Upload app.py

app.py

@@ -11,6 +11,7 @@ import os
 import time
 import gc # メモリ解放
 import re # 正規表現で文章をクリーンアップ
+import regex # 漢字抽出で利用
 
 # HuggingFace
 import torch
@@ -115,6 +116,55 @@ class SessionState:
 
 		self.cache_clear()
 
+# --------------------------------------
+# メモリを使用しない ConversationChainを自作
+# --------------------------------------
+from typing import Dict, List
+
+from langchain.chains.conversation.prompt import PROMPT
+from langchain.chains.llm import LLMChain
+from langchain.pydantic_v1 import Extra, Field, root_validator
+from langchain.schema import BasePromptTemplate
+
+class ConversationChain(LLMChain):
+    """Chain to have a conversation without loading context from memory.
+
+    Example:
+        .. code-block:: python
+
+            from langchain import ConversationChainWithoutMemory, OpenAI
+
+            conversation = ConversationChainWithoutMemory(llm=OpenAI())
+    """
+
+    prompt: BasePromptTemplate = PROMPT
+    """Default conversation prompt to use."""
+
+    input_key: str = "input"  #: :meta private:
+    output_key: str = "response"  #: :meta private:
+
+    class Config:
+        """Configuration for this pydantic object."""
+
+        extra = Extra.forbid
+        arbitrary_types_allowed = True
+
+    @property
+    def input_keys(self) -> List[str]:
+        """Use this since so some prompt vars come from history."""
+        return [self.input_key]
+
+    @root_validator()
+    def validate_prompt_input_variables(cls, values: Dict) -> Dict:
+        """Validate that prompt input variables are consistent without memory."""
+        input_key = values["input_key"]
+        prompt_variables = values["prompt"].input_variables
+        if input_key not in prompt_variables:
+            raise ValueError(
+                f"The prompt expects {prompt_variables}, but {input_key} is not found."
+            )
+        return values
+
 # --------------------------------------
 # 自作TextSplitter（テキストをLLMのトークン数内に分割）
 # （参考）https://www.sato-susumu.com/entry/2023/04/30/131338
@@ -157,10 +207,21 @@ def name_detector(text: str) -> list:
 
     node = node.next
 
-  names = list(set(names))
+  # ユニークな値を抽出し、その後漢字を含む値のみとする
+  names = filter_kanji(list(set(names)))
 
   return names
 
+# --------------------------------------
+# リストから漢字を含む値だけを抽出する
+# --------------------------------------
+def filter_kanji(lst) -> list:
+  def contains_kanji(s):
+    p = regex.compile(r'\p{Script=Han}+')
+    return bool(p.search(s))
+
+  return [item for item in lst if contains_kanji(item)]
+
 # --------------------------------------
 # DeepL でメモリを翻訳しトークン数を削減(OpenAIモデル利用時)
 # --------------------------------------
@@ -207,21 +268,12 @@ def deepl_memory(ss: SessionState) -> (SessionState):
 def web_search(ss: SessionState, query) -> (SessionState, str):
 
   search =  DuckDuckGoSearchRun(verbose=True)
+  names = []
+  names.extend(name_detector(query))
 
   for i in range(3):
     web_result = search(query)
 
-    # 人名の抽出
-    names = []
-    names.extend(name_detector(query))
-    names.extend(name_detector(web_result))
-    if len(names)==0:
-      names = ""
-    elif len(names)==1:
-      names = names[0]
-    else:
-      names = ", ".join(names)
-
     if ss.current_model == "gpt-3.5-turbo":
       text = [query, web_result]
       params = {
@@ -235,21 +287,33 @@ def web_search(ss: SessionState, query) -> (SessionState, str):
       request = requests.post(DEEPL_API_ENDPOINT, data=params)
       response = request.json()
 
