botrun-flow-lang 5.9.301__py3-none-any.whl → 5.10.82__py3-none-any.whl

botrun_flow_lang/langgraph_agents/agents/agent_runner.py

@@ -1,174 +1,174 @@
-from typing import AsyncGenerator, Dict, List, Optional, Union, Any
-import logging
-
-from pydantic import BaseModel
-
-
-class StepsUpdateEvent(BaseModel):
-    """
-    for step in steps:
-        print("Description:", step.get("description", ""))
-        print("Status:", step.get("status", ""))
-        print("Updates:", step.get("updates", ""))
-    """
-
-    steps: List = []
-
-
-class OnNodeStreamEvent(BaseModel):
-    chunk: str
-
-
-class ChatModelEndEvent(BaseModel):
-    """
-    Chat Model End Event 資料模型
-    提供 on_chat_model_end 事件的原始資料，供呼叫端自行處理
-    """
-
-    # 原始事件資料
-    raw_output: Any = None  # 來自 event["data"]["output"]
-    raw_input: Any = None  # 來自 event["data"]["input"]
-
-    # 基本 metadata
-    langgraph_node: str = ""
-    usage_metadata: Dict = {}
-    model_name: str = ""
-
-    # chunk
-    chunk: str = ""
-
-
-MAX_RECURSION_LIMIT = 25
-
-
-# graph 是 CompiledStateGraph，不傳入型別的原因是，loading import 需要 0.5秒
-async def langgraph_runner(
-    thread_id: str,
-    init_state: dict,
-    graph,
-    need_resume: bool = False,
-    extra_config: Optional[Dict] = None,
-) -> AsyncGenerator:
-    """
-    這個 function 與 agent_runner 的差別在於，langgraph_runner 是回傳原原本本 LangGraph 的 event，而 agent_runner 是回傳 LangGraph 的 event 經過處理後的 event。
-    """
-    invoke_state = init_state
-    config = {
-        "configurable": {"thread_id": thread_id},
-        "recursion_limit": MAX_RECURSION_LIMIT,
-    }
-    if extra_config:
-        config["configurable"].update(extra_config)
-    if need_resume:
-        state_history = []
-        async for state in graph.aget_state_history(config):
-            state_history.append(state)
-
-        # 如果 state_history 的長度超過 MAX_RECURSION_LIMIT，動態調整 recursion_limit
-        if len(state_history) > MAX_RECURSION_LIMIT:
-            # 計算超出的倍數
-            multiplier = (len(state_history) - 1) // MAX_RECURSION_LIMIT
-            # 設定新的 recursion_limit 為 (multiplier + 1) * MAX_RECURSION_LIMIT
-            config["recursion_limit"] = (multiplier + 1) * MAX_RECURSION_LIMIT
-
-    async for event in graph.astream_events(
-        invoke_state,
-        config,
-        version="v2",
-    ):
-        # state = await graph.aget_state(config)
-        # print(state.config)
-
-        yield event
-
-
-# graph 是 CompiledStateGraph，不傳入型別的原因是，loading import 需要 0.5秒
-async def agent_runner(
-    thread_id: str,
-    init_state: dict,
-    graph,
-    need_resume: bool = False,
-    extra_config: Optional[Dict] = None,
-) -> AsyncGenerator[
-    Union[StepsUpdateEvent, OnNodeStreamEvent, ChatModelEndEvent], None
-]:
-    invoke_state = init_state
-    config = {
-        "configurable": {"thread_id": thread_id},
-        "recursion_limit": MAX_RECURSION_LIMIT,
-    }
-    if extra_config:
-        config["configurable"].update(extra_config)
-    if need_resume:
-        state_history = []
-        async for state in graph.aget_state_history(config):
-            state_history.append(state)
-
-        # 如果 state_history 的長度超過 MAX_RECURSION_LIMIT，動態調整 recursion_limit
-        if len(state_history) > MAX_RECURSION_LIMIT:
-            # 計算超出的倍數
-            multiplier = (len(state_history) - 1) // MAX_RECURSION_LIMIT
-            # 設定新的 recursion_limit 為 (multiplier + 1) * MAX_RECURSION_LIMIT
-            config["recursion_limit"] = (multiplier + 1) * MAX_RECURSION_LIMIT
-
-    async for event in graph.astream_events(
-        invoke_state,
-        config,
-        version="v2",
-    ):
-        if event["event"] == "on_chain_end":
-            pass
-        if event["event"] == "on_chat_model_end":
-            data = event.get("data", {})
-            logging.info(f"[Agent Runner] on_chat_model_end data: {data}")
-            metadata = event.get("metadata", {})
-            langgraph_node = metadata.get("langgraph_node", {})
-            output = data.get("output", {})
-
-            # 提取常用的 metadata（為了向後相容性）
-            usage_metadata = {}
