Skip to content

J
jinchat-server
Unverified 提交 55504fcd 作者: shrimp 提交者： GitHub
操作

新增加知识库测试能力 (#302)

上级 466bfb7a
正在显示 包含 294 行增加 和 125 行删除
chains/local_doc_qa.py
......@@ -11,43 +11,33 @@ import numpy as np
from utils import torch_gc
from tqdm import tqdm
DEVICE_ = EMBEDDING_DEVICE
DEVICE_ID = "0" if torch.cuda.is_available() else None
DEVICE = f"{DEVICE_}:{DEVICE_ID}" if DEVICE_ID else DEVICE_
def load_file(filepath):
def load_file(filepath, sentence_size=SENTENCE_SIZE):
if filepath.lower().endswith(".md"):
loader = UnstructuredFileLoader(filepath, mode="elements")
docs = loader.load()
elif filepath.lower().endswith(".pdf"):
loader = UnstructuredFileLoader(filepath)
textsplitter = ChineseTextSplitter(pdf=True)
textsplitter = ChineseTextSplitter(pdf=True, sentence_size=sentence_size)
docs = loader.load_and_split(textsplitter)
else:
loader = UnstructuredFileLoader(filepath, mode="elements")
textsplitter = ChineseTextSplitter(pdf=False)
textsplitter = ChineseTextSplitter(pdf=False, sentence_size=sentence_size)
docs = loader.load_and_split(text_splitter=textsplitter)
return docs
def generate_prompt(related_docs: List[str],
query: str,
def generate_prompt(related_docs: List[str], query: str,
prompt_template=PROMPT_TEMPLATE) -> str:
context = "\n".join([doc.page_content for doc in related_docs])
prompt = prompt_template.replace("{question}", query).replace("{context}", context)
return prompt
def get_docs_with_score(docs_with_score):
docs = []
for doc, score in docs_with_score:
doc.metadata["score"] = score
docs.append(doc)
return docs
def seperate_list(ls: List[int]) -> List[List[int]]:
lists = []
ls1 = [ls[0]]
......@@ -62,34 +52,44 @@ def seperate_list(ls: List[int]) -> List[List[int]]:
def similarity_search_with_score_by_vector(
self, embedding: List[float], k: int = 4,
self, embedding: List[float], k: int = 4
) -> List[Tuple[Document, float]]:
scores, indices = self.index.search(np.array([embedding], dtype=np.float32), k)
docs = []
id_set = set()
store_len = len(self.index_to_docstore_id)
for j, i in enumerate(indices[0]):
if i == -1:
if i == -1 or 0 < self.score_threshold < scores[0][j]:
# This happens when not enough docs are returned.
continue
_id = self.index_to_docstore_id[i]
doc = self.docstore.search(_id)
if not self.chunk_conent:
if not isinstance(doc, Document):
raise ValueError(f"Could not find document for id {_id}, got {doc}")
doc.metadata["score"] = int(scores[0][j])
docs.append(doc)
continue
id_set.add(i)
docs_len = len(doc.page_content)
for k in range(1, max(i, store_len-i)):
for k in range(1, max(i, store_len - i)):
break_flag = False
for l in [i + k, i - k]:
if 0 <= l < len(self.index_to_docstore_id):
_id0 = self.index_to_docstore_id[l]
doc0 = self.docstore.search(_id0)
if docs_len + len(doc0.page_content) > self.chunk_size:
break_flag=True
break_flag = True
break
elif doc0.metadata["source"] == doc.metadata["source"]:
docs_len += len(doc0.page_content)
id_set.add(l)
if break_flag:
break
if not self.chunk_conent:
return docs
if len(id_set) == 0 and self.score_threshold > 0:
return []
id_list = sorted(list(id_set))
id_lists = seperate_list(id_list)
for id_seq in id_lists:
......@@ -104,7 +104,8 @@ def similarity_search_with_score_by_vector(
if not isinstance(doc, Document):
raise ValueError(f"Could not find document for id {_id}, got {doc}")
doc_score = min([scores[0][id] for id in [indices[0].tolist().index(i) for i in id_seq if i in indices[0]]])
docs.append((doc, doc_score))
doc.metadata["score"] = int(doc_score)
docs.append(doc)
torch_gc()
return docs
......@@ -114,6 +115,8 @@ class LocalDocQA:
embeddings: object = None
top_k: int = VECTOR_SEARCH_TOP_K
chunk_size: int = CHUNK_SIZE
chunk_conent: bool = True
score_threshold: int = VECTOR_SEARCH_SCORE_THRESHOLD
def init_cfg(self,