-      query = response["translations"][0]["text"]
-      web_result = response["translations"][1]["text"]
-      web_result = ss.web_summary_chain({'query': query, 'context': web_result})['text']
-      if web_result != "NO INFO":
+      query_eng = response["translations"][0]["text"]
+      web_result_eng = response["translations"][1]["text"]
+      web_result_eng = ss.web_summary_chain({'query': query_eng, 'context': web_result_eng})['text']
+      if "$$NO INFO$$" in web_result_eng:
+        web_result_eng = ss.web_summary_chain({'query': query_eng, 'context': web_result_eng})['text']
+      if "$$NO INFO$$" not in web_result_eng:
         break
 
+  # 検索結果から人名を抽出し、テキスト化
+  names.extend(name_detector(web_result))
+  if len(names)==0:
+    names = ""
+  elif len(names)==1:
+    names = names[0]
+  else:
+    names = ", ".join(names)
+
+  # Web検索結果を含むQueryを渡す。
   if names != "":
     web_query = f"""
-    {query}
-    Use the following Suggested Answer Source as a reliable reference to answer the question above in Japanese. When translating names of people, refer to Names as a translation guide.
-    Suggested Answer Source: {web_result}
+    {query_eng}
+    Use the following Suggested Answer as a reference to answer the question above in Japanese. When translating names of people, refer to Names as a translation guide.
+    Suggested Answer: {web_result_eng}
     Names: {names}
     """.strip()
   else:
-    web_query = query + "\nUse the following Suggested Answer Source as a reliable reference to answer the question above in the Japanese.\n===\nSuggested Answer Source: " + web_result + "\n"
+    web_query = query_eng + "\nUse the following Suggested Answer as a reference to answer the question above in the Japanese.\n===\nSuggested Answer: " + web_result_eng + "\n"
 
 
   return ss, web_query
@@ -265,29 +329,19 @@ def web_search(ss: SessionState, query) -> (SessionState, str):
 # --------------------------------------
 # Conversation Chain Template
 # --------------------------------------
-
 # Tokens: OpenAI 104/ Llama 105 <- In Japanese: Tokens: OpenAI 191/ Llama 162
-# sys_chat_message = """
-# You are an outstanding AI concierge. Understand the intent of the customer's questions based on
-# the conversation history. Then, answer them with many specific and detailed information in Japanese.
-# If you do not know the answer to a question, do make up an answer and says
-# "誠に申し訳ございませんが、その点についてはわかりかねます".
-# """.replace("\n", "")
 
 sys_chat_message = """
-You are an outstanding AI concierge.
-1) Understand the intent of the customer's questions based on the conversation history.
-2) Then, by using references if available, answer the question with many specific and detailed information in Japanese.
-3) If the reference does not provide answer to the question at all, and you do not know the answer, do make up an answer and says "誠に申し訳ございませんが、その点についてはわかりかねます".
-""".strip()
+You are an AI concierge who carefully answers questions from customers based on references.
+You understand what the customer wants to know, and give many specific details in Japanese
+using sentences extracted from the following references when available. If you do not know
+the answer, do not make up an answer and reply, "誠に申し訳ございませんが、その点についてはわかりかねます".
+""".replace("\n", "")
 
 chat_common_format = """
 ===
 Question: {query}
-===
-Conversation History:
-{chat_history}
-===
+
 日本語の回答: """
 
 chat_template_std = f"{sys_chat_message}{chat_common_format}"
@@ -297,35 +351,46 @@ chat_template_llama2 = f"<s>[INST] <<SYS>>{sys_chat_message}<</SYS>>{chat_common
 # QA Chain Template (Stuff)
 # --------------------------------------
 # Tokens: OpenAI 113/ Llama 111 <- In Japanese: Tokens: OpenAI 256/ Llama 225
-sys_qa_message = """
-You are an AI concierge who carefully answers questions from customers based on references.
-Understand the intent of the customer's questions based on the conversation history. Then, give
-a specific answer in Japanese using sentences extracted from the following references. If you do
-not know the answer, do not make up an answer and reply, "誠に申し訳ございませんが、その点についてはわかりかねます".
-""".replace("\n", "")
+# sys_qa_message = """
+# You are an AI concierge who carefully answers questions from customers based on references.
+# You understand what the customer wants to know from the Conversation History and Question,
+# and give a specific answer in Japanese using sentences extracted from the following references.
+# If you do not know the answer, do not make up an answer and reply,
+# "誠に申し訳ございませんが、その点についてはわかりかねます".
+# """.replace("\n", "")
+
+# qa_common_format = """
+# ===
+# Question: {query}
+# References: {context}
+# ===
+# Conversation History:
+# {chat_history}
+# ===
+# 日本語の回答: """
 
