diff --git a/src/backend/bisheng/workflow/nodes/input/input.py b/src/backend/bisheng/workflow/nodes/input/input.py index 17d16aaca..258e201a8 100644 --- a/src/backend/bisheng/workflow/nodes/input/input.py +++ b/src/backend/bisheng/workflow/nodes/input/input.py @@ -1,10 +1,16 @@ +import json import os import shutil import tempfile -from typing import Any +from typing import Any, Dict, List from loguru import logger + +from langchain_core.messages import HumanMessage, SystemMessage +from langchain_core.runnables import RunnableConfig + + from bisheng.api.services.knowledge import KnowledgeService from bisheng.api.services.knowledge_imp import decide_vectorstores, read_chunk_text from bisheng.api.services.llm import LLMService @@ -12,6 +18,7 @@ from bisheng.api.v1.schemas import FileProcessBase from bisheng.cache.utils import file_download from bisheng.chat.types import IgnoreException +from bisheng.workflow.callback.event import GuideQuestionData from bisheng.workflow.nodes.base import BaseNode @@ -35,11 +42,24 @@ def __init__(self, *args, **kwargs): if self.is_dialog_input(): new_node_params['user_input'] = self.node_params['user_input'] new_node_params['dialog_files_content'] = self.node_params.get('dialog_files_content', []) + predict_config = self.node_params.get("input_prediction_config", {}) + if predict_config.get("open", False): + self._predict_input_open = True + self._predict_count = predict_config.get("predict_count", 3) + self._history_limit = predict_config.get("history_limit", 10) + predict_model = predict_config.get("model_id") + self._predict_llm = LLMService.get_bisheng_llm( + model_id=predict_model, + temperature=predict_config.get("temperature", 0.7), + cache=False, + ) + else: + self._predict_input_open = False else: for value_info in self.node_params['form_input']: new_node_params[value_info['key']] = value_info['value'] self._node_params_map[value_info['key']] = value_info - + self.node_params = new_node_params self._image_ext = ['png', 'jpg', 'jpeg', 'bmp'] @@ -60,6 +80,8 @@ def is_dialog_input(self): def get_input_schema(self) -> Any: if self.is_dialog_input(): + if self._predict_input_open: + self._predict_input(self.exec_unique_id) user_input_info = self.node_data.get_variable_info('user_input') user_input_info.value = [ self.node_data.get_variable_info('dialog_files_content'), @@ -255,3 +277,135 @@ def parse_upload_file(self, key: str, key_info: dict, value: str) -> dict | None key_info['file_path']: original_file_path, key_info['image_file']: image_files_path } + + def _build_system_prompt(self) -> str: + """构建系统提示词""" + + return f"""你是一个智能助手，专门负责分析用户的对话历史，预测用户可能会问的下一个问题。 + +请基于以下对话历史，分析用户的意图、兴趣点和对话发展趋势，预测用户最可能问的{self._predict_count}个问题。 + +分析要点： +1. 用户的关注焦点和兴趣方向 +2. 对话的逻辑发展趋势 +3. 用户可能的深入需求 +4. 相关的延伸话题 + +请确保预测的问题： +- 与对话上下文相关 +- 符合用户的交流风格 +- 具有实际价值和可操作性 +- 按照可能性从高到低排序""" + + def _build_output_format_prompt(self) -> str: + """构建输出格式说明""" + return """ +请以JSON格式返回结果： +{ + "predicted_questions": [ + { + "question": "预测的问题内容", + "probability": "可能性评分(0-1)", + "reason": "预测理由" + } + ], + "analysis": "对话趋势分析总结" +} + +确保返回有效的JSON格式。""" + + def _get_chat_history(self) -> List[Dict]: + """获取聊天历史""" + try: + # 从图状态获取历史消息 + history_list = self.graph_state.get_history_list(self._history_limit) + + # 转换为更易处理的格式 + formatted_history = [] + for msg in history_list: + if hasattr(msg, "content"): + content = msg.content + if isinstance(content, list) and len(content) > 0: + # 处理多模态消息，只取文本部分 + text_content = "" + for item in content: + if isinstance(item, dict) and item.get("type") == "text": + text_content += item.get("text", "") + content = text_content + + role = "user" if hasattr(msg, "__class__") and "Human" in msg.__class__.