embedding_model: str = EMBEDDING_MODEL,
......@@ -136,7 +139,8 @@ class LocalDocQA:
def init_knowledge_vector_store(self,
filepath: str or List[str],
vs_path: str or os.PathLike = None):
vs_path: str or os.PathLike = None,
sentence_size=SENTENCE_SIZE):
loaded_files = []
failed_files = []
if isinstance(filepath, str):
......@@ -146,40 +150,41 @@ class LocalDocQA:
elif os.path.isfile(filepath):
file = os.path.split(filepath)[-1]
try:
docs = load_file(filepath)
print(f"{file} 已成功加载")
docs = load_file(filepath, sentence_size)
logger.info(f"{file} 已成功加载")
loaded_files.append(filepath)
except Exception as e:
print(e)
print(f"{file} 未能成功加载")
logger.error(e)
logger.info(f"{file} 未能成功加载")
return None
elif os.path.isdir(filepath):
docs = []
for file in tqdm(os.listdir(filepath), desc="加载文件"):
fullfilepath = os.path.join(filepath, file)
try:
docs += load_file(fullfilepath)
docs += load_file(fullfilepath, sentence_size)
loaded_files.append(fullfilepath)
except Exception as e:
logger.error(e)
failed_files.append(file)
if len(failed_files) > 0:
print("以下文件未能成功加载：")
logger.info("以下文件未能成功加载：")
for file in failed_files:
print(file,end="\n")
logger.info(file, end="\n")
else:
docs = []
for file in filepath:
try:
docs += load_file(file)
print(f"{file} 已成功加载")
logger.info(f"{file} 已成功加载")
loaded_files.append(file)
except Exception as e:
print(e)
print(f"{file} 未能成功加载")
logger.error(e)
logger.info(f"{file} 未能成功加载")
if len(docs) > 0:
print("文件加载完毕，正在生成向量库")
logger.info("文件加载完毕，正在生成向量库")
if vs_path and os.path.isdir(vs_path):
vector_store = FAISS.load_local(vs_path, self.embeddings)
vector_store.add_documents(docs)
......@@ -188,38 +193,46 @@ class LocalDocQA:
if not vs_path:
vs_path = os.path.join(VS_ROOT_PATH,
f"""{os.path.splitext(file)[0]}_FAISS_{datetime.datetime.now().strftime("%Y%m%d_%H%M%S")}""")
vector_store = FAISS.from_documents(docs, self.embeddings)
vector_store = FAISS.from_documents(docs, self.embeddings) ##docs 为Document列表
torch_gc()
vector_store.save_local(vs_path)
return vs_path, loaded_files
else:
print("文件均未成功加载，请检查依赖包或替换为其他文件再次上传。")
logger.info("文件均未成功加载，请检查依赖包或替换为其他文件再次上传。")
return None, loaded_files
def get_knowledge_based_answer(self,
query,
vs_path,
chat_history=[],
streaming: bool = STREAMING):
def one_knowledge_add(self, vs_path, one_title, one_conent, one_content_segmentation, sentence_size):
try:
if not vs_path or not one_title or not one_conent:
logger.info("知识库添加错误，请确认知识库名字、标题、内容是否正确！")
return None, [one_title]
docs = [Document(page_content=one_conent+"\n", metadata={"source": one_title})]
if not one_content_segmentation:
text_splitter = ChineseTextSplitter(pdf=False, sentence_size=sentence_size)
docs = text_splitter.split_documents(docs)
if os.path.isdir(vs_path):
vector_store = FAISS.load_local(vs_path, self.embeddings)
vector_store.add_documents(docs)
else:
vector_store = FAISS.from_documents(docs, self.embeddings) ##docs 为Document列表
torch_gc()
vector_store.save_local(vs_path)
return vs_path, [one_title]
except Exception as e:
logger.error(e)
return None, [one_title]
def get_knowledge_based_answer(self, query, vs_path, chat_history=[], streaming: bool = STREAMING):
vector_store = FAISS.load_local(vs_path, self.embeddings)
FAISS.similarity_search_with_score_by_vector = similarity_search_with_score_by_vector
vector_store.chunk_size = self.chunk_size
related_docs_with_score = vector_store.similarity_search_with_score(query,
k=self.top_k)
related_docs = get_docs_with_score(related_docs_with_score)
vector_store.chunk_conent = self.chunk_conent
vector_store.score_threshold = self.score_threshold
related_docs_with_score = vector_store.similarity_search_with_score(query, k=self.top_k)
torch_gc()
prompt = generate_prompt(related_docs, query)
