feat: add GraphRAG pipeline

fastapi_app/config/settings.py

@@ -75,6 +75,48 @@ class AppSettings(BaseSettings):
     LOCAL_EMBEDDING_CUDA_VISIBLE_DEVICES: Optional[str] = None
     LOCAL_EMBEDDING_GPU_MEMORY_UTILIZATION: float = 0.3
 
+    # ── GraphRAG ──────────────────────────────────────────────────────────────
+    # 查询侧默认用较快聊天模型；索引抽取亦走同一 default_chat_model（见 settings.yaml）
+    GRAPHRAG_LLM_MODEL: str = "deepseek-v3.2"
+    GRAPHRAG_EMBEDDING_MODEL: str = "text-embedding-3-small"
+    # True 且 USE_LOCAL_EMBEDDING=1 时，查询/索引写入的 embedding 指向本地 vLLM（须与向量维度一致；从 OpenAI 维度过来的库需重建索引）
+    GRAPHRAG_USE_LOCAL_EMBEDDING_RUNTIME: bool = True
+    GRAPHRAG_OUTPUT_DIR: str = "outputs/graphrag_kb"  # workspace root, layout: {dir}/{email}/{nb_id}/
+    GRAPHRAG_CMD: str = ""                          # graphrag CLI path; auto-detected from PATH if empty
+    GRAPHRAG_CHUNK_SIZE: int = 384                  # chars per chunk; also written to settings.yaml chunks.size
+    GRAPHRAG_CHUNK_OVERLAP: int = 48
+    # 写入 prompt，偏短输出可缩短 local_search 生成时间
+    GRAPHRAG_RESPONSE_TYPE: str = "At most 4 bullet points; be concise."
+    GRAPHRAG_COMMUNITY_LEVEL: int = 1               # 低于 2 通常更快，社区上下文更少
+    GRAPHRAG_LOCAL_SEARCH_CONTEXT_MAX_TOKENS: int = 12000  # 低于 24000 可加快检索上下文组装与生成
+    GRAPHRAG_SUBGRAPH_PRUNE_ENABLED: bool = True    # run LLM subgraph pruning after each query
+    GRAPHRAG_SUBGRAPH_PRUNE_MODEL: str = "deepseek-v3.2"  # 仅单独裁剪路径使用
+    GRAPHRAG_SUBGRAPH_PRUNE_MAX_EDGES_INPUT: int = 28   # 裁剪+Judge 合并路径：输入边数上限（越小越快）
+    GRAPHRAG_SUBGRAPH_PRUNE_MAX_TOKENS: int = 512      # 单独裁剪 LLM 输出上限（若仍启用旧路径）
+    # 裁剪与 Judge 合并为一次 LLM（graphrag_chat / graphrag_kb query）
+    GRAPHRAG_PRUNE_JUDGE_MODEL: Optional[str] = None   # 为空则用 JUDGE_MODEL
+    GRAPHRAG_PRUNE_JUDGE_MAX_TOKENS: int = 768        # 合并调用输出上限（含 analysis + judge JSON）
+    GRAPHRAG_MAX_HIGHLIGHT_HINTS: int = 10          # max highlight_hints returned (0 = unlimited)
+    # 子图实体名 → Wikidata 搜索 → 在 GraphRAG query/chat 答案末尾附加简短参考（需出网）
+    GRAPHRAG_WIKIDATA_ENRICH_ENABLED: bool = True
+    GRAPHRAG_WIKIDATA_LANG: str = "zh"              # wbsearchentities + 标签/描述优先语言
+    GRAPHRAG_WIKIDATA_MAX_ENTITIES: int = 8         # 最多解析的不重复实体数
+    # HTTPS 读超时（秒）；弱网可再加大或通过 HTTP 代理访问 wikidata.org
+    GRAPHRAG_WIKIDATA_TIMEOUT_SEC: float = 45.0
+    GRAPHRAG_WIKIDATA_CONNECT_TIMEOUT_SEC: float = 10.0
+    # 对 Read timeout / 连接错误额外重试次数（每次递增短暂退避）
+    GRAPHRAG_WIKIDATA_HTTP_RETRIES: int = 2
+    GRAPHRAG_WIKIDATA_API_URL: str = "https://www.wikidata.org/w/api.php"
+
+    # ── KGGen (optional triple extraction, disabled by default) ───────────────
+    KGGEN_MODEL: str = "deepseek-v3.2"
+    KGGEN_PER_CHUNK: bool = True                    # True = per-chunk calls; False = full-text single call
+    KGGEN_LOG_CHUNK_INTERVAL: int = 10              # log every N chunks (0 = first/last only)
+
+    # ── Judge (answer confidence scoring) ─────────────────────────────────────
+    JUDGE_MODEL: str = "deepseek-v3.2"              # 单独 Judge；合并路径默认同此模型
+    JUDGE_MAX_TOKENS: int = 256                     # 弱化：更短 judge 输出
+
     class Config:
         env_file = ".env"
         env_file_encoding = "utf-8"

fastapi_app/main.py

@@ -41,6 +41,7 @@
 from fastapi.responses import FileResponse
 
 from fastapi_app.routers import auth, data_extract, files, kb, kb_embedding, paper2drawio, paper2ppt
+from fastapi_app.routers import graphrag_kb
 from fastapi_app.middleware.api_key import APIKeyMiddleware
 from fastapi_app.middleware.logging import LoggingMiddleware
 from workflow_engine.utils import get_project_root
@@ -428,6 +429,17 @@ async def _lifespan(app: FastAPI):
         os.environ["LOCAL_MINERU_API_URL"] = mineru_base_url
         os.environ["LOCAL_MINERU_MODEL"] = resolved_mineru_model
 
+    def _warmup_graphrag_imports() -> None:
+        try:
+            import graphrag.config.load_config  # noqa: F401
+            from graphrag import api as _graphrag_api  # noqa: F401
+            from graphrag.cli.query import _resolve_output_files  # noqa: F401
+            log.info("GraphRAG 相关 Python 包已预导入，可降低首次查询的 import 冷启动")
+        except ImportError as exc:
+            log.debug("GraphRAG 预导入跳过: %s", exc)
+
+    _warmup_graphrag_imports()
+
     yield
     for proc in managed_procs:
         if proc.poll() is None:
@@ -476,6 +488,8 @@ def create_app() -> FastAPI:
     app.include_router(paper2drawio.router, prefix="/api/v1", tags=["Paper2Drawio"])
     app.include_router(paper2ppt.router, prefix="/api/v1", tags=["Paper2PPT"])
     app.include_router(auth.router, prefix="/api/v1", tags=["Auth"])
+    # GraphRAG 知识库：/api/v1/graphrag-kb/{index,query,merge,chunk-snippet} → wa_graphrag_kb → wf_graphrag_kb
+    app.include_router(graphrag_kb.router, prefix="/api/v1", tags=["GraphRAG KB"])
 
     # 静态文件：/outputs 下的文件（兼容 URL 中 %40 与 磁盘 @ 两种路径）
     project_root = get_project_root()