 qa_common_format = """
 ===
 Question: {query}
 References: {context}
-===
-Conversation History:
-{chat_history}
-===
+
 日本語の回答: """
 
+qa_template_std = f"{sys_chat_message}{qa_common_format}"
+qa_template_llama2 = f"<s>[INST] <<SYS>>{sys_chat_message}<</SYS>>{qa_common_format}[/INST]"
 
-qa_template_std = f"{sys_qa_message}{qa_common_format}"
-qa_template_llama2 = f"<s>[INST] <<SYS>>{sys_qa_message}<</SYS>>{qa_common_format}[/INST]"
+# qa_template_std = f"{sys_qa_message}{qa_common_format}"
+# qa_template_llama2 = f"<s>[INST] <<SYS>>{sys_qa_message}<</SYS>>{qa_common_format}[/INST]"
 
 # --------------------------------------
 # QA Chain Template (Map Reduce)
 # --------------------------------------
 # 1. 会話履歴と最新の質問から、質問文を生成するchain のプロンプト
 query_generator_message = """
-Referring to the "Conversation History", reformat the user's "Additional Question"
-to a specific question by filling in the missing subject, verb, objects, complements,
-and other necessary information to get a better search result. Answer in Japanese.
+Referring to the "Conversation History", especially to the most recent conversation,
+reformat the user's "Additional Question" into a specific question in Japanese by
+filling in the missing subject, verb, objects, complements,and other necessary
+information to get a better search result. Answer in 日本語(Japanese).
 """.replace("\n", "")
 
 query_generator_common_format = """
@@ -334,30 +399,25 @@ query_generator_common_format = """
 {chat_history}
 
 [Additional Question] {query}
-明確な日本語の質問文: """
+明確な質問文: """
 
 query_generator_template_std = f"{query_generator_message}{query_generator_common_format}"
 query_generator_template_llama2 = f"<s>[INST] <<SYS>>{query_generator_message}<</SYS>>{query_generator_common_format}[/INST]"
 
 
 # 2. 生成された質問文を用いて、参考文献を要約するchain のプロンプト
-# question_prompt_message = """
-# From the following references, extract key information relevant to the question
-# and summarize it in a natural English sentence with clear subject, verb, object,
-# and complement. If there is no information in the reference that answers the question,
-# do not summarize and simply answer "NO INFO"
-# """.replace("\n", "")
 
 question_prompt_message = """
-1. Determine if any of the following references provide information that answers the Question, and if there is no information, answer "NO INFO" and stop.
-2. From the following references, extract key information relevant to the question and summarize it in a natural English sentence with clear subject, verb, object,　and complement.
-""".strip()
+From the following references, extract key information relevant to the question
+and summarize it in a natural English sentence with clear subject, verb, object,
+and complement.
+""".replace("\n", "")
 
 question_prompt_common_format = """
 ===
 [Question] {query}
-[references] {context}
-[Answer]"""
+[References] {context}
+[Key Information] """
 
 question_prompt_template_std = f"{question_prompt_message}{question_prompt_common_format}"
 question_prompt_template_llama2 = f"<s>[INST] <<SYS>>{question_prompt_message}<</SYS>>{question_prompt_common_format}[/INST]"
@@ -578,11 +638,12 @@ def set_chains(ss: SessionState, summarization_mode) -> SessionState:
     ss.query_generator      = LLMChain(llm=ss.llm, prompt=query_generator_prompt, verbose=True)
 