# if streaming:
# for result, history in self.llm._stream_call(prompt=prompt,
# history=chat_history):
# history[-1][0] = query
# response = {"query": query,
# "result": result,
# "source_documents": related_docs}
# yield response, history
# else:
prompt = generate_prompt(related_docs_with_score, query)
for result, history in self.llm._call(prompt=prompt,
history=chat_history,
streaming=streaming):
......@@ -227,10 +240,35 @@ class LocalDocQA:
history[-1][0] = query
response = {"query": query,
"result": result,
"source_documents": related_docs}
"source_documents": related_docs_with_score}
yield response, history
torch_gc()
# query 查询内容
# vs_path 知识库路径
# chunk_conent 是否启用上下文关联
# score_threshold 搜索匹配score阈值
# vector_search_top_k 搜索知识库内容条数，默认搜索5条结果
# chunk_sizes 匹配单段内容的连接上下文长度
def get_knowledge_based_conent_test(self, query, vs_path, chunk_conent,
score_threshold=VECTOR_SEARCH_SCORE_THRESHOLD,
vector_search_top_k=VECTOR_SEARCH_TOP_K, chunk_size=CHUNK_SIZE):
vector_store = FAISS.load_local(vs_path, self.embeddings)
FAISS.similarity_search_with_score_by_vector = similarity_search_with_score_by_vector
vector_store.chunk_conent = chunk_conent
vector_store.score_threshold = score_threshold
vector_store.chunk_size = chunk_size
related_docs_with_score = vector_store.similarity_search_with_score(query, k=vector_search_top_k)
if not related_docs_with_score:
response = {"query": query,
"source_documents": []}
return response, ""
torch_gc()
prompt = "\n".join([doc.page_content for doc in related_docs_with_score])
response = {"query": query,
"source_documents": related_docs_with_score}
return response, prompt
if __name__ == "__main__":
local_doc_qa = LocalDocQA()
......@@ -242,11 +280,11 @@ if __name__ == "__main__":
vs_path=vs_path,
chat_history=[],
streaming=True):
print(resp["result"][last_print_len:])
logger.info(resp["result"][last_print_len:], end="", flush=True)
last_print_len = len(resp["result"])
source_text = [f"""出处 [{inum + 1}] {os.path.split(doc.metadata['source'])[-1]}：\n\n{doc.page_content}\n\n"""
# f"""相关度：{doc.metadata['score']}\n\n"""
for inum, doc in
enumerate(resp["source_documents"])]
print("\n\n" + "\n\n".join(source_text))
logger.info("\n\n" + "\n\n".join(source_text))
pass
configs/model_config.py
......@@ -69,6 +69,9 @@ LLM_HISTORY_LEN = 3
# return top-k text chunk from vector store
VECTOR_SEARCH_TOP_K = 5
# 如果为0，则不生效，经测试小于500值的结果更精准
VECTOR_SEARCH_SCORE_THRESHOLD = 0
NLTK_DATA_PATH = os.path.join(os.path.dirname(os.path.dirname(__file__)), "nltk_data")
FLAG_USER_NAME = uuid.uuid4().hex
......@@ -79,4 +82,4 @@ llm device: {LLM_DEVICE}
embedding device: {EMBEDDING_DEVICE}
dir: {os.path.dirname(os.path.dirname(__file__))}
flagging username: {FLAG_USER_NAME}
""")
\ No newline at end of file
""")
models/chatglm_llm.py
......@@ -125,7 +125,7 @@ class ChatGLM(LLM):
prefix_encoder_file.close()
model_config.pre_seq_len = prefix_encoder_config['pre_seq_len']
model_config.prefix_projection = prefix_encoder_config['prefix_projection']
except Exception as e:
except Exception as e:
logger.error(f"加载PrefixEncoder config.json失败: {e}")
self.model = AutoModel.from_pretrained(model_name_or_path, config=model_config, trust_remote_code=True,
**kwargs)
......@@ -163,7 +163,7 @@ class ChatGLM(LLM):
new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
self.model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
self.model.transformer.prefix_encoder.float()
except Exception as e:
except Exception as e:
logger.error(f"加载PrefixEncoder模型参数失败:{e}")
self.model = self.model.eval()
......@@ -175,7 +175,7 @@ if __name__ == "__main__":
llm_device=LLM_DEVICE, )
last_print_len = 0
for resp, history in llm._call("你好", streaming=True):
logger.info(resp[last_print_len:])
logger.info(resp[last_print_len:], end="", flush=True)
last_print_len = len(resp)
for resp, history in llm._call("你好", streaming=False):
logger.info(resp)
......
