add index

app.py

@@ -3,30 +3,33 @@ import os
 import json
 import requests
 from langchain import FAISS
-from langchain.prompts import PromptTemplate
-from langchain.embeddings import CohereEmbeddings, HuggingFaceInstructEmbeddings
+from langchain.embeddings import CohereEmbeddings
+from langchain import VectorDBQA
+from langchain.chat_models import ChatOpenAI
+from prompts import MyTemplate
+from langchain.prompts.chat import (
+    ChatPromptTemplate,
+    SystemMessagePromptTemplate,
+    HumanMessagePromptTemplate,
+)
 
 # Streaming endpoint
 API_URL = "https://api.openai.com/v1/chat/completions"  # os.getenv("API_URL") + "/generate_stream"
-embedding_key = '5IRbILAbjTI0VcqTsktBfKsr13Lych9iBAFbLpkj'
-faiss_store = './output/数字经济'
-
-
-# load prompt template
-with open("prompts/combine_prompt.txt", "r") as f:
-    template = f.read()
-
-with open("prompts/combine_prompt_hist.txt", "r") as f:
-    template_hist = f.read()
-
-with open("prompts/question_prompt.txt", "r") as f:
-    template_quest = f.read()
-
-with open("prompts/chat_combine_prompt.txt", "r") as f:
-    chat_combine_template = f.read()
-
-with open("prompts/chat_reduce_prompt.txt", "r") as f:
-    chat_reduce_template = f.read()
+embeddings_key = '5IRbILAbjTI0VcqTsktBfKsr13Lych9iBAFbLpkj'
+faiss_store = './indexer'
+
+def gen_conversation(conversations):
+    messages = []
+    for data in conversations:
+        temp1 = {}
+        temp1["role"] = "user"
+        temp1["content"] = data[0]
+        temp2 = {}
+        temp2["role"] = "assistant"
+        temp2["content"] = data[1]
+        messages.append(temp1)
+        messages.append(temp2)
+    return messages
 
 
 def predict(inputs, top_p, temperature, openai_api_key, enable_index,
@@ -37,71 +40,67 @@ def predict(inputs, top_p, temperature, openai_api_key, enable_index,
     }
 
     print(f"chat_counter - {chat_counter}")
-    # 如果有历史对话，把对话拼接进入上下文
+    #Debugging
+    if enable_index:
+        # Faiss 检索最近的embedding
+        docsearch = FAISS.load_local(faiss_store, CohereEmbeddings(cohere_api_key=embeddings_key))
+        llm = ChatOpenAI(openai_api_key=openai_api_key)
+        messages_combine = [
+            SystemMessagePromptTemplate.from_template(MyTemplate['chat_combine_template']),
+            HumanMessagePromptTemplate.from_template("{question}")
+        ]
+        p_chat_combine = ChatPromptTemplate.from_messages(messages_combine)
+        messages_reduce = [
+            SystemMessagePromptTemplate.from_template(MyTemplate['chat_reduce_template']),
+            HumanMessagePromptTemplate.from_template("{question}")
+        ]
+        p_chat_reduce = ChatPromptTemplate.from_messages(messages_reduce)
+        chain = VectorDBQA.from_chain_type(llm=llm, chain_type="map_reduce", vectorstore=docsearch,
+                                           k=4,
+                                           chain_type_kwargs={"question_prompt": p_chat_reduce,
+                                                              "combine_prompt": p_chat_combine}
+                                           )
+        result = chain({"query": inputs})
+        print(result)
+
     if chat_counter == 0:
-        payload = {
-            "model": "gpt-3.5-turbo",
-            "messages": [{"role": "user", "content": f"{inputs}"}],
-            "temperature": 1.0,
-            "top_p": 1.0,
-            "n": 1,
-            "stream": True,
-            "presence_penalty": 0,
-            "frequency_penalty": 0,
-        }
-    else:
-        messages = []
-        if enable_index:
-            pass
-            # history = json.loads(history)
-            # template_temp = template_hist.replace("{historyquestion}", history[0]).replace("{historyanswer}",
-            #                                                                                history[1])
-            # c_prompt = PromptTemplate(input_variables=["summaries", "question"], template=template_temp,
-            #                           template_format="jinja2")
-        else:
-            for data in chatbot:
-                temp1 = {}
-                temp1["role"] = "user"
-                temp1["content"] = data[0]
-                temp2 = {}
-                temp2["role"] = "assistant"
-                temp2["content"] = data[1]
-                messages.append(temp1)
-                messages.append(temp2)
+        messages = [{"role": "user", "content": f"{inputs}"}]
 