-            model_name = ""
-
-            if hasattr(output, "usage_metadata"):
-                usage_metadata = output.usage_metadata if output.usage_metadata else {}
-
-            if hasattr(output, "response_metadata"):
-                model_name = output.response_metadata.get("model_name", "")
-
-            chat_model_end_event = ChatModelEndEvent(
-                raw_output=data.get("output", {}),
-                raw_input=data.get("input", {}),
-                langgraph_node=langgraph_node,
-                usage_metadata=usage_metadata,
-                model_name=model_name,
-            )
-            yield chat_model_end_event
-        if event["event"] == "on_chat_model_stream":
-            data = event["data"]
-            if (
-                data["chunk"].content
-                and isinstance(data["chunk"].content[0], dict)
-                and data["chunk"].content[0].get("text", "")
-            ):
-                yield OnNodeStreamEvent(chunk=data["chunk"].content[0].get("text", ""))
-            elif data["chunk"].content and isinstance(data["chunk"].content, str):
-                yield OnNodeStreamEvent(chunk=data["chunk"].content)
-
-
-# def handle_copilotkit_intermediate_state(event: dict):
-#     print("Handling copilotkit intermediate state")
-#     copilotkit_intermediate_state = event["metadata"].get(
-#         "copilotkit:emit-intermediate-state"
-#     )
-#     print(f"Intermediate state: {copilotkit_intermediate_state}")
-#     if copilotkit_intermediate_state:
-#         for intermediate_state in copilotkit_intermediate_state:
-#             if intermediate_state.get("state_key", "") == "steps":
-#                 for tool_call in event["data"]["output"].tool_calls:
-#                     if tool_call.get("name", "") == intermediate_state.get("tool", ""):
-#                         steps = tool_call["args"].get(
-#                             intermediate_state.get("tool_argument")
-#                         )
-#                         print(f"Yielding steps: {steps}")
-#                         yield StepsUpdateEvent(steps=steps)
-#     print("--------------------------------")
+from typing import AsyncGenerator, Dict, List, Optional, Union, Any
+import logging
+
+from pydantic import BaseModel
+
+
+class StepsUpdateEvent(BaseModel):
+    """
+    for step in steps:
+        print("Description:", step.get("description", ""))
+        print("Status:", step.get("status", ""))
+        print("Updates:", step.get("updates", ""))
+    """
+
+    steps: List = []
+
+
+class OnNodeStreamEvent(BaseModel):
+    chunk: str
+
+
+class ChatModelEndEvent(BaseModel):
+    """
+    Chat Model End Event 資料模型
+    提供 on_chat_model_end 事件的原始資料，供呼叫端自行處理
+    """
+
+    # 原始事件資料
+    raw_output: Any = None  # 來自 event["data"]["output"]
+    raw_input: Any = None  # 來自 event["data"]["input"]
+
+    # 基本 metadata
+    langgraph_node: str = ""
+    usage_metadata: Dict = {}
+    model_name: str = ""
+
+    # chunk
+    chunk: str = ""
+
+
+MAX_RECURSION_LIMIT = 25
+
+
+# graph 是 CompiledStateGraph，不傳入型別的原因是，loading import 需要 0.5秒
+async def langgraph_runner(
+    thread_id: str,
+    init_state: dict,
+    graph,
+    need_resume: bool = False,
+    extra_config: Optional[Dict] = None,
+) -> AsyncGenerator:
+    """
+    這個 function 與 agent_runner 的差別在於，langgraph_runner 是回傳原原本本 LangGraph 的 event，而 agent_runner 是回傳 LangGraph 的 event 經過處理後的 event。
+    """
+    invoke_state = init_state
+    config = {
+        "configurable": {"thread_id": thread_id},
+        "recursion_limit": MAX_RECURSION_LIMIT,
+    }
+    if extra_config:
+        config["configurable"].update(extra_config)
+    if need_resume:
+        state_history = []
+        async for state in graph.aget_state_history(config):
+            state_history.append(state)
+
+        # 如果 state_history 的長度超過 MAX_RECURSION_LIMIT，動態調整 recursion_limit
+        if len(state_history) > MAX_RECURSION_LIMIT:
+            # 計算超出的倍數
+            multiplier = (len(state_history) - 1) // MAX_RECURSION_LIMIT