textsplitter/chinese_text_splitter.py
......@@ -5,9 +5,10 @@ from configs.model_config import SENTENCE_SIZE
class ChineseTextSplitter(CharacterTextSplitter):
def __init__(self, pdf: bool = False, **kwargs):
def __init__(self, pdf: bool = False, sentence_size: int = None, **kwargs):
super().__init__(**kwargs)
self.pdf = pdf
self.sentence_size = sentence_size
def split_text1(self, text: str) -> List[str]:
if self.pdf:
......@@ -23,7 +24,7 @@ class ChineseTextSplitter(CharacterTextSplitter):
sent_list.append(ele)
return sent_list
def split_text(self, text: str) -> List[str]:
def split_text(self, text: str) -> List[str]: ##此处需要进一步优化逻辑
if self.pdf:
text = re.sub(r"\n{3,}", r"\n", text)
text = re.sub('\s', " ", text)
......@@ -38,15 +39,15 @@ class ChineseTextSplitter(CharacterTextSplitter):
# 很多规则中会考虑分号;，但是这里我把它忽略不计，破折号、英文双引号等同样忽略，需要的再做些简单调整即可。
ls = [i for i in text.split("\n") if i]
for ele in ls:
if len(ele) > SENTENCE_SIZE:
if len(ele) > self.sentence_size:
ele1 = re.sub(r'([,，.]["’”」』]{0,2})([^,，.])', r'\1\n\2', ele)
ele1_ls = ele1.split("\n")
for ele_ele1 in ele1_ls:
if len(ele_ele1) > SENTENCE_SIZE:
if len(ele_ele1) > self.sentence_size:
ele_ele2 = re.sub(r'([\n]{1,}| {2,}["’”」』]{0,2})([^\s])', r'\1\n\2', ele_ele1)
ele2_ls = ele_ele2.split("\n")
for ele_ele2 in ele2_ls:
if len(ele_ele2) > SENTENCE_SIZE:
if len(ele_ele2) > self.sentence_size:
ele_ele3 = re.sub('( ["’”」』]{0,2})([^ ])', r'\1\n\2', ele_ele2)
ele2_id = ele2_ls.index(ele_ele2)
ele2_ls = ele2_ls[:ele2_id] + [i for i in ele_ele3.split("\n") if i] + ele2_ls[
......
webui.py
......@@ -4,9 +4,10 @@ import shutil
from chains.local_doc_qa import LocalDocQA
from configs.model_config import *
import nltk
nltk.data.path = [NLTK_DATA_PATH] + nltk.data.path
def get_vs_list():
lst_default = ["新建知识库"]
if not os.path.exists(VS_ROOT_PATH):
......@@ -28,14 +29,13 @@ local_doc_qa = LocalDocQA()
flag_csv_logger = gr.CSVLogger()
def get_answer(query, vs_path, history, mode,
streaming: bool = STREAMING):
if mode == "知识库问答" and vs_path:
def get_answer(query, vs_path, history, mode, score_threshold=VECTOR_SEARCH_SCORE_THRESHOLD,
vector_search_top_k=VECTOR_SEARCH_TOP_K, chunk_conent: bool = True,
chunk_size=CHUNK_SIZE, streaming: bool = STREAMING):
if mode == "知识库问答" and os.path.exists(vs_path):
for resp, history in local_doc_qa.get_knowledge_based_answer(
query=query,
vs_path=vs_path,
chat_history=history,
streaming=streaming):
query=query, vs_path=vs_path, chat_history=history, streaming=streaming):
source = "\n\n"
source += "".join(
[f"""<details> <summary>出处 [{i + 1}] {os.path.split(doc.metadata["source"])[-1]}</summary>\n"""
......@@ -45,15 +45,34 @@ def get_answer(query, vs_path, history, mode,
enumerate(resp["source_documents"])])
history[-1][-1] += source
yield history, ""
elif mode == "知识库测试" and os.path.exists(vs_path):
resp, prompt = local_doc_qa.get_knowledge_based_conent_test(query=query, vs_path=vs_path,
score_threshold=score_threshold,
vector_search_top_k=vector_search_top_k,
chunk_conent=chunk_conent,
chunk_size=chunk_size)
if not resp["source_documents"]:
yield history + [[query,
"根据您的设定，没有匹配到任何内容，请确认您设置的score阈值是否过小或其他参数是否正确！"]], ""
else:
source = "".join(
[
f"""<details> <summary>[score值]：{doc.metadata["score"]} - （{i + 1}）[出处]： {os.path.split(doc.metadata["source"])[-1]}</summary>\n"""
