feat(write): 优化文档撰写对话输出功能并添加图片支持
- 更新 write_dialog 函数，增加相似度阈值、关键词相似度权重和温度参数 - 在对话服务中添加图片 Markdown 支持 - 调整 Web 前端的写入功能，支持流式输出和图片显示 - 优化模板内容和插入点标记
This commit is contained in:
zstar
parent
e4a4786ca3
commit
c7b162d1e8
|
|
@ -262,7 +262,7 @@ def writechat():
|
|||
def stream():
|
||||
nonlocal req, uid
|
||||
try:
|
||||
for ans in write_dialog(req["question"], req["kb_ids"], uid):
|
||||
for ans in write_dialog(req["question"], req["kb_ids"], uid, req["similarity_threshold"], req["keyword_similarity_weight"], req["temperature"]):
|
||||
yield "data:" + json.dumps({"code": 0, "message": "", "data": ans}, ensure_ascii=False) + "\n\n"
|
||||
except Exception as e:
|
||||
yield "data:" + json.dumps({"code": 500, "message": str(e), "data": {"answer": "**ERROR**: " + str(e), "reference": []}}, ensure_ascii=False) + "\n\n"
|
||||
|
|
|
|||
|
|
@ -307,8 +307,7 @@ def chat(dialog, messages, stream=True, **kwargs):
|
|||
img_url = chunk.get("image_id")
|
||||
if img_url and img_url not in processed_image_urls:
|
||||
# 生成 Markdown 字符串，alt text 可以简单设为 "image" 或 chunk ID
|
||||
alt_text = f"image_chunk_{chunk.get('chunk_id', i_int)}"
|
||||
image_markdowns.append(f"\n")
|
||||
image_markdowns.append(f"\n")
|
||||
processed_image_urls.add(img_url) # 标记为已处理
|
||||
|
||||
idx = set([kbinfos["chunks"][int(i)]["doc_id"] for i in idx])
|
||||
|
|
|
|||
|
|
@ -6,7 +6,7 @@ from rag.app.tag import label_question
|
|||
from rag.prompts import kb_prompt
|
||||
|
||||
|
||||
def write_dialog(question, kb_ids, tenant_id):
|
||||
def write_dialog(question, kb_ids, tenant_id, similarity_threshold, keyword_similarity_weight, temperature):
|
||||
"""
|
||||
处理用户搜索请求，从知识库中检索相关信息并生成回答
|
||||
|
||||
|
|
@ -44,15 +44,16 @@ def write_dialog(question, kb_ids, tenant_id):
|
|||
# 获取所有知识库的租户ID并去重
|
||||
tenant_ids = list(set([kb.tenant_id for kb in kbs]))
|
||||
# 调用检索器检索相关文档片段
|
||||
kbinfos = retriever.retrieval(question, embd_mdl, tenant_ids, kb_ids, 1, 12, 0.1, 0.3, aggs=False, rank_feature=label_question(question, kbs))
|
||||
kbinfos = retriever.retrieval(question, embd_mdl, tenant_ids, kb_ids, 1, 12, similarity_threshold, 1 - keyword_similarity_weight, aggs=False, rank_feature=label_question(question, kbs))
|
||||
# 将检索结果格式化为提示词，并确保不超过模型最大token限制
|
||||
knowledges = kb_prompt(kbinfos, max_tokens)
|
||||
|
||||
prompt = """
|
||||
角色：你是一个聪明的助手。
|
||||
任务：总结知识库中的信息并回答用户的问题。
|
||||
要求与限制：
|
||||
- 绝不要捏造内容，尤其是数字。
|
||||
- 如果知识库中的信息与用户问题无关，**只需回答：对不起，未提供相关信息。
|
||||
- 如果知识库中的信息与用户问题无关，只需回答：对不起，未提供相关信息。
|
||||
- 使用Markdown格式进行回答。
|
||||
- 使用用户提问所用的语言作答。
|
||||
- 绝不要捏造内容，尤其是数字。
|
||||
|
|
@ -65,27 +66,23 @@ def write_dialog(question, kb_ids, tenant_id):
|
|||
""" % "\n".join(knowledges)
|
||||
