Upload 2 files

app.py

@@ -22,6 +22,7 @@ from langchain.embeddings.openai import OpenAIEmbeddings
 from langchain.chat_models import ChatOpenAI
 
 # LangChain
+import langchain
 from langchain.llms import HuggingFacePipeline
 from transformers import pipeline
 
@@ -45,8 +46,8 @@ import gradio as gr
 from pypdf import PdfReader
 import requests # DeepL API request
 
-# test
-import langchain # (debug=Trueにするため)
+# Mecab
+import MeCab
 
 # --------------------------------------
 #  ユーザ別セッションの変数値を記録するクラス
@@ -69,6 +70,7 @@ class SessionState:
 		self.conversation_chain		= None		# ConversationChain
 		self.query_generator			= None		# Query Refiner with Chat history
 		self.qa_chain							= None		# load_qa_chain
+		self.web_summary_chain		= None		# Summarize web search result
 		self.embedded_urls				= []
 		self.similarity_search_k	= None		# No. of similarity search documents to find.
 		self.summarization_mode		=	None		# Stuff / Map Reduce / Refine
@@ -132,6 +134,33 @@ text_splitter = JPTextSplitter(
     chunk_overlap = chunk_overlap,  # オーバーラップの最大文字数
 )
 
+# --------------------------------------
+# 文中から人名を抽出
+# --------------------------------------
+def name_detector(text: str) -> list:
+  mecab = MeCab.Tagger()
+  mecab.parse('')  # ←バグ対応
+  node = mecab.parseToNode(text).next
+  names = []
+
+  while node:
+    if node.feature.split(',')[3] == "姓":
+      if node.next and node.next.feature.split(',')[3] == "名":
+        names.append(str(node.surface) + str(node.next.surface))
+      else:
+        names.append(node.surface)
+    if node.feature.split(',')[3] == "名":
+      if node.prev and node.prev.feature.split(',')[3] == "姓":
+        pass
+      else:
+        names.append(str(node.surface))
+
+    node = node.next
+
+  names = list(set(names))
+
+  return names
+
 # --------------------------------------
 # DeepL でメモリを翻訳しトークン数を削減(OpenAIモデル利用時)
 # --------------------------------------
@@ -175,11 +204,22 @@ def deepl_memory(ss: SessionState) -> (SessionState):
 # DEEPL_API_ENDPOINT = "https://api-free.deepl.com/v2/translate"
 # DEEPL_API_KEY = os.getenv("DEEPL_API_KEY")
 
-def web_search(query, current_model) -> str:
-  search =  DuckDuckGoSearchRun()
+def web_search(ss: SessionState, query) -> (SessionState, str):
+  search =  DuckDuckGoSearchRun(verbose=True)
   web_result = search(query)
 
-  if current_model == "gpt-3.5-turbo":
+  # 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 = {
         "auth_key": DEEPL_API_KEY,
@@ -193,19 +233,28 @@ def web_search(query, current_model) -> str:
     response = request.json()
 
     query = response["translations"][0]["text"]
-    web_result =  response["translations"][1]["text"]
+    web_result = response["translations"][1]["text"]
+    web_result = ss.web_summary_chain({'query': query, 'context': web_result})['text']
+
+  if names != "":
+    web_query = f"""
+    {query}
+    Use the following information as a reference to answer the question above in Japanese. When translating names of Japanese people, refer to Japanese Names as a translation guide.
+    Reference: {web_result}
+    Japanese Names: {names}
+    """.strip()
+  else:
+    web_query = query + "\nUse the following information as a reference to answer the question above in the Japanese.\n===\nReference: " + web_result + "\n==="
 
-  web_query = query + "\nUse the following information as a reference to answer the question above in the Japanese.\n===\nReference: " + web_result + "\n==="
 
-  return web_query
+  return ss, web_query
 
 # --------------------------------------
 # LangChain カスタムプロンプト各種
 #   llama tokenizer
-#   https://belladoreai.github.io/llama-tokenizer-js/example-demo/build/
-
+#     https://belladoreai.github.io/llama-tokenizer-js/example-demo/build/
 #   OpenAI tokenizer
-#   https://platform.openai.com/tokenizer
+#     https://platform.openai.com/tokenizer
 # --------------------------------------
 
 # --------------------------------------
@@ -214,19 +263,18 @@ def web_search(query, current_model) -> str:
 
 # Tokens: OpenAI 104/ Llama 105 <- In Japanese: Tokens: OpenAI 191/ Llama 162
 sys_chat_message = """
-The following is a conversation between an AI concierge and a customer.
-The AI understands what the customer wants to know from the conversation history and the latest question,
-and gives many specific details in Japanese. If the AI does not know the answer to a question, it does not
-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".
+You are an outstanding AI concierge. You understand your customers' needs from their questions and 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". Ignore Conversation History.
 """.replace("\n", "")
 
 chat_common_format = """
 ===
 Question: {query}
-
-Conversation History:
+===
+Conversation History(Ignore):
 {chat_history}
-
+===
 日本語の回答: """
 
 chat_template_std = f"{sys_chat_message}{chat_common_format}"
@@ -238,21 +286,23 @@ chat_template_llama2 = f"<s>[INST] <<SYS>>{sys_chat_message}<</SYS>>{chat_common
 # 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.
-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".
+You understand what the customer wants to know from 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".
+Ignore Conversation History.
 """.replace("\n", "")
 
 qa_common_format = """
 ===
 Question: {query}
 References: {context}
-Conversation History:
+===
+Conversation History(Ignore):
 {chat_history}
-
+===
 日本語の回答: """
 