f"""{doc.page_content}\n"""
f"""</details>"""
for i, doc in
enumerate(resp["source_documents"])])
history.append([query, prompt + source])
yield history, ""
else:
for resp, history in local_doc_qa.llm._call(query, history,
streaming=streaming):
for resp, history in local_doc_qa.llm._call(query, history, streaming=streaming):
history[-1][-1] = resp + (
"\n\n当前知识库为空，如需基于知识库进行问答，请先加载知识库后，再进行提问。" if mode == "知识库问答" else "")
yield history, ""
logger.info(f"flagging: username={FLAG_USER_NAME},query={query},vs_path={vs_path},mode={mode},history={history}")
flag_csv_logger.flag([query, vs_path, history, mode], username=FLAG_USER_NAME)
def init_model():
try:
local_doc_qa.init_cfg()
......@@ -66,7 +85,7 @@ def init_model():
reply = """模型未成功加载，请到页面左上角"模型配置"选项卡中重新选择后点击"加载模型"按钮"""
if str(e) == "Unknown platform: darwin":
logger.info("该报错可能因为您使用的是 macOS 操作系统，需先下载模型至本地后执行 Web UI，具体方法请参考项目 README 中本地部署方法及常见问题："
" https://github.com/imClumsyPanda/langchain-ChatGLM")
" https://github.com/imClumsyPanda/langchain-ChatGLM")
else:
logger.info(reply)
return reply
......@@ -89,19 +108,23 @@ def reinit_model(llm_model, embedding_model, llm_history_len, use_ptuning_v2, us
return history + [[None, model_status]]
def get_vector_store(vs_id, files, history):
def get_vector_store(vs_id, files, sentence_size, history, one_conent, one_content_segmentation):
vs_path = os.path.join(VS_ROOT_PATH, vs_id)
filelist = []
if not os.path.exists(os.path.join(UPLOAD_ROOT_PATH, vs_id)):
os.makedirs(os.path.join(UPLOAD_ROOT_PATH, vs_id))
for file in files:
filename = os.path.split(file.name)[-1]
shutil.move(file.name, os.path.join(UPLOAD_ROOT_PATH, vs_id, filename))
filelist.append(os.path.join(UPLOAD_ROOT_PATH, vs_id, filename))
if local_doc_qa.llm and local_doc_qa.embeddings:
vs_path, loaded_files = local_doc_qa.init_knowledge_vector_store(filelist, vs_path)
if isinstance(files, list) and one_conent is None:
for file in files:
filename = os.path.split(file.name)[-1]
shutil.move(file.name, os.path.join(UPLOAD_ROOT_PATH, vs_id, filename))
filelist.append(os.path.join(UPLOAD_ROOT_PATH, vs_id, filename))
vs_path, loaded_files = local_doc_qa.init_knowledge_vector_store(filelist, vs_path, sentence_size)
else:
vs_path, loaded_files = local_doc_qa.one_knowledge_add(vs_path, files, one_conent, one_content_segmentation,
sentence_size)
if len(loaded_files):
file_status = f"已上传 {'、'.join([os.path.split(i)[-1] for i in loaded_files])} 至知识库，并已加载知识库，请开始提问"
file_status = f"已添加 {'、'.join([os.path.split(i)[-1] for i in loaded_files])} 内容至知识库，并已加载知识库，请开始提问"
else:
file_status = "文件未成功加载，请重新上传文件"
else:
......@@ -111,7 +134,6 @@ def get_vector_store(vs_id, files, history):
return vs_path, None, history + [[None, file_status]]
def change_vs_name_input(vs_id, history):
if vs_id == "新建知识库":
return gr.update(visible=True), gr.update(visible=True), gr.update(visible=False), None, history
......@@ -119,25 +141,45 @@ def change_vs_name_input(vs_id, history):
file_status = f"已加载知识库{vs_id}，请开始提问"
return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True), os.path.join(VS_ROOT_PATH,
vs_id), history + [
[None, file_status]]
[None, file_status]]
def change_mode(mode):
def change_mode(mode, history):
if mode == "知识库问答":
return gr.update(visible=True)
return gr.update(visible=True), gr.update(visible=False), history + [[None,