msg = [{"role": "user", "content": question}]
|
||||
|
||||
# 生成完成后添加回答中的引用标记
|
||||
# def decorate_answer(answer):
|
||||
# nonlocal knowledges, kbinfos, prompt
|
||||
# answer, idx = retriever.insert_citations(answer, [ck["content_ltks"] for ck in kbinfos["chunks"]], [ck["vector"] for ck in kbinfos["chunks"]], embd_mdl, tkweight=0.7, vtweight=0.3)
|
||||
# idx = set([kbinfos["chunks"][int(i)]["doc_id"] for i in idx])
|
||||
# recall_docs = [d for d in kbinfos["doc_aggs"] if d["doc_id"] in idx]
|
||||
# if not recall_docs:
|
||||
# recall_docs = kbinfos["doc_aggs"]
|
||||
# kbinfos["doc_aggs"] = recall_docs
|
||||
# refs = deepcopy(kbinfos)
|
||||
# for c in refs["chunks"]:
|
||||
# if c.get("vector"):
|
||||
# del c["vector"]
|
||||
|
||||
# if answer.lower().find("invalid key") >= 0 or answer.lower().find("invalid api") >= 0:
|
||||
# answer += " Please set LLM API-Key in 'User Setting -> Model Providers -> API-Key'"
|
||||
# refs["chunks"] = chunks_format(refs)
|
||||
# return {"answer": answer, "reference": refs}
|
||||
|
||||
answer = ""
|
||||
for ans in chat_mdl.chat_streamly(prompt, msg, {"temperature": 0.1}):
|
||||
final_answer = ""
|
||||
for ans in chat_mdl.chat_streamly(prompt, msg, {"temperature": temperature}):
|
||||
answer = ans
|
||||
final_answer = answer
|
||||
yield {"answer": answer, "reference": {}}
|
||||
# yield decorate_answer(answer)
|
||||
|
||||
# 流式返回完毕后，追加图片
|
||||
image_markdowns = []
|
||||
image_urls = set()
|
||||
for chunk in kbinfos["chunks"]:
|
||||
img_url = chunk.get("image_id")
|
||||
if img_url and img_url not in image_urls:
|
||||
image_urls.add(img_url)
|
||||
image_markdowns.append(f"\n")
|
||||
|
||||
if image_markdowns:
|
||||
final_answer += "".join(image_markdowns) + "123223"
|
||||
print("追加后的答案：", final_answer)
|
||||
yield {"answer": final_answer, "reference": {}}
|
||||
|
|
|
|||
|
|
@ -93,6 +93,7 @@ export const useSendMessageWithSse = (url: string = api.writeChat) => {
|
|||
console.info('data:', d);
|
||||
setAnswer({
|
||||
...d,
|
||||
answer: d.answer,
|
||||
conversationId: body?.conversation_id,
|
||||
});
|
||||
}
|
||||
|
|
|
|||
|
|
@ -61,7 +61,7 @@ type MarkedSpaceToken = Tokens.Space;
|
|||
|
||||
// 定义插入点标记，以便在onChange时识别并移除
|
||||
// const INSERTION_MARKER = '【AI内容将插入此处】';
|
||||
const INSERTION_MARKER = ''; // 保持为空字符串，不显示实际标记
|
||||
const INSERTION_MARKER = ''; // 改成空字符串，发现使用插入点标记体验不佳；
|
||||
|
||||
const Write = () => {
|
||||
const { t } = useTranslate('write');
|
||||
|
|
@ -72,7 +72,6 @@ const Write = () => {
|
|||
// cursorPosition 存储用户点击设定的插入点位置
|
||||
const [cursorPosition, setCursorPosition] = useState<number | null>(null);
|
||||
// showCursorIndicator 现在仅用于控制文档中是否显示 'INSERTION_MARKER'，
|
||||
// 并且一旦设置了光标位置，就希望它保持为 true，除非内容被清空或主动重置。