+            # 設定新的 recursion_limit 為 (multiplier + 1) * MAX_RECURSION_LIMIT
+            config["recursion_limit"] = (multiplier + 1) * MAX_RECURSION_LIMIT
+
+    async for event in graph.astream_events(
+        invoke_state,
+        config,
+        version="v2",
+    ):
+        # state = await graph.aget_state(config)
+        # print(state.config)
+
+        yield event
+
+
+# graph 是 CompiledStateGraph，不傳入型別的原因是，loading import 需要 0.5秒
+async def agent_runner(
+    thread_id: str,
+    init_state: dict,
+    graph,
+    need_resume: bool = False,
+    extra_config: Optional[Dict] = None,
+) -> AsyncGenerator[
+    Union[StepsUpdateEvent, OnNodeStreamEvent, ChatModelEndEvent], None
+]:
+    invoke_state = init_state
+    config = {
+        "configurable": {"thread_id": thread_id},
+        "recursion_limit": MAX_RECURSION_LIMIT,
+    }
+    if extra_config:
+        config["configurable"].update(extra_config)
+    if need_resume:
+        state_history = []
+        async for state in graph.aget_state_history(config):
+            state_history.append(state)
+
+        # 如果 state_history 的長度超過 MAX_RECURSION_LIMIT，動態調整 recursion_limit
+        if len(state_history) > MAX_RECURSION_LIMIT:
+            # 計算超出的倍數
+            multiplier = (len(state_history) - 1) // MAX_RECURSION_LIMIT
+            # 設定新的 recursion_limit 為 (multiplier + 1) * MAX_RECURSION_LIMIT
+            config["recursion_limit"] = (multiplier + 1) * MAX_RECURSION_LIMIT
+
+    async for event in graph.astream_events(
+        invoke_state,
+        config,
+        version="v2",
+    ):
+        if event["event"] == "on_chain_end":
+            pass
+        if event["event"] == "on_chat_model_end":
+            data = event.get("data", {})
+            logging.info(f"[Agent Runner] on_chat_model_end data: {data}")
+            metadata = event.get("metadata", {})
+            langgraph_node = metadata.get("langgraph_node", {})
+            output = data.get("output", {})
+
+            # 提取常用的 metadata（為了向後相容性）
+            usage_metadata = {}
+            model_name = ""
+
+            if hasattr(output, "usage_metadata"):
+                usage_metadata = output.usage_metadata if output.usage_metadata else {}
+
+            if hasattr(output, "response_metadata"):
+                model_name = output.response_metadata.get("model_name", "")
+
+            chat_model_end_event = ChatModelEndEvent(
+                raw_output=data.get("output", {}),
+                raw_input=data.get("input", {}),
+                langgraph_node=langgraph_node,
+                usage_metadata=usage_metadata,
+                model_name=model_name,
+            )
+            yield chat_model_end_event
+        if event["event"] == "on_chat_model_stream":
+            data = event["data"]
+            if (
+                data["chunk"].content
+                and isinstance(data["chunk"].content[0], dict)
+                and data["chunk"].content[0].get("text", "")
+            ):
+                yield OnNodeStreamEvent(chunk=data["chunk"].content[0].get("text", ""))
+            elif data["chunk"].content and isinstance(data["chunk"].content, str):
+                yield OnNodeStreamEvent(chunk=data["chunk"].content)
+
+
+# def handle_copilotkit_intermediate_state(event: dict):
+#     print("Handling copilotkit intermediate state")
+#     copilotkit_intermediate_state = event["metadata"].get(
+#         "copilotkit:emit-intermediate-state"
+#     )
+#     print(f"Intermediate state: {copilotkit_intermediate_state}")
+#     if copilotkit_intermediate_state:
+#         for intermediate_state in copilotkit_intermediate_state:
+#             if intermediate_state.get("state_key", "") == "steps":
+#                 for tool_call in event["data"]["output"].tool_calls:
+#                     if tool_call.get("name", "") == intermediate_state.get("tool", ""):
+#                         steps = tool_call["args"].get(
+#                             intermediate_state.get("tool_argument")
+#                         )
+#                         print(f"Yielding steps: {steps}")