"【注意】：现在是知识库问答模式，您输入的任何查询都将进行知识库查询，然后会自动整理知识库关联内容进入模型查询！！！"]]
elif mode == "知识库测试":
return gr.update(visible=True), gr.update(visible=True), [[None,
"【注意】：现在是知识库测试模式，您输入的任何查询都将进行知识库查询，并仅输出知识库匹配出的内容及相似度分值和及输入的文本源路径，查询的内容并不会进入模型查询！！！如果单条内容入库，内容如未分段，则内容越多越会稀释各查询内容与之关联的score阈值。单条内容长度在100-150左右较为合理。"]]
else:
return gr.update(visible=False)
return gr.update(visible=False), gr.update(visible=False), history
def change_chunk_conent(mode, label_conent, history):
conent = ""
if "chunk_conent" in label_conent:
conent = "搜索结果上下文关联"
elif "one_content_segmentation" in label_conent: # 这里没用上，可以先留着
conent = "内容分段入库"
if mode:
return gr.update(visible=True), history + [[None, f"【已开启{conent}】"]]
else:
return gr.update(visible=False), history + [[None, f"[已关闭{conent}]"]]
def add_vs_name(vs_name, vs_list, chatbot):
if vs_name in vs_list:
vs_status = "与已有知识库名称冲突，请重新选择其他名称后提交"
chatbot = chatbot + [[None, vs_status]]
return gr.update(visible=True), vs_list,gr.update(visible=True), gr.update(visible=True), gr.update(visible=False), chatbot
return gr.update(visible=True), vs_list, gr.update(visible=True), gr.update(visible=True), gr.update(
visible=False), chatbot
else:
vs_status = f"""已新增知识库"{vs_name}",将在上传文件并载入成功后进行存储。请在开始对话前，先完成文件上传。 """
chatbot = chatbot + [[None, vs_status]]
return gr.update(visible=True, choices= [vs_name] + vs_list, value=vs_name), [vs_name]+vs_list, gr.update(visible=False), gr.update(visible=False), gr.update(visible=True),chatbot
return gr.update(visible=True, choices=[vs_name] + vs_list, value=vs_name), [vs_name] + vs_list, gr.update(
visible=False), gr.update(visible=False), gr.update(visible=True), chatbot
block_css = """.importantButton {
background: linear-gradient(45deg, #7e0570,#5d1c99, #6e00ff) !important;
......@@ -152,7 +194,7 @@ webui_title = """
# 🎉langchain-ChatGLM WebUI🎉
👍 [https://github.com/imClumsyPanda/langchain-ChatGLM](https://github.com/imClumsyPanda/langchain-ChatGLM)
"""
default_vs = vs_list[0] if len(vs_list) > 0 else "为空"
default_vs = vs_list[0] if len(vs_list) > 1 else "为空"
init_message = f"""欢迎使用 langchain-ChatGLM Web UI！
请在右侧切换模式，目前支持直接与 LLM 模型对话或基于本地知识库问答。
......@@ -163,29 +205,30 @@ init_message = f"""欢迎使用 langchain-ChatGLM Web UI！
"""
model_status = init_model()
default_path = os.path.join(VS_ROOT_PATH, vs_list[0]) if len(vs_list) > 0 else ""
default_path = os.path.join(VS_ROOT_PATH, vs_list[0]) if len(vs_list) > 1 else ""
with gr.Blocks(css=block_css) as demo:
vs_path, file_status, model_status, vs_list = gr.State(default_path), gr.State(""), gr.State(
model_status), gr.State(vs_list)
gr.Markdown(webui_title)
with gr.Tab("对话"):
with gr.Row():
with gr.Column(scale=10):
chatbot = gr.Chatbot([[None, init_message], [None, model_status.value]],
elem_id="chat-box",
show_label=False).style(height=750)
show_label=False).style(height=650)
query = gr.Textbox(show_label=False,
placeholder="请输入提问内容，按回车进行提交").style(container=False)
with gr.Column(scale=5):
mode = gr.Radio(["LLM 对话", "知识库问答"],
label="请选择使用模式",
value="知识库问答", )
knowledge_set = gr.Accordion("知识库设定", visible=False)
vs_setting = gr.Accordion("配置知识库")
mode.change(fn=change_mode,
inputs=mode,
outputs=vs_setting)
inputs=[mode, chatbot],
outputs=[vs_setting, knowledge_set, chatbot])
with vs_setting:
select_vs = gr.Dropdown(vs_list.value,