|
||||
const [showCursorIndicator, setShowCursorIndicator] = useState(false);
|
||||
const textAreaRef = useRef<any>(null); // Ant Design Input.TextArea 的 ref 类型
|
||||
|
||||
|
|
@ -87,16 +86,15 @@ const Write = () => {
|
|||
const [selectedKnowledgeBases, setSelectedKnowledgeBases] = useState<
|
||||
string[]
|
||||
>([]);
|
||||
const [similarityThreshold, setSimilarityThreshold] = useState<number>(0.7);
|
||||
const [similarityThreshold, setSimilarityThreshold] = useState<number>(0.2);
|
||||
const [keywordSimilarityWeight, setKeywordSimilarityWeight] =
|
||||
useState<number>(0.5);
|
||||
const [modelTemperature, setModelTemperature] = useState<number>(0.7);
|
||||
useState<number>(0.7);
|
||||
const [modelTemperature, setModelTemperature] = useState<number>(1.0);
|
||||
const [knowledgeBases, setKnowledgeBases] = useState<KnowledgeBaseItem[]>([]);
|
||||
const [isLoadingKbs, setIsLoadingKbs] = useState(false);
|
||||
const [isStreaming, setIsStreaming] = useState(false); // 标记AI是否正在流式输出
|
||||
|
||||
// 新增状态和 useRef，用于流式输出管理
|
||||
// currentStreamedAiOutput 现在将直接接收 useSendMessageWithSse 返回的累积内容
|
||||
// currentStreamedAiOutput 直接接收 useSendMessageWithSse 返回的累积内容
|
||||
const [currentStreamedAiOutput, setCurrentStreamedAiOutput] = useState('');
|
||||
// 使用 useRef 存储 AI 插入点前后的内容，以及插入点位置，避免在流式更新中出现闭包陷阱
|
||||
const contentBeforeAiInsertionRef = useRef('');
|
||||
|
|
@ -120,7 +118,7 @@ const Write = () => {
|
|||
{
|
||||
id: 'default_1_v4f',
|
||||
name: t('defaultTemplate'),
|
||||
content: `# ${t('defaultTemplateTitle')}\n ## ${t('introduction')}\n ## ${t('mainContent')}\n \n ## ${t('conclusion')}\n `,
|
||||
content: `# ${t('defaultTemplateTitle')}\n \n ## ${t('introduction')}\n \n ## ${t('mainContent')}\n \n ## ${t('conclusion')}\n `,
|
||||
isCustom: false,
|
||||
},
|
||||
{
|
||||
|
|
@ -191,8 +189,7 @@ const Write = () => {
|
|||
setContent(fallbackDefaults[0].content);
|
||||
}
|
||||
}
|
||||
// eslint-disable-next-line react-hooks/exhaustive-deps
|
||||
}, [t, getInitialDefaultTemplateDefinitions]);
|
||||
}, [selectedTemplate, getInitialDefaultTemplateDefinitions, t]);
|
||||
|
||||
useEffect(() => {
|
||||
loadOrInitializeTemplates();
|
||||
|
|
@ -213,7 +210,6 @@ const Write = () => {
|
|||
|
||||
// --- 调整流式响应处理逻辑 ---
|
||||
// 阶段1: 累积 AI 输出片段，用于实时显示（包括 <think> 标签）
|
||||
// 这个 useEffect 确保 currentStreamedAiOutput 始终是实时更新的、包含 <think> 标签的完整内容
|
||||
useEffect(() => {
|
||||
if (isStreaming && answer && answer.answer) {
|
||||
setCurrentStreamedAiOutput(answer.answer);
|
||||
|
|
@ -221,15 +217,10 @@ const Write = () => {
|
|||
}, [isStreaming, answer]);
|
||||
|
||||
// 阶段2: 当流式输出完成时 (done 为 true)
|
||||
// 这个 useEffect 负责在流式输出结束时执行清理和最终内容更新