__name__ else "assistant" + formatted_history.append({"role": role, "content": str(content)}) + + return formatted_history + except Exception as e: + logger.warning(f"获取聊天历史失败: {e}") + return [] + + def _format_history_for_prompt(self, history: List[Dict]) -> str: + """将历史消息格式化为提示词""" + formatted = "对话历史：\n" + for i, msg in enumerate(history, 1): + role_name = "用户" if msg["role"] == "user" else "助手" + formatted += f"{i}. {role_name}: {msg['content']}\n" + + return formatted + + def _parse_llm_output(self, output: str) -> Dict: + """解析LLM输出""" + try: + # 尝试解析JSON + # 清理可能的markdown代码块标记 + cleaned_output = output.strip() + if cleaned_output.startswith("```json"): + cleaned_output = cleaned_output[7:] + if cleaned_output.endswith("```"): + cleaned_output = cleaned_output[:-3] + cleaned_output = cleaned_output.strip() + + result = json.loads(cleaned_output) + return { + "questions": result.get("predicted_questions", []), + "analysis": result.get("analysis", ""), + } + + except Exception as e: + logger.error(f"解析LLM输出失败: {e}") + return { + "questions": [{"question": "解析预测结果失败", "probability": 0.0, "reason": str(e)}], + "analysis": "输出解析出错", + "format": "error", + "raw_output": output, + } + + def _predict_input(self, unique_id: str) -> Dict: + chat_history = self._get_chat_history() + if not chat_history: + return + self._history_used = chat_history + + # 构建提示词 + system_prompt = self._build_system_prompt() + self._system_prompt_used = system_prompt + + history_text = self._format_history_for_prompt(chat_history) + output_format = self._build_output_format_prompt() + + user_prompt = f"{history_text}\n\n{output_format}" + + # 构建消息 + messages = [SystemMessage(content=system_prompt), HumanMessage(content=user_prompt)] + + # 调用LLM + result = self._predict_llm.invoke(messages, config=RunnableConfig()) + + # 解析结果 + parsed_result = self._parse_llm_output(result.content) + + + self.callback_manager.on_guide_question( + GuideQuestionData( + node_id=self.id, + guide_question=[q["question"] for q in parsed_result["questions"]], + unique_id=unique_id, + ) + ) \ No newline at end of file diff --git a/src/backend/bisheng/workflow/nodes/node_manage.py b/src/backend/bisheng/workflow/nodes/node_manage.py index 25df59c28..44c938557 100644 --- a/src/backend/bisheng/workflow/nodes/node_manage.py +++ b/src/backend/bisheng/workflow/nodes/node_manage.py @@ -24,7 +24,7 @@ NodeType.CONDITION.value: ConditionNode, NodeType.AGENT.value: AgentNode, NodeType.CODE.value: CodeNode, - NodeType.LLM.value: LLMNode + NodeType.LLM.value: LLMNode, } diff --git a/src/frontend/platform/public/locales/en/flow.json b/src/frontend/platform/public/locales/en/flow.json index e55f7f772..8dddc35eb 100644 --- a/src/frontend/platform/public/locales/en/flow.json +++ b/src/frontend/platform/public/locales/en/flow.json @@ -165,5 +165,13 @@ "storeFilesSentInCurrentSession": "Store files sent in the current session", "displayNameTooltip": "Displayed on the user conversation page", "variableNameTooltipFile": "Used to store content entered on the user conversation page; can be selected from the temporary session file list", - "variableNameTooltipOther": "Used to store content entered on the user conversation page; can be referenced in other nodes" + "variableNameTooltipOther": "Used to store content entered on the user conversation page; can be referenced in other nodes", + "modelSettings": "Model Settings", + "model": "Model", + "modelRequired": "Model cannot be empty", + "selectModel": "Select Model", + "predictCount": "Prediction Count", + "historyCount": "History Count", + "temperature": "Temperature", + "predictionSettings": "Prediction Settings" } \ No newline at end of file diff --git a/src/frontend/platform/public/locales/zh/flow.json b/src/frontend/platform/public/locales/zh/flow.json index 03f974739..f0eaef892 100644 --- a/src/frontend/platform/public/locales/zh/flow.json +++ b/src/frontend/platform/public/locales/zh/flow.json @@ -165,5 +165,13 @@ "storeFilesSentInCurrentSession": "存储当前会话中发送的文件", "displayNameTooltip": "用户会话页面展示此名称", "variableNameTooltipFile": "用于存储用户会话页面填写的内容，可在临时会话文件列表中选择此变量", - "variableNameTooltipOther": "用于存储用户会话页面填写的内容，可在其他节点中引用此变量" + "variableNameTooltipOther": "用于存储用户会话页面填写的内容，可在其他节点中引用此变量", + "modelSettings": "模型设置", + "model": "模型", + "modelRequired": "模型不可为空", + "selectModel": "选择模型", + "predictCount": "预测问题数量", + "historyCount": "历史消息数量", + "temperature": "温度", + "predictionSettings": "预测配置" } \ No newline at end of file diff --git a/src/frontend/platform/src/controllers/API/workflow.ts b/src/frontend/platform/src/controllers/API/workflow.ts index 6cae73937..f27c54cb6 100644 --- a/src/frontend/platform/src/controllers/API/workflow.ts +++ b/src/frontend/platform/src/controllers/API/workflow.ts @@ -220,6 +220,20 @@ const workflowTemplate = [ "tab": "dialog_input", "help": "提取上传文件中的图片文件，当助手或大模型节点使用多模态大模型时，可传入此图片。" }, + { + "key": "input_prediction_config", + "label": "输入预测配置", + "type": "input_prediction_config", + "tab": "dialog_input", + "value": { + "open": false, + "model_id": "", + "temperature": 0.7, + "predict_count": 3, + "history_count": 10 + }, + "help": "根据历史消息预测用户下一个可能的输入内容。" + }, { "key": "form_input", "global": "item:form_input", @@ -955,6 +969,20 @@ const workflowTemplateEN = [ "tab": "dialog_input", "help": "Extract the image file from the uploaded file. When the assistant or large model node uses the MultiModal Machine Learning large model, this image can be passed in." }, + { + "key": "input_prediction_config", + "label": "Input Prediction Configuration", + "type": "input_prediction_config", + "tab": "dialog_input", + "value": { + "open": false, + "model_id": "", + "temperature": 0.7, + "predict_count": 3, + "history_count": 10 + }, + "help": "Predict the user's next possible input based on message history." + }, { "global": "item:form_input", "key": "form_input", diff --git a/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/NodeLogo.tsx b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/NodeLogo.tsx index e50f5e097..001daa272 100644 --- a/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/NodeLogo.tsx +++ b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/NodeLogo.tsx @@ -1,5 +1,5 @@ import { cname } from "@/components/bs-ui/utils"; -import { BookOpenTextIcon, Bot, Brain, Code2, FileDown, FileSearch, FlagTriangleRight, Hammer, Home, Keyboard, MessagesSquareIcon, Split } from "lucide-react"; +import { BookOpenTextIcon, Bot, Brain, Code2, FileDown, FileSearch, FlagTriangleRight, Hammer, Home, Keyboard, MessagesSquareIcon, Split, HelpCircle } from "lucide-react"; export const Icons = { 'start': Home, 'input': Keyboard, @@ -11,7 +11,7 @@ export const Icons = { 'agent': Bot, 'end': FlagTriangleRight, 'condition': Split, - 'report': FileDown + 'report': FileDown, } export const Colors = { 'start': 'bg-[#FFD89A]', @@ -24,7 +24,7 @@ export const Colors = { 'agent': 'bg-[#FFD89A]', 'end': 'bg-red-400', 'condition': 'bg-[#EDC9E9]', - 'report': 'bg-[#9CE4F4]' + 'report': 'bg-[#9CE4F4]', } export default function NodeLogo({ type, className = '', colorStr = '' }) { diff --git a/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/Parameter.tsx b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/Parameter.tsx index b4b2d0e11..ce60acece 100644 --- a/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/Parameter.tsx +++ b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/Parameter.tsx @@ -8,6 +8,7 @@ import HistoryNumItem from "./component/HistoryNumItem"; import InputFormItem from "./component/InputFormItem"; import InputItem from "./component/InputItem"; import InputListItem from "./component/InputListItem"; +import InputPredictionConfigItem