label="请选择要加载的知识库",
......@@ -195,12 +238,96 @@ with gr.Blocks(css=block_css) as demo:
vs_name = gr.Textbox(label="请输入新建知识库名称",
lines=1,
interactive=True,
visible=True if default_path=="" else False)
vs_add = gr.Button(value="添加至知识库选项", visible=True if default_path=="" else False)
file2vs = gr.Column(visible=False if default_path=="" else True)
visible=True)
vs_add = gr.Button(value="添加至知识库选项", visible=True)
file2vs = gr.Column(visible=False)
with file2vs:
# load_vs = gr.Button("加载知识库")
gr.Markdown("向知识库中添加文件")
sentence_size = gr.Number(value=SENTENCE_SIZE, precision=0,
label="文本入库分句长度限制",
interactive=True, visible=True)
with gr.Tab("上传文件"):
files = gr.File(label="添加文件",
file_types=['.txt', '.md', '.docx', '.pdf'],
file_count="multiple",
show_label=False)
load_file_button = gr.Button("上传文件并加载知识库")
with gr.Tab("上传文件夹"):
folder_files = gr.File(label="添加文件",
# file_types=['.txt', '.md', '.docx', '.pdf'],
file_count="directory",
show_label=False)
load_folder_button = gr.Button("上传文件夹并加载知识库")
vs_add.click(fn=add_vs_name,
inputs=[vs_name, vs_list, chatbot],
outputs=[select_vs, vs_list, vs_name, vs_add, file2vs, chatbot])
select_vs.change(fn=change_vs_name_input,
inputs=[select_vs, chatbot],
outputs=[vs_name, vs_add, file2vs, vs_path, chatbot])
load_file_button.click(get_vector_store,
show_progress=True,
inputs=[select_vs, files, sentence_size, chatbot, vs_setting, file2vs],
outputs=[vs_path, files, chatbot], )
load_folder_button.click(get_vector_store,
show_progress=True,
inputs=[select_vs, folder_files, sentence_size, chatbot, vs_setting,
file2vs],
outputs=[vs_path, folder_files, chatbot], )
flag_csv_logger.setup([query, vs_path, chatbot, mode], "flagged")
query.submit(get_answer,
[query, vs_path, chatbot, mode],
[chatbot, query])
with gr.Tab("知识库测试"):
with gr.Row():
with gr.Column(scale=10):
chatbot = gr.Chatbot([[None,
"【注意】：现在是知识库测试模式，您输入的任何查询都将进行知识库查询，并仅输出知识库匹配出的内容及相似度分值和及输入的文本源路径，查询的内容并不会进入模型查询！！！如果单条内容入库，内容如未分段，则内容越多越会稀释各查询内容与之关联的score阈值。单条内容长度在100-150左右较为合理。"]],
elem_id="chat-box",
show_label=False).style(height=750)
query = gr.Textbox(show_label=False,
placeholder="请输入提问内容，按回车进行提交").style(container=False)
with gr.Column(scale=5):
mode = gr.Radio(["知识库问答", "知识库测试"],
label="请选择使用模式",
value="知识库测试", )
knowledge_set = gr.Accordion("知识库设定", visible=True)
vs_setting = gr.Accordion("配置知识库", visible=True)
mode.change(fn=change_mode,
inputs=[mode, chatbot],
outputs=[vs_setting, knowledge_set, chatbot])
with knowledge_set:
score_threshold = gr.Number(value=VECTOR_SEARCH_SCORE_THRESHOLD,
label="score阈值，分值越低匹配度越高",
precision=0, interactive=True)
vector_search_top_k = gr.Number(value=VECTOR_SEARCH_TOP_K, precision=0,
label="获取知识库内容条数", interactive=True)
chunk_conent = gr.Checkbox(value=False,
label="是否启用上下文关联",
interactive=True)
chunk_sizes = gr.Number(value=CHUNK_SIZE, precision=0,
label="匹配单段内容的连接上下文长度",
interactive=True, visible=False)
chunk_conent.change(fn=change_chunk_conent,
inputs=[chunk_conent, gr.Textbox(value="chunk_conent", visible=False), chatbot],
outputs=[chunk_sizes, chatbot])
with vs_setting:
select_vs = gr.Dropdown(vs_list.value,
label="请选择要加载的知识库",
interactive=True,
value=vs_list.value[0] if len(vs_list.value) > 0 else None)
vs_name = gr.Textbox(label="请输入新建知识库名称",
lines=1,
interactive=True,
visible=True)
