none/sha256/simhash）。
 """
 
 import asyncio
-import hashlib
 import uuid
 from datetime import UTC, datetime
 from enum import StrEnum
@@ -18,6 +20,7 @@ from typing import Any
 import structlog
 
 from app.config import Settings
+from app.core.dedup import DEDUP_KEY_PREFIX, get_dedup_strategy
 from app.core.ingestion import Ingester, IngestionError
 from app.models.document import DocumentInput
 from app.services.redis import RedisCache
@@ -26,8 +29,6 @@ logger = structlog.get_logger()
 
 # Redis 任务状态 key 前缀
 REDIS_KEY_PREFIX = "ingest_task:"
-# Redis 文本去重 key 前缀（value 为已入库文档的 IngestionResult JSON）
-DEDUP_KEY_PREFIX = "dedup:sha256:"
 
 
 class IngestTaskStatus(StrEnum):
@@ -71,17 +72,17 @@ class IngestTaskManager:
     async def submit(self, doc: DocumentInput) -> str:
         """登记入库任务并后台执行，立即返回 task_id
 
-        文本去重：基于 doc.text 的 sha256 在 Redis 中查重；命中则直接复用旧
-        IngestionResult（仅置 deduplicated=True），不重跑流水线；未命中走原
-        异步入库流程，完成后写入去重记录供后续命中复用。Redis 不可用时跳过
-        去重，按原流程执行，不影响主流程。
+        文本去重：按 runtime_settings.dedup 选择策略（none/sha256/simhash）；
+        命中则直接复用旧 IngestionResult（仅置 deduplicated=True），不重跑流水线；
+        未命中走原异步入库流程，完成后写入去重记录供后续命中复用。
+        Redis 不可用或策略=none 时跳过，按原流程执行。
         """
         task_id = uuid.uuid4().hex
         now = _utc_now_iso()
-        text_hash = hashlib.sha256(doc.text.encode("utf-8")).hexdigest()
+        dedup = get_dedup_strategy(self._redis)
 
         # 1. 去重命中：直接置 done，复用旧结果，不调 _run
-        dedup_record = await self._lookup_dedup(text_hash)
+        dedup_record = await dedup.lookup(doc.text)
         if dedup_record is not None:
             result_dict = dict(dedup_record)
             result_dict["deduplicated"] = True
@@ -111,7 +112,7 @@ class IngestTaskManager:
             "error": None,
         }
         self._schedule_mirror(task_id)
-        background = asyncio.create_task(self._run(task_id, doc, text_hash))
+        background = asyncio.create_task(self._run(task_id, doc, dedup))
         self._background_tasks.add(background)
         background.add_done_callback(self._background_tasks.discard)
         logger.info("入库任务已登记", task_id=task_id, title=doc.title)
@@ -151,7 +152,7 @@ class IngestTaskManager:
                 raise TimeoutError(f"入库任务 {task_id} 在 {timeout}s 内未进入终态")
             await asyncio.sleep(0.01)
 
-    async def _run(self, task_id: str, doc: DocumentInput, text_hash: str) -> None:
+    async def _run(self, task_id: str, doc: DocumentInput, dedup: Any) -> None:
         """后台执行入库：并发限流 + 阶段状态推进 + 结果/错误落账 + 去重记录写入"""
         async with self._semaphore:
             try:
@@ -185,34 +186,9 @@ class IngestTaskManager:
                     result=result_dict,
                 )
                 self._schedule_mirror(task_id)
-                await self._record_dedup(text_hash, result_dict)
+                await dedup.record(doc.text, result_dict)
                 logger.info("入库任务完成", task_id=task_id)
 
-    async def _lookup_dedup(self, text_hash: str) -> dict[str, Any] | None:
-        """查询文本去重记录；Redis 不可用或异常时降级为未命中"""
-        if self._redis is None:
-            return None
-        try:
-            return await self._redis.get_json(f"{DEDUP_KEY_PREFIX}{text_hash}")
-        except Exception:
-            logger.warning(
-                "去重记录查询失败，降级为未命中", text_hash=text_hash, exc_info=True
-            )
-            return None
-
-    async def _record_dedup(self, text_hash: str, result_dict: dict[str, Any]) -> None:
-        """写入文本去重记录（含完整 IngestionResult），供后续命中复用；失败仅告警"""
-        if self._redis is None:
-            return
-        try:
-            await self._redis.set_json(
-                f"{DEDUP_KEY_PREFIX}{text_hash}",
-                result_dict,
-                ttl=self._settings.ingest_task_ttl_done,
-            )
-        except Exception:
-            logger.warning("去重记录写入失败", text_hash=text_hash, exc_info=True)
-
     def _finish_failed(self, task_id: str, error: dict[str, Any]) -> None:
         """将任务置为 failed 并记录错误信息"""
         self._tasks[task_id].update(