-        # Faiss 检索最近的embedding
-        if enable_index:
-            docsearch = FAISS.load_local(faiss_store, CohereEmbeddings(cohere_api_key=embedding_key))
-        else:
-            temp3 = {}
-            temp3["role"] = "user"
-            temp3["content"] = inputs
-            messages.append(temp3)
-
-        # messages
-        payload = {
-            "model": "gpt-3.5-turbo",
-            "messages": messages,  # [{"role": "user", "content": f"{inputs}"}],
-            "temperature": temperature,  # 1.0,
-            "top_p": top_p,  # 1.0,
-            "n": 1,
-            "stream": True,
-            "presence_penalty": 0,
-            "frequency_penalty": 0,
-        }
+    else:
+        # 如果有历史对话，把对话拼接进入上下文
+        messages = gen_conversation(chatbot)
+
+        temp3 = {}
+        temp3["role"] = "user"
+        temp3["content"] = inputs
+        messages.append(temp3)
+
+    # messages
+    payload = {
+        "model": "gpt-3.5-turbo",
+        "messages": messages,  # [{"role": "user", "content": f"{inputs}"}],
+        "temperature": temperature,  # 1.0,
+        "top_p": top_p,  # 1.0,
+        "n": 1,
+        "stream": True,
+        "presence_penalty": 0,
+        "frequency_penalty": 0,
+    }
 
     chat_counter += 1
 
-    # list of user input
+    # History: Original Input and Output
     history.append(inputs)
     print(f"payload is - {payload}")
+    #上一轮回复的[[user, AI]]
     print(f'chatbot - {chatbot}')
-    print(f'chatbot - {chatbot}')
-
+    print(f'Histroy - {history}')
+    # 请求OpenAI
     response = requests.post(API_URL, headers=headers, json=payload, stream=True)
     token_counter = 0
     partial_words = ""
 
+    # 逐字返回
     counter = 0
     for chunk in response.iter_lines():
         if counter == 0:
@@ -111,9 +110,11 @@ def predict(inputs, top_p, temperature, openai_api_key, enable_index,
         # check whether each line is non-empty
         if chunk:
             # decode each line as response data is in bytes
-            if len(json.loads(chunk.decode()[6:])['choices'][0]["delta"]) == 0:
+            delta = json.loads(chunk.decode()[6:])['choices'][0]["delta"]
+            if len(delta) == 0:
                 break
-            partial_words = partial_words + json.loads(chunk.decode()[6:])['choices'][0]["delta"]["content"]
+            partial_words += delta["content"]
+            # Keep Updating history
             if token_counter == 0:
                 history.append(" " + partial_words)
             else:
@@ -136,7 +137,7 @@ with gr.Blocks(css="""#col_container {width: 1000px; margin-left: auto; margin-r
         inputs = gr.Textbox(placeholder="您有什么问题可以问我", label="输入数字经济，两会，硅谷银行相关的提问")
         state = gr.State([])
 
-        clear = gr.Button("Clear")
+        clear = gr.Button("Clear Conversation")
         run = gr.Button("Run")
 
         # inputs, top_p, temperature, top_k, repetition_penalty
@@ -147,15 +148,17 @@ with gr.Blocks(css="""#col_container {width: 1000px; margin-left: auto; margin-r
                                     label="Temperature", )
             # top_k = gr.Slider( minimum=1, maximum=50, value=4, step=1, interactive=True, label="Top-k",)
             # repetition_penalty = gr.Slider( minimum=0.1, maximum=3.0, value=1.03, step=0.01, interactive=True, label="Repetition Penalty", )
-            chat_counter = gr.Number(value=0, visible=False, precision=0)
+            chat_counter = gr.Number(value=0, precision=0)
             enable_index = gr.Checkbox(label='是', info='是否使用研报等金融数据')
+            # 后续考虑加入搜索结果
+            enable_search = gr.Checkbox(label='是', info='是否使用搜索结果')
 
     inputs.submit(predict, [inputs, top_p, temperature, openai_api_key, enable_index, chat_counter, chatbot, state],
                   [chatbot, state, chat_counter], )
     run.click(predict, [inputs, top_p, temperature, openai_api_key, enable_index, chat_counter, chatbot, state],
               [chatbot, state, chat_counter], )
 
-    # 每次对话结束都重置对话框
+    # 每次对话结束都重置对话
     clear.click(reset_textbox, [], [inputs], queue=False)
     inputs.submit(reset_textbox, [], [inputs])
 

indexer/index.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e6006bd06e30b017ce77e5859f1c7b0abcad6c69ac81c9067e1adeb448ac273
+size 17619

prompts/__init__.py

@@ -0,0 +1,22 @@
+# -*-coding:utf-8 -*-
+
+# load prompt template
+with open("prompts/combine_prompt.txt", "r") as f:
+    template = f.read()
+
+with open("prompts/combine_prompt_hist.txt", "r") as f:
+    template_hist = f.read()
+
+with open("prompts/chat_combine_prompt.txt", "r") as f:
+    chat_combine_template = f.read()
+
+with open("prompts/chat_reduce_prompt.txt", "r") as f:
+    chat_reduce_template = f.read()
+
+
+MyTemplate ={
+    'chat_reduce_template': chat_reduce_template,
+    'chat_combine_template': chat_combine_template,
+    'template_hist': template_hist,
+    'template':template
+}

prompts/chat_reduce_prompt.txt

@@ -1,3 +1,3 @@
 Use the following portion of a long document to see if any of the text is relevant to answer the question.
 {context}
-Provide all relevant text to the question verbatim. Summarize if needed. If nothing relevant return "-".
+Provide all relevant text to the question verbatim. Summarize if needed, Answer in Chinese. If nothing relevant return "-".