AI重构

tools/rag/README.md

@@ -1,39 +1,62 @@
-# RAG & Map Tools
+# RAG 工具说明
 
-该目录包含了地图构建、知识库生成和向量数据库管理的相关工具。
+该目录负责知识构建与向量入库。
 
 ## 目录结构
 
-- **knowledge_base/**: 存放源文档数据。
-  - 支持格式: `.txt`, `.md`, `.pdf`
-  - 生成格式: `.json`, `.ndjson` (由 `build_knowledge_base.py` 生成)
-  
-- **map/**: 存放地图原始数据。
-  - `.osm` (OpenStreetMap 数据)
-  - `.world` (Gazebo 仿真环境数据)
-  
-- **vector_store/**: ChromaDB 向量数据库的持久化存储目录。
-
-## 脚本说明
-
-### 1. `build_knowledge_base.py`
-**功能**: 处理 `map/` 目录下的地图文件，提取地理信息和语义描述，生成知识库文件到 `knowledge_base/` 目录。
-**使用方法**:
-```bash
-python build_knowledge_base.py
+```text
+tools/rag/
+├── map/                     # 原始地图数据
+├── knowledge_base/
+│   ├── location/            # 地点知识（可选分目录）
+│   ├── pattern/             # 模式知识（任务模板）
+│   ├── rules/               # 规则知识
+│   └── *.ndjson
+├── vector_store/            # Chroma 持久化目录
+├── build_knowledge_base.py  # 从 map 构建 ndjson
+└── ingest.py                # 入库到 Chroma（多集合 + 兼容集合）
 ```
 
-### 2. `ingest.py`
-**功能**: 读取 `knowledge_base/` 中的所有文档，调用嵌入模型（Embedding Model）将其向量化，并存入 `vector_store/` 中的 ChromaDB 数据库。
-**使用方法**:
+## 运行前准备
+
 ```bash
-python ingest.py
+cd /home/huangfukk/DronePlanning
+source backend_service/venv/bin/activate
+export ORIN_IP="localhost"
 ```
-**依赖**: 需要确保后端嵌入服务（如 `llama-server`）已启动，或者配置正确的 `ORIN_IP` 环境变量。
 
-## 工作流
-1. 将地图文件放入 `map/`。
-2. 运行 `build_knowledge_base.py` 生成文本描述。
-3. 将其他补充文档放入 `knowledge_base/`。
-4. 运行 `ingest.py` 构建向量索引。
+并确保 embedding 服务可用（默认 `8090`）：
+
+```bash
+curl -s http://localhost:8090/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -d '{"input":["test"]}'
+```
+
+## 完整流程（可直接执行）
+
+```bash
+cd /home/huangfukk/DronePlanning
+source backend_service/venv/bin/activate
+
+# 1) 由 map 构建知识文本
+python tools/rag/build_knowledge_base.py
+
+# 2) 入库
+python tools/rag/ingest.py
+```
+
+## 入库结果
+
+`ingest.py` 会同时写入：
+
+- 兼容集合：`drone_docs`
+- 新集合：`location_kb`、`pattern_kb`、`rules_kb`
+
+并在 metadata 中写入 `kb_type`，支持检索时按知识域筛选。
+
+## 常见问题
+
+- **Chroma 初始化异常**：先确认使用的是项目 venv，再检查 `tools/rag/vector_store/` 是否损坏（备份后重建）。
+- **连接 embedding 失败**：检查 `ORIN_IP`、端口 8090、以及模型服务是否已启动。
 

tools/rag/ingest.py

@@ -5,7 +5,6 @@ import chromadb
 # from chromadb.utils import embedding_functions - 不再需要
 from chromadb.api.types import Documents, EmbeddingFunction, Embeddings, Embeddable
 from unstructured.partition.auto import partition
-from rich.progress import track
 import logging
 import requests # 导入requests
 import json # 导入json模块
@@ -19,6 +18,11 @@ SCRIPT_DIR = Path(__file__).resolve().parent
 KNOWLEDGE_BASE_DIR = SCRIPT_DIR / "knowledge_base"
 VECTOR_STORE_DIR = SCRIPT_DIR / "vector_store"
 COLLECTION_NAME = "drone_docs"
+COLLECTIONS_BY_KB_TYPE = {
+    "location": "location_kb",
+    "pattern": "pattern_kb",
+    "rules": "rules_kb",
+}
 # EMBEDDING_MODEL_NAME = "bge-small-zh-v1.5" # 不再需要，模型名在函数内部处理
 
 # --- 自定义远程嵌入函数 ---
@@ -66,6 +70,18 @@ class RemoteEmbeddingFunction(EmbeddingFunction[Embeddable]):
             return []
 
 
+def _infer_kb_type(file_path: Path, base_dir: Path) -> str:
+    try:
+        rel = file_path.relative_to(base_dir)
+    except ValueError:
+        return "location"
+    if len(rel.parts) > 1:
+        first = rel.parts[0].lower()
+        if first in COLLECTIONS_BY_KB_TYPE:
+            return first
+    return "location"
+
+
 def get_documents(directory: Path):
     """从知识库目录加载所有文档并进行切分"""
     documents = []
@@ -74,11 +90,12 @@ def get_documents(directory: Path):
         if file_path.is_file() and not file_path.name.startswith('.'):
