改进

document_ai/agent.py

@@ -0,0 +1,83 @@
+from langchain.agents import Tool, load_tools
+from langchain.memory import ConversationBufferMemory
+from langchain.chat_models import ChatOpenAI
+from langchain.schema import HumanMessage, BaseMessage
+from langchain.utilities import SerpAPIWrapper
+from langchain.agents import initialize_agent
+from langchain.agents import AgentType
+from getpass import getpass
+import proto.documents_pb2
+
+import init, doc_client
+
+
+# def fake_result(str: str) -> str:
+#     print(str)
+#     return "博客名称: iVampireSP.com"
+#
+
+def search_document(question: str) -> str:
+    print("搜索请求：" + question)
+    vec = init.text_to_vector(question)
+
+    search_param = {
+        "data": [vec],
+        "anns_field": "vector",
+        "param": {"metric_type": "L2"},
+        "limit": 5,
+        "expr": "user_id == 2",
+        "output_fields": ["todo_id", "title", "source", "todo_description", "language", "text", "user_id"],
+    }
+    res = init.collection.search(**search_param)
+
+    plain_text = ""
+
+    for i in range(len(res[0])):
+        _chunk_id = res[0][i].id
+        print("正在获取分块 " + str(_chunk_id) + " 的内容...")
+
+        try:
+            _chunk_content = doc_client.stub.GetDocumentChunk(proto.documents_pb2.GetDocumentChunkByIdRequest(
+                id=_chunk_id
+            ))
+
+            plain_text += "=== \n" + _chunk_content.content + " ===\n"
+
+        except Exception as e:
+            print(e)
+    return plain_text
+
+
+tools = [
+    # Tool(
+    #     name="Get Blog Name",
+    #     func=fake_result,
+    #     description="Get user's blog name from the Internet.",
+    # ),
+    Tool(
+        name="Search user's Library Document",
+        func=search_document,
+        description="优先使用 Search user's Library Document.",
+    )
+]
+
+llm = ChatOpenAI(temperature=0)
+loaded_tools = load_tools(["llm-math"], llm=llm)
+tools.extend(loaded_tools)
+
+memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
+memory.chat_memory.messages.append(HumanMessage(content="必须使用中文回复。"))
+# memory.clear()
+
+agent_chain = initialize_agent(tools, llm,
+                               agent=AgentType.CHAT_CONVERSATIONAL_REACT_DESCRIPTION,
+                               verbose=False,
+                               memory=memory)
+
+while True:
+    question = input("请输入问题：")
+    question = "必须使用中文回复:" + question
+    result = agent_chain.run(input=question)
+    print(result)
+
+

document_ai/chat.py

@@ -0,0 +1,121 @@
+import json
+from langchain.tools import BaseTool
+import proto.documents_pb2
+
+import init
+import doc_client
+import openai
+
+# class Eva
+
+
+# 连续对话
+messages = [
+    {
+        "role": "system",
+        "content": """
+回答问题使用文档，并以提问的语言和Markdown回答，并告诉来源。
+你得用“你”的身份指代用户。请辨别文档中的内容，有一些是不相干的。
+"""
+    },
+    #     {
+    #         "role": "system",
+    #         "content": f"""
+    # Context: {plain_text}
+    # """
+    #     },
+    # {
+    #         "role": "user",
+    #         "content": f"""
+    # {question}
+    # """
+    #     }
+]
+
+
+def ask_question(question):
+    messages.append({
+        "role": "user",
+        "content": f"""
+    {question}
+    """
+    })
+
+    question_vec = init.text_to_vector(question)
+    search_param = {
+        "data": [question_vec],
+        "anns_field": "vector",
+        "param": {"metric_type": "L2"},
+        "limit": 5,
+        "expr": "user_id == 2",
+        "output_fields": ["todo_id", "title", "source", "todo_description", "language", "text", "user_id"],
+    }
+    res = init.collection.search(**search_param)
+
+    plain_text = ""
+
+    for i in range(len(res[0])):
+        _chunk_id = res[0][i].id
+
+        if _chunk_id == "0":
+            continue
+
+        print("正在获取分块 " + str(_chunk_id) + " 的内容...")
+
+        try:
+            _chunk_content = doc_client.stub.GetDocumentChunk(proto.documents_pb2.GetDocumentChunkByIdRequest(
+                id=_chunk_id
+            ))
+
+            print(_chunk_content)
+
+            plain_text += ("=== \n" + f"文档 ID:{_chunk_content.document.id}\n"
+                           + f"文档内容: {_chunk_content.content}" + "===\n")
+
+        except Exception as e:
+            print(e)
+
+    messages.append({
+        "role": "system",
+        "content": f"""
+文档: {plain_text}
+"""
+    })
+
+    print("正在调用 LLM...")
+    result = openai.ChatCompletion.create(
+        messages=messages, model="gpt-3.5-turbo", temperature=0
+    )
+    res = result["choices"][0]["message"].to_dict_recursive()
+
+    # add to
+    messages.append({
+        "role": "assistant",
+        "content": res["content"]
+    })
+
+
+    # 删除多余的 system 消息
+    for i in range(len(messages)):
+        if messages[i]["role"] == "system":
+
+            if i == 0:
+                continue
+
+            messages.pop(i)
+            break
+
+    return res["content"]
+
+
+while True:
+
+    print(messages)
+
+    if len(messages) > 10:
+        messages = messages[-10:]
+        print("很抱歉，我只能记住最近 10 条上下文数据。让我们重新开始吧。")
+
+    question = input("请输入问题：")
+    resp = ask_question(question)
+    print(resp)

document_ai/search.py

@@ -1,9 +1,15 @@
 import json
+
+import openai
+from langchain.chains import LLMChain
+from langchain.prompts import PromptTemplate
+
 import proto.documents_pb2
 from langchain import text_splitter
 from langchain.chains.qa_with_sources import load_qa_with_sources_chain
 from langchain.embeddings import OpenAIEmbeddings
-from langchain.llms.openai import OpenAI
+# from langchain.llms.openai import OpenAI;
+from langchain.chat_models.openai import ChatOpenAI
 from langchain.schema.document import Document
 from pymilvus import (
     connections,
@@ -16,6 +22,8 @@ from pymilvus import (
 
 import init
 import doc_client
+import openai
+from langchain.adapters import openai as lc_openai
 