vs_add = gr.Button(value="添加至知识库选项", visible=True)
file2vs = gr.Column(visible=False)
with file2vs:
# load_vs = gr.Button("加载知识库")
gr.Markdown("向知识库中添加单条内容或文件")
sentence_size = gr.Number(value=SENTENCE_SIZE, precision=0,
label="文本入库分句长度限制",
interactive=True, visible=True)
with gr.Tab("上传文件"):
files = gr.File(label="添加文件",
file_types=['.txt', '.md', '.docx', '.pdf'],
......@@ -212,38 +339,46 @@ with gr.Blocks(css=block_css) as demo:
folder_files = gr.File(label="添加文件",
# file_types=['.txt', '.md', '.docx', '.pdf'],
file_count="directory",
show_label=False
)
show_label=False)
load_folder_button = gr.Button("上传文件夹并加载知识库")
# load_vs.click(fn=)
with gr.Tab("添加单条内容"):
one_title = gr.Textbox(label="标题", placeholder="请输入要添加单条段落的标题", lines=1)
one_conent = gr.Textbox(label="内容", placeholder="请输入要添加单条段落的内容", lines=5)
one_content_segmentation = gr.Checkbox(value=True, label="禁止内容分句入库",
interactive=True)
load_conent_button = gr.Button("添加内容并加载知识库")
# 将上传的文件保存到content文件夹下,并更新下拉框
vs_add.click(fn=add_vs_name,
inputs=[vs_name, vs_list, chatbot],
outputs=[select_vs, vs_list,vs_name,vs_add, file2vs,chatbot])
outputs=[select_vs, vs_list, vs_name, vs_add, file2vs, chatbot])
select_vs.change(fn=change_vs_name_input,
inputs=[select_vs, chatbot],
outputs=[vs_name, vs_add, file2vs, vs_path, chatbot])
# 将上传的文件保存到content文件夹下,并更新下拉框
load_file_button.click(get_vector_store,
show_progress=True,
inputs=[select_vs, files, chatbot],
outputs=[vs_path, files, chatbot],
)
inputs=[select_vs, files, sentence_size, chatbot, vs_setting, file2vs],
outputs=[vs_path, files, chatbot], )
load_folder_button.click(get_vector_store,
show_progress=True,
inputs=[select_vs, folder_files, chatbot],
outputs=[vs_path, folder_files, chatbot],
)
inputs=[select_vs, folder_files, sentence_size, chatbot, vs_setting,
file2vs],
outputs=[vs_path, folder_files, chatbot], )
load_conent_button.click(get_vector_store,
show_progress=True,
inputs=[select_vs, one_title, sentence_size, chatbot,
one_conent, one_content_segmentation],
outputs=[vs_path, files, chatbot], )
flag_csv_logger.setup([query, vs_path, chatbot, mode], "flagged")
query.submit(get_answer,
[query, vs_path, chatbot, mode],
[query, vs_path, chatbot, mode, score_threshold, vector_search_top_k, chunk_conent,
chunk_sizes],
[chatbot, query])
with gr.Tab("模型配置"):
llm_model = gr.Radio(llm_model_dict_list,
label="LLM 模型",
value=LLM_MODEL,
interactive=True)
llm_history_len = gr.Slider(0,
10,
llm_history_len = gr.Slider(0, 10,
value=LLM_HISTORY_LEN,
step=1,
label="LLM 对话轮数",
......@@ -258,19 +393,12 @@ with gr.Blocks(css=block_css) as demo:
label="Embedding 模型",
value=EMBEDDING_MODEL,
interactive=True)
top_k = gr.Slider(1,
20,
value=VECTOR_SEARCH_TOP_K,
step=1,
label="向量匹配 top k",
interactive=True)
top_k = gr.Slider(1, 20, value=VECTOR_SEARCH_TOP_K, step=1,
label="向量匹配 top k", interactive=True)
load_model_button = gr.Button("重新加载模型")
load_model_button.click(reinit_model,
show_progress=True,
inputs=[llm_model, embedding_model, llm_history_len, use_ptuning_v2, use_lora, top_k,
chatbot],
outputs=chatbot
)
load_model_button.click(reinit_model, show_progress=True,
inputs=[llm_model, embedding_model, llm_history_len, use_ptuning_v2, use_lora,
top_k, chatbot], outputs=chatbot)
(demo
.queue(concurrency_count=3)
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论