             try:
                 # 对简单文本文件直接读取
+                kb_type = _infer_kb_type(file_path, directory)
                 if file_path.suffix in ['.txt', '.md']:
                     text = file_path.read_text(encoding='utf-8')
                     documents.append({
                         "text": text,
-                        "metadata": {"source": str(file_path.name)}
+                        "metadata": {"source": str(file_path.name), "kb_type": kb_type}
                     })
                     logging.info(f"成功处理文本文件: {file_path.name}")
                 # 特别处理常规的JSON文件
@@ -89,13 +106,13 @@ def get_documents(directory: Path):
                             for element in data['elements']:
                                 documents.append({
                                     "text": json.dumps(element, ensure_ascii=False),
-                                    "metadata": {"source": str(file_path.name)}
+                                    "metadata": {"source": str(file_path.name), "kb_type": kb_type}
                                 })
                             logging.info(f"成功处理JSON文件: {file_path.name}, 提取了 {len(data['elements'])} 个元素。")
                         else:
                             documents.append({
                                 "text": json.dumps(data, ensure_ascii=False),
-                                "metadata": {"source": str(file_path.name)}
+                                "metadata": {"source": str(file_path.name), "kb_type": kb_type}
                             })
                             logging.info(f"成功处理JSON文件: {file_path.name} (作为单个文档)")
                 # 新增：专门处理我们生成的 NDJSON 文件
@@ -115,7 +132,7 @@ def get_documents(directory: Path):
                                     text = json.dumps(record, ensure_ascii=False)
                                 documents.append({
                                     "text": text,
-                                    "metadata": {"source": str(file_path.name)}
+                                    "metadata": {"source": str(file_path.name), "kb_type": kb_type}
                                 })
                                 count += 1
                             except json.JSONDecodeError:
@@ -128,7 +145,7 @@ def get_documents(directory: Path):
                     for element in elements:
                         documents.append({
                             "text": element.text,
-                            "metadata": {"source": str(file_path.name)}
+                            "metadata": {"source": str(file_path.name), "kb_type": kb_type}
                         })
                     logging.info(f"成功处理文件: {file_path.name} (使用unstructured)")
             except Exception as e:
@@ -159,38 +176,48 @@ def main():
 
     client = chromadb.PersistentClient(path=str(VECTOR_STORE_DIR))
     
-    # 3. 创建或获取集合
-    logging.info(f"正在访问ChromaDB集合: {COLLECTION_NAME}")
-    collection = client.get_or_create_collection(
-        name=COLLECTION_NAME,
-        embedding_function=embedding_func
-    )
+    # 3. 创建或获取集合（兼容旧集合 + 新多知识库集合）
+    collections = {
+        "fallback": client.get_or_create_collection(name=COLLECTION_NAME, embedding_function=embedding_func)
+    }
+    for kb_type, coll_name in COLLECTIONS_BY_KB_TYPE.items():
+        collections[kb_type] = client.get_or_create_collection(name=coll_name, embedding_function=embedding_func)
 
     # 4. 将文档向量化并存入数据库
     logging.info(f"开始将 {len(docs_to_ingest)} 个文档块入库...")