app/core/query_parser.py

@@ -1,6 +1,7 @@
 """query 解析与分类路由模块
 
-分层 RAG 在线侧第一步：用 Ollama 小模型将用户 query 解析为结构化 JSON
+分层 RAG 在线侧第一步：用 LLM（Ollama 或 OpenAI 兼容服务，由
+runtime_settings.models.query 决定）将用户 query 解析为结构化 JSON
 （命中类目+置信度、rewrite 后 query、关键词），再由纯函数做路由决策：
 - 高置信且命中类目数 <= 上限 → 按类目过滤检索
 - 低置信 / 解析失败 / 命中类目过多 → 全库兜底（不丢召回）
@@ -16,8 +17,8 @@ import structlog
 from pydantic import BaseModel, Field, ValidationError
 
 from app.config import settings
-from app.models.knowledge import UNCATEGORIZED, TaxonomyCategory
-from app.services.ollama import OllamaClient
+from app.models.knowledge import UNCATEGORIZED, TaxonomyCategory, load_taxonomy
+from app.services.llm import LLMClient, create_llm_client
 from app.services.redis import RedisCache, get_cache
 
 logger = structlog.get_logger()
@@ -106,17 +107,22 @@ def decide_route(parsed: ParsedQuery, threshold: float, max_categories: int) ->
 
 
 class QueryParser:
-    """query 解析器：调用 Ollama 小模型将 query 解析为结构化 JSON"""
+    """query 解析器：调用 LLM 将 query 解析为结构化 JSON"""
 
     def __init__(
-        self, ollama: OllamaClient, taxonomy: list[TaxonomyCategory], cache: RedisCache | None = None
+        self,
+        ollama: LLMClient | None = None,
+        taxonomy: list[TaxonomyCategory] | None = None,
+        cache: RedisCache | None = None,
     ) -> None:
-        self.ollama = ollama
-        self.taxonomy = taxonomy
+        # 默认按 runtime_settings.models.query 选择 LLM 实现；
+        # 测试可通过 ollama 参数注入替身。
+        self.ollama = ollama or create_llm_client("query")
+        self.taxonomy = taxonomy if taxonomy is not None else load_taxonomy()
         # 解析结果缓存，缺省用全局单例；RedisCache 全操作容错，缓存不可用时退化为无缓存行为
         self.cache = cache if cache is not None else get_cache()
         # 可作为路由命中类目的名字集合（uncategorized 不可作为路由命中类目）
-        self._routable_names = {c.name for c in taxonomy if c.name != UNCATEGORIZED}
+        self._routable_names = {c.name for c in self.taxonomy if c.name != UNCATEGORIZED}
 
     async def parse(self, query: str) -> ParsedQuery:
         """调用 LLM 将 query 解析为结构化结果

app/core/result_summarizer.py

@@ -1,6 +1,7 @@
 """检索结果 AI 总结
 
-对检索返回的 chunk 命中结果，调用 Ollama 本地模型生成一段针对用户 query 的总结回答。
+对检索返回的 chunk 命中结果，调用 LLM（Ollama 或 OpenAI 兼容服务，由
+runtime_settings.models.query 决定）生成一段针对用户 query 的总结回答。
 仅基于检索结果内容，不编造未提及的信息。
 """
 
@@ -8,7 +9,7 @@ import structlog
 
 from app.config import settings
 from app.models.search import SearchHit
-from app.services.ollama import OllamaClient
+from app.services.llm import LLMClient, create_llm_client
 
 logger = structlog.get_logger()
 
@@ -16,8 +17,10 @@ logger = structlog.get_logger()
 class ResultSummarizer:
     """检索结果总结器"""
 