+
 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]"
 
@@ -262,8 +312,8 @@ qa_template_llama2 = f"<s>[INST] <<SYS>>{sys_qa_message}<</SYS>>{qa_common_forma
 # 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 in Japanese by filling in the missing subject, verb, objects,
-complements, and other necessary information to get a better search result.
+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.
 """.replace("\n", "")
 
 query_generator_common_format = """
@@ -272,7 +322,7 @@ query_generator_common_format = """
 {chat_history}
 
 [Additional Question] {query}
-明確な質問文: """
+明確な日本語の質問文: """
 
 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]"
@@ -287,8 +337,8 @@ and complement.
 
 question_prompt_common_format = """
 ===
-[references] {context}
 [Question] {query}
+[references] {context}
 [Summary] """
 
 question_prompt_template_std = f"{question_prompt_message}{question_prompt_common_format}"
@@ -305,17 +355,14 @@ If you do not know the answer, do not make up an answer and reply,
 
 combine_prompt_common_format = """
 ===
-Question:
-{query}
-===
+Question: {query}
 Reference: {summaries}
-===
 日本語の回答: """
 
+
 combine_prompt_template_std = f"{combine_prompt_message}{combine_prompt_common_format}"
 combine_prompt_template_llama2 = f"<s>[INST] <<SYS>>{combine_prompt_message}<</SYS>>{combine_prompt_common_format}[/INST]"
 
-
 # --------------------------------------
 # ConversationSummaryBufferMemoryの要約プロンプト
 # ソース　→　https://github.com/langchain-ai/langchain/blob/894c272a562471aadc1eb48e4a2992923533dea0/langchain/memory/prompt.py#L26-L49
@@ -508,6 +555,10 @@ def set_chains(ss: SessionState, summarization_mode) -> SessionState:
   # --------------------------------------
   # Conversation/QAチェーンの設定
   # --------------------------------------
+  if ss.query_generator is None:
+    query_generator_prompt  = PromptTemplate(template=query_generator_template, input_variables = ["chat_history", "query"])
+    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)
     ss.conversation_chain = ConversationChain(
@@ -525,13 +576,14 @@ def set_chains(ss: SessionState, summarization_mode) -> SessionState:
       ss.qa_chain             = load_qa_chain(ss.llm, chain_type="stuff", memory=ss.memory, prompt=qa_prompt)
 
     elif summarization_mode == "map_reduce":
-      query_generator_prompt  = PromptTemplate(template=query_generator_template, input_variables = ["chat_history", "query"])
-      ss.query_generator      = LLMChain(llm=ss.llm, prompt=query_generator_prompt)
-
       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)
 
+  if ss.web_summary_chain is None:
+    question_prompt           = PromptTemplate(template=question_template, input_variables=["context", "query"])
+    ss.web_summary_chain      = LLMChain(llm=ss.llm, prompt=question_prompt, verbose=True)
+
   return ss
 
 def initialize_db(ss: SessionState) -> SessionState:
@@ -761,16 +813,16 @@ def bot(ss: SessionState, query, qa_flag, web_flag, summarization_mode) -> (Sess
   # QA Model
   if qa_flag is True and ss.embeddings is not None and ss.db is not None:
     if web_flag:
-      web_query = web_search(query, ss.current_model)
+      ss, web_query = web_search(ss, query)
       ss = qa_predict(ss, web_query)
       ss.memory.chat_memory.messages[-2].content = query
     else:
-      ss = qa_predict(ss, query)      # LLMで回答を生成
+      ss = qa_predict(ss, query)
 
   # Chat Model
   else:
     if web_flag:
-      web_query = web_search(query, ss.current_model)
+      ss, web_query = web_search(ss, query)
       ss = chat_predict(ss, web_query)
       ss.memory.chat_memory.messages[-2].content = query
     else:
@@ -788,6 +840,8 @@ def chat_predict(ss: SessionState, query) -> SessionState:
 
 def qa_predict(ss: SessionState, query) -> SessionState:
 
+  original_query = query
+
   # Rinnaモデル向けの設定（クエリの改行コード修正）
   if ss.current_model == "rinna/bilingual-gpt-neox-4b-instruction-sft":
     query = query.strip().replace("\n", "<NL>")
@@ -829,7 +883,7 @@ def qa_predict(ss: SessionState, query) -> SessionState:
     response += "参考文献の抽出には成功していますので、言語モデルを変えてお試しください。"
 
   # ユーザーメッセージと AI メッセージの追加
-  ss.memory.chat_memory.add_user_message(query.replace("<NL>", "\n"))
+  ss.memory.chat_memory.add_user_message(original_query.replace("<NL>", "\n"))
   ss.memory.chat_memory.add_ai_message(response)
   ss.dialogue[-1] = (ss.dialogue[-1][0], response)  # 会話履歴
   return ss
@@ -1028,4 +1082,5 @@ with gr.Blocks() as demo:
 
 if __name__ == "__main__":
     demo.queue(concurrency_count=5)
-    demo.launch(debug=True)
+    demo.launch(debug=True,)
+

requirements.txt

@@ -21,4 +21,6 @@ numpy==1.23.5
 pandas==1.5.3
 chromedriver-autoinstaller
 chromedriver-binary
-duckduckgo-search==3.8.5
+duckduckgo-search==3.8.5
+mecab-python3==1.0.6
+unidic-lite==1.0.8