-    
-    # 为了避免重复添加，可以先检查
-    # (这里为了简单，我们每次都重新添加，生产环境需要更复杂的逻辑)
-    
-    doc_texts = [doc['text'] for doc in docs_to_ingest]
-    metadatas = [doc['metadata'] for doc in docs_to_ingest]
-    ids = [f"doc_{KNOWLEDGE_BASE_DIR.name}_{i}" for i in range(len(doc_texts))]
+    grouped_docs = {}
+    for doc in docs_to_ingest:
+        kb_type = doc["metadata"].get("kb_type", "location")
+        grouped_docs.setdefault(kb_type, []).append(doc)
 
+    # 先入旧集合，保持兼容
+    all_texts = [doc['text'] for doc in docs_to_ingest]
+    all_metadatas = [doc['metadata'] for doc in docs_to_ingest]
+    all_ids = [f"doc_{KNOWLEDGE_BASE_DIR.name}_{i}" for i in range(len(all_texts))]
     try:
-        # ChromaDB的add方法会自动处理嵌入
-        collection.add(
-            documents=doc_texts,
-            metadatas=metadatas,
-            ids=ids
-        )
-        logging.info("所有文档块已成功入库！")
+        collections["fallback"].add(documents=all_texts, metadatas=all_metadatas, ids=all_ids)
+        logging.info("兼容集合 drone_docs 入库完成。")
     except Exception as e:
-        logging.error(f"向ChromaDB添加文档时出错: {e}")
+        logging.error(f"向兼容集合添加文档时出错: {e}")
 
+    for kb_type, docs in grouped_docs.items():
+        if kb_type not in collections:
+            continue
+        doc_texts = [doc["text"] for doc in docs]
+        metadatas = [doc["metadata"] for doc in docs]
+        ids = [f"doc_{kb_type}_{i}" for i in range(len(doc_texts))]
+        try:
+            collections[kb_type].add(documents=doc_texts, metadatas=metadatas, ids=ids)
+            logging.info(f"{kb_type} 集合入库完成，条目数: {len(doc_texts)}")
+        except Exception as e:
+            logging.error(f"向 {kb_type} 集合添加文档时出错: {e}")
 
-    # 验证一下
-    count = collection.count()
-    logging.info(f"数据库中现在有 {count} 个条目。")
+    for name, collection in collections.items():
+        try:
+            count = collection.count()
+            logging.info(f"集合 {name} 当前条目数: {count}")
+        except Exception:
+            pass
     
     print("\n✅ 数据入库完成！")
     print(f"知识库位于: {KNOWLEDGE_BASE_DIR}")

tools/rag/knowledge_base/pattern/scene_patterns.ndjson

@@ -0,0 +1,6 @@
+{"scene_id":"scene1_perimeter_window_ground","intent_type":"patrol_or_monitor","trigger_terms":["面前大楼","12米","外围","窗户"],"text":"地面起飞后到面前大楼约12米高度，沿外围巡查打开窗户；如发现窗户则拍照回传。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"system_checks","params":{"check_level":"comprehensive"}},{"type":"action","name":"takeoff","params":{"altitude":12}},{"type":"action","name":"rotate_search","params":{"target_class":"window"}},{"type":"condition","name":"object_detected","params":{"target_class":"window"}},{"type":"action","name":"take_photos","params":{"target_class":"window","track_time":10}}]}}}