-    def __init__(self, ollama: OllamaClient | None = None) -> None:
-        self.ollama = ollama or OllamaClient()
+    def __init__(self, ollama: LLMClient | None = None) -> None:
+        # 默认按 runtime_settings.models.query 选择 LLM 实现；
+        # 测试可通过 ollama 参数注入替身。
+        self.ollama = ollama or create_llm_client("query")
 
     async def summarize(self, query: str, hits: list[SearchHit]) -> str:
         """对检索结果生成针对 query 的总结

app/core/retriever.py

@@ -21,7 +21,7 @@ from app.core.result_summarizer import ResultSummarizer
 from app.core.sparse import SparseEncoder
 from app.models.knowledge import load_taxonomy
 from app.models.search import ExtractedInfo, SearchHit, SearchRequest, SearchResponse
-from app.services.ollama import OllamaClient
+from app.services.llm import create_llm_client
 from app.services.qdrant import (
     COLLECTION_CHUNKS,
     COLLECTION_L1,
@@ -52,7 +52,7 @@ class Retriever:
     ) -> None:
         self.qdrant = qdrant or QdrantService()
         self.query_parser = query_parser or QueryParser(
-            ollama=OllamaClient(),
+            ollama=create_llm_client("query"),
             taxonomy=load_taxonomy(settings.taxonomy_path),
         )
         self.embedding = embedding or create_embedding_service()