|
||||
useEffect(() => {
|
||||
if (done) {
|
||||
setIsStreaming(false);
|
||||
setIsAiLoading(false);
|
||||
|
||||
// --- Process the final streamed AI output before committing ---
|
||||
// 关键修改：这里**必须**使用 currentStreamedAiOutput，因为它是在流式过程中积累的、包含 <think> 标签的内容
|
||||
// answer.answer 可能在 done 阶段已经提前被钩子内部清理过，所以不能依赖它来获取带标签的原始内容。
|
||||
let processedAiOutput = currentStreamedAiOutput;
|
||||
if (processedAiOutput) {
|
||||
// Regex to remove <think>...</think> including content
|
||||
|
|
@ -238,9 +229,8 @@ const Write = () => {
|
|||
'',
|
||||
);
|
||||
}
|
||||
// --- END NEW ---
|
||||
|
||||
// 将最终累积的AI内容（已处理移除<think>标签）和初始文档内容拼接，更新到主内容状态
|
||||
// 将最终累积的AI内容和初始文档内容拼接，更新到主内容状态
|
||||
setContent((prevContent) => {
|
||||
if (aiInsertionStartPosRef.current !== null) {
|
||||
// 使用 useRef 中存储的初始内容和最终处理过的 AI 输出
|
||||
|
|
@ -390,15 +380,16 @@ const Write = () => {
|
|||
const getContextContent = (
|
||||
cursorPos: number,
|
||||
currentDocumentContent: string,
|
||||
maxLength: number = 4000,
|
||||
maxLength: number = 4000, // 截取插入点上下文总共4000个字符
|
||||
) => {
|
||||
// 注意: 这里的 currentDocumentContent 传入的是 AI 提问时编辑器里的总内容，
|
||||
// 而不是 contentBeforeAiInsertionRef + contentAfterAiInsertionRef，因为可能包含标记
|
||||
const beforeCursor = currentDocumentContent.substring(0, cursorPos);
|
||||
const afterCursor = currentDocumentContent.substring(cursorPos);
|
||||
|
||||
// 使用更明显的插入点标记，这个标记是给AI看的，不是给用户看的
|
||||
const insertMarker = '[AI 内容插入点]';
|
||||
// 使用更明显的插入点标记
|
||||
// const insertMarker = '[AI 内容插入点]';
|
||||
const insertMarker = '';
|
||||
const availableLength = maxLength - insertMarker.length;
|
||||
|
||||
if (currentDocumentContent.length <= availableLength) {
|
||||
|
|
@ -431,6 +422,7 @@ const Write = () => {
|
|||
};
|
||||
};
|
||||
|
||||
// 处理问题请求提交
|
||||
const handleAiQuestionSubmit = async (
|
||||
e: React.KeyboardEvent<HTMLTextAreaElement>,
|
||||
) => {
|
||||
|
|
@ -453,9 +445,6 @@ const Write = () => {
|
|||
setIsStreaming(false); // 立即设置为false，中断流
|
||||
setIsAiLoading(false); // 确保加载状态也停止
|
||||
|
||||
// 中断时立即清除流中的 <think> 标签，并更新主内容
|
||||
// 这里使用 currentStreamedAiOutput 作为基准来构建中断时的内容，
|
||||
// 因为它是屏幕上实际显示的，包含了 <think> 标签。
|
||||
const contentToCleanOnInterrupt =
|
||||
contentBeforeAiInsertionRef.current +
|
||||
currentStreamedAiOutput +
|
||||
|
|
@ -854,8 +843,6 @@ const Write = () => {
|
|||
setContent(finalContent); // 更新主内容状态
|
||||
setCursorPosition(finalCursorPosition); // 更新光标位置状态
|
||||
// 手动设置光标位置
|
||||
// 这里不能直接操作 DOM，因为是在 setState 之后，DOM 尚未更新
|
||||
// Ant Design Input.TextArea 会在 value 更新后自动处理光标位置
|
||||
setShowCursorIndicator(true); // 用户输入时，表明已设置光标位置，持续显示标记
|
||||
}}
|
||||
onClick={(e) => {
|
||||
|
|
|
|||
Loading…
Reference in New Issue