 #
 # question = """
@@ -31,15 +39,15 @@ import doc_client
 # """
 
 question = """
-为什么我会在 	WHMCS 下开发摸不着头脑
+错误 yarn 什么了遇到我
 """
-
-vec = init.text_to_vector(question)
-
-# vec = ""
 #
-# with open("../question_vec.json", "r") as f:
-#     vec = json.load(f)
+# vec = init.text_to_vector(question)
+
+vec = ""
+
+with open("../question_vec.json", "r") as f:
+    vec = json.load(f)
 
 search_param = {
     "data": [vec],
@@ -51,8 +59,13 @@ search_param = {
 }
 res = init.collection.search(**search_param)
 
-document_chunk_ids = []
-real_document = []
+# 保留 5 个
+if len(res[0]) > 5:
+    res[0] = res[0][:5]
+
+# document_chunk_ids = []
+# real_document = []
+plain_text = ""
 
 for i in range(len(res[0])):
     _chunk_id = res[0][i].id
@@ -65,23 +78,104 @@ for i in range(len(res[0])):
 
         # print(_chunk_content)
 
-        _doc_content_full = _chunk_content.content
+        # _doc_content_full = _chunk_content.content
+        # print("DOC OBJ:" + _doc_content_full)
+        plain_text += "=== \n" + _chunk_content.content + " ===\n"
 
         # real_document.append(_doc_content)
         # doc_obj = Document(page_content=_doc_content_full, metadata={"source": _chunk_content.title})
-        doc_obj = Document(page_content=_doc_content_full, metadata={"source": "chunked content"})
+        # doc_obj = Document(page_content=_doc_content_full, metadata={"source": "chunked content"})
 
-        real_document.append(doc_obj)
+        # real_document.append(doc_obj)
 
     except Exception as e:
         print(e)
 
-print(real_document)
+# print(real_document)
 
 print("正在调用 LLM...")
-chain = load_qa_with_sources_chain(OpenAI(temperature=0), chain_type="map_reduce", return_intermediate_steps=True,
-                                   verbose=True)
 
-question = "必须使用中文回复:" + question
-output = chain({"input_documents": real_document, "question": question}, return_only_outputs=False)
-print("回复：" + output["output_text"])
+# prompt_template = f"""Answer questions use the following context and reply in question spoken language and answer
+# with Markdown format, you can fix syntax errors in the context, but do not change the meaning of the context.
+# you can tell user context errors(syntax or meaning) in answer.
+# ---
+# {plain_text}
+# ---
+# Question: {question}
+# Answer:"""
+
+messages = [
+    {
+        "role": "system",
+        "content": """
+回答问题使用以下上下文，并以提问的语言和Markdown回答，并告诉来源。
+你得用“你”的身份指代用户。如果用户的问题有语法错误或者上下文的意思不对，你可以告诉用户。
+请辨别上下文中的内容，有一些是不相干的。
+"""
+    },
+    {
+        "role": "system",
+        "content": f"""
+Context: {plain_text}
+"""
+    },
+{
+        "role": "user",
+        "content": f"""
+{question}
+"""
+    }
+]
+
+result = openai.ChatCompletion.create(
+    messages=messages, model="gpt-3.5-turbo", temperature=0
+)
+res = result["choices"][0]["message"].to_dict_recursive()
+print(res)
+
+
+# prompt_template = f"""
+# ---
+# {plain_text}
+# ---
+# Question: {question}
+# Answer:"""
+#
+# print(prompt_template)
+# # PROMPT = PromptTemplate(
+# #     template=prompt_template, input_variables=["real_document", "question"]
+# # )
+#
+#
+# ChatOpenAI
+
+# llm = OpenAI(temperature=0, model_name="gpt-3.5-turbo")
+# # chain = LLMChain(llm=llm, prompt=PROMPT)
+#
+# output = llm(prompt_template)
+
+# gpt = openai.Completion.create(
+#     engine="gpt-3.5-turbo",
+#     prompt=prompt_template,
+#     max_tokens=150,
+#     temperature=0,
+#     top_p=1,
+#     frequency_penalty=0,
+#     presence_penalty=0,
+#     stop=["==="]
+#     )
+
+# output = gpt["choices"][0]["text"]
+# print(output)
+
+# output = chain({"real_document": real_document, "question": question}, return_only_outputs=True)
+
+
+# print(output)
+
+# chain = load_qa_with_sources_chain(OpenAI(temperature=0), chain_type="map_reduce", return_intermediate_steps=True,
+#                                    verbose=True)
+#
+# question = "必须使用中文回复:" + question
+# output = chain({"input_documents": real_document, "question": question}, return_only_outputs=False)
+# print("回复：" + output["output_text"])

document_ai/server.py

@@ -42,7 +42,7 @@ class AIServer(proto.document_query_pb2_grpc.DocumentQuery):
             "data": [vec],
             "anns_field": "vector",
             "param": {"metric_type": "L2"},
-            "limit": 10,
+            "limit": 5,
             "expr": "user_id == " + str(target.user_id),
             "output_fields": ["document_id", "user_id"],
         }