app/core/runtime_settings.py

@@ -0,0 +1,224 @@
+"""运行时可调配置（runtime settings）
+
+与 `app.config.settings`（启动时从环境变量加载、不可变）互补：本模块管理运行时
+可通过管理后台动态调整的子集，持久化到 JSON 文件，启动时加载覆盖到内存单例。
+
+设计要点：
+- RuntimeSettings 只包含「可运行时调整」的字段（模型/解析插件/去重等），
+  不包含敏感或启动期固定的字段（端口、数据库连接等）。
+- 持久化路径默认 `data/runtime_settings.json`，可由 env `RUNTIME_SETTINGS_PATH` 覆盖。
+- 加载失败/文件缺失时回退到默认值，不阻塞启动。
+- 写入采用「先临时文件后 rename」原子替换，避免半写损坏。
+- 全模块只通过 `get_runtime_settings()` 访问单例，避免直接读 JSON。
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import tempfile
+from pathlib import Path
+from threading import RLock
+from typing import Any, Literal
+
+import structlog
+from pydantic import BaseModel, Field
+
+from app.config import settings
+
+logger = structlog.get_logger()
+
+# 默认持久化路径（相对工作目录）；可由 env 覆盖
+DEFAULT_RUNTIME_SETTINGS_PATH = "./data/runtime_settings.json"
+
+_lock = RLock()
+
+
+# ---------------------------------------------------------------------------- #
+# 模型配置（文档总结 / 查询 / 分类 各一份独立配置）
+# ---------------------------------------------------------------------------- #
+
+class LlmProviderConfig(BaseModel):
+    """单个用途的 LLM 提供方配置"""
+
+    provider: Literal["ollama", "openai_compatible"] = Field(
+        default="ollama", description="提供方：ollama 走 Ollama HTTP；openai_compatible 走 OpenAI 兼容 chat/completions"
+    )
+    base_url: str = Field(default="", description="服务地址；空则 ollama 用 settings.ollama_base_url，openai_compatible 用 settings.openai_base_url")
+    api_key: str = Field(default="", description="API Key（仅 openai_compatible 需要；ollama 忽略）")
+    model: str = Field(default="", description="模型名；空则 ollama 用 settings.ollama_model，openai_compatible 用 settings.embedding_model 同级（如 gpt-4o-mini）")
+    timeout: float = Field(default=120.0, description="请求超时秒数")
+    temperature: float = Field(default=0.3, description="采样温度（0~2）")
+
+
+class ModelSettings(BaseModel):
+    """模型相关运行时配置：三种用途独立配置"""
+
+    summarize: LlmProviderConfig = Field(default_factory=LlmProviderConfig, description="文档三级总结")
+    query: LlmProviderConfig = Field(default_factory=LlmProviderConfig, description="检索链路 query 解析与结果总结")
+    classify: LlmProviderConfig = Field(default_factory=LlmProviderConfig, description="文档分类判定")
+
+
+# ---------------------------------------------------------------------------- #
+# 解析插件配置
+# ---------------------------------------------------------------------------- #
+
+class PluginConfig(BaseModel):
+    """单个解析插件的选择与参数"""
+
+    plugin: str = Field(default="", description="插件名；空则用默认")
+    params: dict[str, Any] = Field(default_factory=dict, description="插件参数（透传给插件实现）")
+
+
+class ParserSettings(BaseModel):
+    """解析插件运行时配置"""
+
+    ocr: PluginConfig = Field(default_factory=lambda: PluginConfig(plugin="rapidocr"), description="OCR 插件")
+    pdf: PluginConfig = Field(default_factory=lambda: PluginConfig(plugin="pypdf"), description="PDF 文本层提取插件")
+    docx: PluginConfig = Field(default_factory=lambda: PluginConfig(plugin="python_docx"), description="DOCX 解析插件")
+
+
+# ---------------------------------------------------------------------------- #
+# 去重策略配置
+# ---------------------------------------------------------------------------- #
+
+class DedupSettings(BaseModel):
+    """文本去重策略运行时配置"""
+
+    strategy: Literal["none", "sha256", "simhash"] = Field(
+        default="sha256", description="去重策略：none 关闭；sha256 精确匹配；simhash 近似匹配"
+    )
+    simhash_threshold: int = Field(default=3, description="simhash 海明距离阈值（仅 strategy=simhash 生效，0~64）")
+    ttl_seconds: int = Field(default=86400, description="去重记录 Redis 保留秒数")
+
+
+# ---------------------------------------------------------------------------- #
+# 顶层 RuntimeSettings
+# ---------------------------------------------------------------------------- #
+
+class RuntimeSettings(BaseModel):
+    """运行时可调配置顶层模型"""
+
+    models: ModelSettings = Field(default_factory=ModelSettings)
+    parsers: ParserSettings = Field(default_factory=ParserSettings)
+    dedup: DedupSettings = Field(default_factory=DedupSettings)
+
+
+# ---------------------------------------------------------------------------- #
+# 单例 + 持久化
+# ---------------------------------------------------------------------------- #
+
+_runtime_settings: RuntimeSettings | None = None
+
+
+def _resolve_path() -> Path:
+    """解析持久化文件路径：env RUNTIME_SETTINGS_PATH > settings 自定义 > 默认"""
+    path_str = os.environ.get("RUNTIME_SETTINGS_PATH", "") or getattr(settings, "runtime_settings_path", "") or DEFAULT_RUNTIME_SETTINGS_PATH
+    return Path(path_str).expanduser().resolve()
+
+
+def _default_with_env_fallback() -> RuntimeSettings:
+    """构造默认 RuntimeSettings，并把启动 env 中已有的模型相关字段填充进去
+
+    这样首次启动（无持久化文件）时，UI 显示的不是空字符串而是 env 当前值。
+    """
+    cfg = RuntimeSettings()
+    # 模型默认值沿用 env
+    cfg.models.summarize.base_url = settings.ollama_base_url
+    cfg.models.summarize.model = settings.ollama_model
+    cfg.models.query.base_url = settings.ollama_base_url
+    cfg.models.query.model = settings.ollama_model
+    cfg.models.classify.base_url = settings.ollama_base_url
+    cfg.models.classify.model = settings.ollama_model
+    # 若 env 提供了 OpenAI key/url，预填到 openai_compatible 字段方便切换
+    if settings.openai_api_key:
+        for usage in ("summarize", "query", "classify"):
+            getattr(cfg.models, usage).api_key = settings.openai_api_key
+            getattr(cfg.models, usage).base_url = settings.openai_base_url if settings.openai_base_url else getattr(cfg.models, usage).base_url
+    return cfg
+
+
+def load_runtime_settings() -> RuntimeSettings:
+    """从磁盘加载 RuntimeSettings；文件缺失或损坏时回退到默认值（带 env 兜底）"""
+    path = _resolve_path()
+    try:
+        if path.exists():
+            data = json.loads(path.read_text(encoding="utf-8"))
+            return RuntimeSettings.model_validate(data)
+    except Exception:
+        logger.warning("RuntimeSettings 加载失败，回退默认值", path=str(path), exc_info=True)
+    return _default_with_env_fallback()
+
+
+def save_runtime_settings(cfg: RuntimeSettings) -> None:
+    """原子写入 RuntimeSettings 到磁盘（先临时文件后 rename）"""
+    path = _resolve_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    data = cfg.model_dump(mode="json")
+    # 写到同目录临时文件再 rename，避免半写损坏
+    fd, tmp_path = tempfile.mkstemp(prefix=".runtime_settings.", suffix=".tmp", dir=str(path.parent))
+    try:
+        with os.fdopen(fd, "w", encoding="utf-8") as f:
+            json.dump(data, f, ensure_ascii=False, indent=2)
+        os.replace(tmp_path, path)
+    except Exception:
+        try:
+            os.unlink(tmp_path)
+        except OSError:
+            pass
+        raise
+
+
+def get_runtime_settings() -> RuntimeSettings:
+    """获取 RuntimeSettings 单例（首次调用时加载）"""
+    global _runtime_settings
+    with _lock:
+        if _runtime_settings is None:
+            _runtime_settings = load_runtime_settings()
+        return _runtime_settings
+
+
+def update_runtime_settings(patch: dict[str, Any]) -> RuntimeSettings:
+    """以 patch 字典更新 RuntimeSettings 并持久化
+
+    支持部分更新（顶层键可缺失，缺省保留原值）。例如：
+        update_runtime_settings({"models": {"summarize": {"model": "qwen2.5:3b"}}})
+    """
+    with _lock:
+        current = get_runtime_settings()
+        merged = current.model_dump(mode="json")
+        _deep_merge(merged, patch)
+        new_cfg = RuntimeSettings.model_validate(merged)
+        save_runtime_settings(new_cfg)
+        # 替换单例后立即返回，新值对所有后续读取生效
+        global _runtime_settings
+        _runtime_settings = new_cfg
+        logger.info("RuntimeSettings 已更新并持久化", path=str(_resolve_path()))
+        return new_cfg
+
+
+def _deep_merge(target: dict[str, Any], patch: dict[str, Any]) -> None:
+    """递归把 patch 合并到 target（同 key 字典则递归，否则覆盖）"""
+    for k, v in patch.items():
+        if k in target and isinstance(target[k], dict) and isinstance(v, dict):
+            _deep_merge(target[k], v)
+        else:
+            target[k] = v
+
+
+def reload_runtime_settings() -> RuntimeSettings:
+    """强制从磁盘重新加载（管理后台触发）"""
+    with _lock:
+        global _runtime_settings
+        _runtime_settings = load_runtime_settings()
+        return _runtime_settings
+
+
+def reset_runtime_settings() -> RuntimeSettings:
+    """重置为默认值并持久化（管理后台触发）"""
+    with _lock:
+        new_cfg = _default_with_env_fallback()
+        save_runtime_settings(new_cfg)
+        global _runtime_settings
+        _runtime_settings = new_cfg
+        return new_cfg