   if ss.conversation_chain is None:
-    chat_prompt = PromptTemplate(input_variables=['query', 'chat_history'], template=chat_template)
+    # chat_prompt = PromptTemplate(input_variables=['query', 'chat_history'], template=chat_template)
+    chat_prompt = PromptTemplate(input_variables=['query'], template=chat_template)
     ss.conversation_chain = ConversationChain(
       llm         = ss.llm,
       prompt      = chat_prompt,
-      memory      = ss.memory,
+      # memory      = ss.memory,
       input_key   = "query",
       output_key  = "output_text",
       verbose     = True,
@@ -590,13 +651,16 @@ def set_chains(ss: SessionState, summarization_mode) -> SessionState:
 
   if ss.qa_chain is None:
     if summarization_mode == "stuff":
-      qa_prompt               = PromptTemplate(input_variables=['context', 'query', 'chat_history'], template=qa_template)
-      ss.qa_chain             = load_qa_chain(ss.llm, chain_type="stuff", memory=ss.memory, prompt=qa_prompt)
+      # qa_prompt               = PromptTemplate(input_variables=['context', 'query', 'chat_history'], template=qa_template)
+      qa_prompt               = PromptTemplate(input_variables=['context', 'query'], template=qa_template)
+      # ss.qa_chain             = load_qa_chain(ss.llm, chain_type="stuff", memory=ss.memory, prompt=qa_prompt)
+      ss.qa_chain             = load_qa_chain(ss.llm, chain_type="stuff", prompt=qa_prompt, verbose=True)
 
     elif summarization_mode == "map_reduce":
       question_prompt         = PromptTemplate(template=question_template, input_variables=["context", "query"])
       combine_prompt          = PromptTemplate(template=combine_template, input_variables=["summaries", "query"])
-      ss.qa_chain             = load_qa_chain(ss.llm, chain_type="map_reduce", return_map_steps=True, memory=ss.memory, question_prompt=question_prompt, combine_prompt=combine_prompt)
+      # ss.qa_chain             = load_qa_chain(ss.llm, chain_type="map_reduce", return_map_steps=True, memory=ss.memory, question_prompt=question_prompt, combine_prompt=combine_prompt)
+      ss.qa_chain             = load_qa_chain(ss.llm, chain_type="map_reduce", return_map_steps=True, question_prompt=question_prompt, combine_prompt=combine_prompt, verbose=True)
 
   if ss.web_summary_chain is None:
     question_prompt           = PromptTemplate(template=question_template, input_variables=["context", "query"])
@@ -853,6 +917,8 @@ def bot(ss: SessionState, query, qa_flag, web_flag, summarization_mode) -> (Sess
 
 def chat_predict(ss: SessionState, query) -> SessionState:
   response = ss.conversation_chain.predict(query=query)
+  ss.memory.chat_memory.add_user_message(query)
+  ss.memory.chat_memory.add_ai_message(response)
   ss.dialogue[-1] = (ss.dialogue[-1][0], response)
   return ss
 
@@ -890,10 +956,12 @@ def qa_predict(ss: SessionState, query) -> SessionState:
       if result["output_text"] != "":
         response = result["output_text"] + sources
         ss.dialogue[-1] = (ss.dialogue[-1][0], response)
+        ss.memory.chat_memory.add_user_message(original_query)
+        ss.memory.chat_memory.add_ai_message(response)
         return ss
-      else:
+      # else:
         # 空欄の場合は直近の履歴を削除してやり直し
-        ss.memory.chat_memory.messages = ss.memory.chat_memory.messages[:-2]
+        # ss.memory.chat_memory.messages = ss.memory.chat_memory.messages[:-2]
 
   # 3回の試行後も空欄の場合
   response = "3回試行しましたが、情報製生成できませんでした。"