from "./component/InputPredictionConfigItem"; import KnowledgeQaSelectItem from "./component/KnowledgeQaSelectItem"; import KnowledgeSelectItem from "./component/KnowledgeSelectItem"; import ModelItem from "./component/ModelItem"; @@ -138,6 +139,8 @@ export default function Parameter({ node, nodeId, item, onOutPutChange, onStatus return ; case 'sql_config': return ; + case 'input_prediction_config': + return ; case 'select_fileaccept': return { // group_params[0] 受input模板影响 diff --git a/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/component/InputPredictionConfigItem.tsx b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/component/InputPredictionConfigItem.tsx new file mode 100644 index 000000000..0560b62f2 --- /dev/null +++ b/src/frontend/platform/src/pages/BuildPage/flow/FlowNode/component/InputPredictionConfigItem.tsx @@ -0,0 +1,183 @@ +import { Label } from "@/components/bs-ui/label"; +import { Slider } from "@/components/bs-ui/slider"; +import { Switch } from "@/components/bs-ui/switch"; +import { QuestionTooltip } from "@/components/bs-ui/tooltip"; +import { useEffect, useState } from "react"; +import { useTranslation } from "react-i18next"; +import ModelItem from "./ModelItem"; + +interface InputPredictionConfigValues { + model_id: string; + predict_count: number; + history_count: number; + temperature: number; + open: boolean; +} + +interface InputPredictionConfigErrors { + model_id?: boolean; + temperature?: boolean; +} + +interface InputPredictionConfigData { + key: string; + label: string; + value: InputPredictionConfigValues; + required?: boolean; + help?: string; +} + +interface InputPredictionConfigItemProps { + data: InputPredictionConfigData; + onChange: (value: InputPredictionConfigValues) => void; + onValidate: (validateFn: () => string | false) => void; +} + +export default function InputPredictionConfigItem({ data, onChange, onValidate }: InputPredictionConfigItemProps) { + const { t } = useTranslation('flow'); + const [values, setValues] = useState(data.value); + const [errors, setErrors] = useState({}); + + const { model_id, predict_count, history_count, temperature, open } = values; + + // 校验方法 + const handleValidate = () => { + if (!open) return; // 开关关闭时无需校验 + + const newErrors: InputPredictionConfigErrors = {}; + const errorMessages: string[] = []; + + if (!model_id) { + newErrors.model_id = true; + errorMessages.push(t("modelRequired")); // 模型不可为空 + } + + setErrors(newErrors); + return errorMessages.length > 0 ? errorMessages[0] : false; + }; + + // 提供校验回调 + useEffect(() => { + onValidate(handleValidate); + return () => onValidate(() => { }); + }, [values, data.required]); + + const handleChange = (key: keyof InputPredictionConfigValues, value: any) => { + const newValues = { ...values, [key]: value }; + setValues(newValues); + setErrors((prev) => ({ ...prev, [key]: false })); // 清除错误状态 + onChange(newValues); + }; + + const handleModelValidate = (validateFn: () => string | false) => { + // 为 ModelItem 提供校验回调 + return () => false; // 目前不需要额外的模型校验 + }; + + return ( +
+
+ + {/* 开关 */} + handleChange("open", checked)} + /> +
+ + {/* 配置表单 */} + {open && ( +
+ {/* 模型配置组 */} +
+

+ {t("modelSettings")} +

+ + {/* 模型选择 */} +
+ handleChange("model_id", value)} + onValidate={handleModelValidate} + /> + {errors.model_id && ( +

{t("modelRequired")}

+ )} +
+ + {/* 温度设置 */} +
+ +
+ handleChange("temperature", v[0])} + className="flex-1" + /> + {temperature.toFixed(1)} +
+
+
+ + {/* 预测配置组 */} +
+

+ {t("predictionSettings")} +

+ + {/* 预测问题数量 */} +
+ +
+ handleChange("predict_count", v[0])} + className="flex-1" + /> + {predict_count} +
+
+ + {/* 历史消息数量 */} +
+ +
+ handleChange("history_count", v[0])} + className="flex-1" + /> + {history_count} +
+
+
+
+ )} +
+ ); +} \ No newline at end of file