app/core/summarizer.py

@@ -1,6 +1,7 @@
 """文档三级总结模块
 
-通过 Ollama 本地小模型对文档进行分级总结：
+通过 LLM（Ollama 或 OpenAI 兼容服务，由 runtime_settings.models.summarize 决定）
+对文档进行分级总结：
 - L1: 总结（一句话高度概括）
 - L2: 大纲（主要章节和关键主题）
 - L3: 内容大纲（每个章节的详细内容摘要）
@@ -12,7 +13,7 @@ import structlog
 
 from app.core.headings import parse_headings, render_outline
 from app.models.document import DocumentSummary, SummaryLevel
-from app.services.ollama import OllamaClient
+from app.services.llm import LLMClient, create_llm_client
 
 logger = structlog.get_logger()
 
@@ -23,8 +24,10 @@ MIN_TEXT_LENGTH_FOR_L3 = 500
 class Summarizer:
     """文档三级总结器"""
 
-    def __init__(self, ollama: OllamaClient | None = None) -> None:
-        self.ollama = ollama or OllamaClient()
+    def __init__(self, ollama: LLMClient | None = None) -> None:
+        # 默认按 runtime_settings 选择 LLM 实现（Ollama 或 OpenAI 兼容）；
+        # 测试可通过 ollama 参数注入替身。
+        self.ollama = ollama or create_llm_client("summarize")
 
     async def summarize(self, text: str, *, title: str = "") -> DocumentSummary:
         """对文档文本进行三级总结