+{"scene_id":"scene1_perimeter_person_air","intent_type":"patrol_or_monitor","trigger_terms":["这栋楼","空中","外围","人"],"text":"无人机已在空中，先调整高度后沿建筑外围巡查人员，发现目标后拍照。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"move_direction","params":{"direction":"up","distance":3}},{"type":"action","name":"rotate_search","params":{"target_class":"person"}},{"type":"condition","name":"object_detected","params":{"target_class":"person"}},{"type":"action","name":"take_photos","params":{"target_class":"person","track_time":10}}]}}}
+{"scene_id":"scene1_perimeter_trash_ground","intent_type":"patrol_or_monitor","trigger_terms":["面前大楼","杂物","外围"],"text":"地面起飞至指定高度后沿楼体外围侦察杂物堆积并拍照。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"takeoff","params":{"altitude":12}},{"type":"action","name":"rotate_search","params":{"target_class":"garbage"}},{"type":"condition","name":"object_detected","params":{"target_class":"garbage"}},{"type":"action","name":"take_photos","params":{"target_class":"garbage","track_time":8}}]}}}
+{"scene_id":"scene4_named_place_search_photo","intent_type":"search_and_photo","trigger_terms":["广场","查找","拍照"],"text":"前往命名地点后搜索目标并拍照。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"fly_to_waypoint","params":{"x":0,"y":0,"z":10,"acceptance_radius":2}},{"type":"action","name":"rotate_search","params":{"target_class":"person"}},{"type":"condition","name":"object_detected","params":{"target_class":"person"}},{"type":"action","name":"take_photos","params":{"target_class":"person","track_time":8}}]}}}
+{"scene_id":"scene4_named_place_monitor_return","intent_type":"return_or_land","trigger_terms":["广场南边","监控","返航"],"text":"先到达命名地点偏移区域执行监控，到时后返航。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"fly_to_waypoint","params":{"x":40,"y":80,"z":15,"acceptance_radius":2}},{"type":"action","name":"loiter","params":{"duration":300}},{"type":"action","name":"return_emergency","params":{"reason":"mission_complete"}}]}}}
+{"scene_id":"scene4_named_place_confirm_then_action","intent_type":"search_and_photo","trigger_terms":["确认后","拍照","返航"],"text":"在命名区域发现目标后等待人工确认，再执行拍照或返航。","pattern_json":{"root":{"type":"Sequence","name":"Sequence","children":[{"type":"action","name":"fly_to_waypoint","params":{"x":10,"y":10,"z":10,"acceptance_radius":2}},{"type":"action","name":"rotate_search","params":{"target_class":"car"}},{"type":"condition","name":"object_detected","params":{"target_class":"car"}},{"type":"action","name":"manual_confirmation","params":{}},{"type":"action","name":"take_photos","params":{"target_class":"car","track_time":8}}]}}}

tools/rag/knowledge_base/rules/flight_rules.ndjson

@@ -0,0 +1,3 @@
+{"rule_id":"rules_manual_confirmation","intent_type":"search_and_photo","trigger_terms":["我确认","等待确认","经允许"],"text":"仅当用户明确要求人工确认时才注入 manual_confirmation 节点；否则禁止主动添加。"}
+{"rule_id":"rules_direction_priority","intent_type":"generic_mission","trigger_terms":["东边","西边","北边","南边","米"],"text":"当只有方向+距离且无具体地点名词时，优先使用 move_direction；不要生成 fly_to_waypoint。"}
+{"rule_id":"rules_return_emergency_scope","intent_type":"return_or_land","trigger_terms":["返航","回到","去"],"text":"有明确目的地时使用 fly_to_waypoint；无明确目的地的立即返航才使用 return_emergency。"}