+#                         yield StepsUpdateEvent(steps=steps)
+#     print("--------------------------------")

botrun_flow_lang/langgraph_agents/agents/agent_tools/step_planner.py

@@ -1,77 +1,77 @@
-from langchain_core.tools import tool
-import litellm
-import json
-
-SYSTEM_PROMPT = """你是一個專業的研究規劃助手。你的工作是：
-1. 分析使用者的研究需求
-2. 規劃完整的研究步驟
-3. 如果使用者指定特定步驟，你要規劃如何執行這些步驟
-
-你必須嚴格按照以下 JSON 格式回覆，不要加入任何其他文字：
-{
-    "analysis": "對使用者需求的分析",
-    "steps": [
-        {
-            "step": "步驟1",
-            "description": "詳細說明",
-            "expected_outcome": "預期結果"
-        }
-    ]
-}"""
-
-JSON_SCHEMA = {
-    "type": "object",
-    "properties": {
-        "analysis": {"type": "string", "description": "對使用者需求的分析"},
-        "steps": {
-            "type": "array",
-            "items": {
-                "type": "object",
-                "properties": {
-                    "step": {"type": "string"},
-                    "description": {"type": "string"},
-                    "expected_outcome": {"type": "string"},
-                },
-                "required": ["step", "description", "expected_outcome"],
-            },
-        },
-    },
-    "required": ["analysis", "steps"],
-}
-
-
-@tool
-def step_planner(user_input: str) -> str:
-    """
-    研究規劃工具 - 負責規劃研究步驟和執行計劃
-
-    這個工具會：
-    1. 分析使用者的研究需求
-    2. 根據現有工具規劃完整的研究步驟
-    3. 處理使用者指定的執行步驟
-
-    Args:
-        user_input (str): 使用者的研究需求或指定步驟
-
-    Returns:
-        str: JSON 格式的研究計劃，包含分析和詳細步驟
-    """
-
-    print("step_planner user_input============>", user_input)
-    response = litellm.completion(
-        model="o3-mini",
-        messages=[
-            {"role": "system", "content": SYSTEM_PROMPT},
-            {"role": "user", "content": user_input},
-        ],
-        response_format={"type": "json_object", "schema": JSON_SCHEMA},
-        reasoning_effort="high",
-    )
-
-    try:
-        # 確保回應是有效的 JSON 格式
-        plan = json.loads(response.choices[0].message.content)
-        return json.dumps(plan, ensure_ascii=False, indent=2)
-    except json.JSONDecodeError:
-        # 如果無法解析 JSON，直接返回原始回應
-        return response.choices[0].message.content
+from langchain_core.tools import tool
+import litellm
+import json
+
+SYSTEM_PROMPT = """你是一個專業的研究規劃助手。你的工作是：
+1. 分析使用者的研究需求
+2. 規劃完整的研究步驟
+3. 如果使用者指定特定步驟，你要規劃如何執行這些步驟
+
+你必須嚴格按照以下 JSON 格式回覆，不要加入任何其他文字：
+{
+    "analysis": "對使用者需求的分析",
+    "steps": [
+        {
+            "step": "步驟1",
+            "description": "詳細說明",
+            "expected_outcome": "預期結果"
+        }
+    ]
+}"""
+
+JSON_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "analysis": {"type": "string", "description": "對使用者需求的分析"},
+        "steps": {
+            "type": "array",
+            "items": {
+                "type": "object",
+                "properties": {
+                    "step": {"type": "string"},
+                    "description": {"type": "string"},
+                    "expected_outcome": {"type": "string"},
+                },
+                "required": ["step", "description", "expected_outcome"],
+            },
+        },
+    },
+    "required": ["analysis", "steps"],
+}
+
+
+@tool
+def step_planner(user_input: str) -> str:
+    """
+    研究規劃工具 - 負責規劃研究步驟和執行計劃
+
+    這個工具會：
+    1. 分析使用者的研究需求
+    2. 根據現有工具規劃完整的研究步驟
+    3. 處理使用者指定的執行步驟
+
+    Args:
+        user_input (str): 使用者的研究需求或指定步驟
+
+    Returns:
+        str: JSON 格式的研究計劃，包含分析和詳細步驟
+    """
+
+    print("step_planner user_input============>", user_input)
+    response = litellm.completion(
+        model="o3-mini",
+        messages=[
+            {"role": "system", "content": SYSTEM_PROMPT},
+            {"role": "user", "content": user_input},
+        ],
+        response_format={"type": "json_object", "schema": JSON_SCHEMA},
+        reasoning_effort="high",
+    )
+
+    try:
+        # 確保回應是有效的 JSON 格式
+        plan = json.loads(response.choices[0].message.content)
+        return json.dumps(plan, ensure_ascii=False, indent=2)
+    except json.JSONDecodeError:
+        # 如果無法解析 JSON，直接返回原始回應
+        return response.choices[0].message.content