[Paddle-pipelines] Add ReAct zh examples (#6095)

* Add ReAct zh examples

* refactor chatglm

* Add chatgpt agents example

* Add Chinese prompts

* Fix agents dead loop bug

* Add semantic search for ReAct

* Update ReAct_example_cn

* Add ernie_bot invocation

* Add unitests

Author: w5688414

pipelines/examples/agents/ReAct_example.py

@@ -77,28 +77,31 @@
 Final Answer: Gainsville, Florida
 ##
 Question: {query}
-Thought:
+Thought:{transcript}
 """
 
 # yapf: disable
 parser = argparse.ArgumentParser()
 parser.add_argument("--search_api_key", default=None, type=str, help="The SerpAPI key.")
+parser.add_argument('--llm_name', choices=['THUDM/chatglm-6b', "THUDM/chatglm-6b-v1.1", "gpt-3.5-turbo", "gpt-4"], default="THUDM/chatglm-6b-v1.1", help="The chatbot models ")
+parser.add_argument("--api_key", default=None, type=str, help="The API Key.")
 args = parser.parse_args()
 # yapf: enable
 
 
 def search_and_action_example():
     pn = PromptNode(
-        "THUDM/chatglm-6b",
-        max_length=512,
+        args.llm_name,
+        max_length=256,
+        api_key=args.api_key,
         default_prompt_template="question-answering-with-document-scores",
     )
 
     # https://serpapi.com/dashboard
     web_retriever = WebRetriever(api_key=args.search_api_key, top_search_results=2)
     pipeline = WebQAPipeline(retriever=web_retriever, prompt_node=pn)
 
-    prompt_node = PromptNode("THUDM/chatglm-6b", max_length=512, stop_words=["Observation:"])
+    prompt_node = PromptNode(args.llm_name, api_key=args.api_key, max_length=512, stop_words=["Observation:"])
 
     web_qa_tool = Tool(
         name="Search",

pipelines/examples/agents/ReAct_example_cn.py

@@ -0,0 +1,218 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import glob
+import os
+
+from pipelines.agents import Agent, Tool
+from pipelines.agents.base import ToolsManager
+from pipelines.document_stores import FAISSDocumentStore
+from pipelines.nodes import (
+    CharacterTextSplitter,
+    DensePassageRetriever,
+    DocxToTextConverter,
+    FileTypeClassifier,
+    PDFToTextConverter,
+    PromptNode,
+    TextConverter,
+    WebRetriever,
+)
+from pipelines.nodes.prompt.prompt_template import PromptTemplate
+from pipelines.pipelines import Pipeline, WebQAPipeline
+from pipelines.utils import fetch_archive_from_http
+
+few_shot_prompt = """
+你是一个乐于助人、知识渊博的人工智能助手。为了实现正确回答复杂问题的目标，您可以使用以下工具:
+搜索: 当你需要用谷歌搜索问题时很有用。你应该问一些有针对性的问题，例如，谁是安东尼·迪雷尔的兄弟？
+要回答问题，你需要经历多个步骤，包括逐步思考和选择合适的工具及其输入；工具将以观察作为回应。当您准备好接受最终答案时，回答"最终答案":
+示例:
+##
+问题: 哈利波特的作者是谁？
+思考: 让我们一步一步地思考。要回答这个问题，我们首先需要了解哈利波特是什么。
+工具: 搜索
+工具输入: 哈利波特是什么？
+观察: 哈利波特是一系列非常受欢迎的魔幻小说，以及后来的电影和衍生作品。
+思考: 我们了解到哈利波特是一系列魔幻小说。现在我们需要找到这些小说的作者是谁。
+工具: 搜索
+工具输入: 哈利波特的作者是谁？
+观察: 哈利波特系列的作者是J.K.罗琳（J.K. Rowling）。
+思考: 根据搜索结果，哈利波特系列的作者是J.K.罗琳。所以最终答案是J.K.罗琳。
+最终答案: J.K.罗琳
+##
+问题: {query}
+思考:{transcript}
+"""
+
+# yapf: disable
+parser = argparse.ArgumentParser()
+parser.add_argument('--device', choices=['cpu', 'gpu'], default="gpu", help="Select which device to run dense_qa system, defaults to gpu.")
+parser.add_argument("--index_name", default='dureader_index', type=str, help="The ann index name of ANN.")
+parser.add_argument("--search_engine", choices=['faiss', 'milvus'], default="faiss", help="The type of ANN search engine.")
+parser.add_argument("--retriever", choices=['dense', 'SerpAPI'], default="dense", help="The type of Retriever.")
+parser.add_argument("--max_seq_len_query", default=64, type=int, help="The maximum total length of query after tokenization.")
+parser.add_argument("--max_seq_len_passage", default=256, type=int, help="The maximum total length of passage after tokenization.")
+parser.add_argument("--retriever_batch_size", default=16, type=int, help="The batch size of retriever to extract passage embedding for building ANN index.")
+parser.add_argument("--query_embedding_model", default="rocketqa-zh-base-query-encoder", type=str, help="The query_embedding_model path")
+parser.add_argument("--passage_embedding_model", default="rocketqa-zh-base-query-encoder", type=str, help="The passage_embedding_model path")
+parser.add_argument("--params_path", default="checkpoints/model_40/model_state.pdparams", type=str, help="The checkpoint path")
+parser.add_argument("--embedding_dim", default=768, type=int, help="The embedding_dim of index")
+parser.add_argument("--search_api_key", default=None, type=str, help="The SerpAPI key.")
+parser.add_argument('--embed_title', default=False, type=bool, help="The title to be  embedded into embedding")
+parser.add_argument('--model_type', choices=['ernie_search', 'ernie', 'bert', 'neural_search'], default="ernie", help="the ernie model types")
+parser.add_argument('--llm_name', choices=['ernie-bot', 'THUDM/chatglm-6b', "gpt-3.5-turbo", "gpt-4"], default="THUDM/chatglm-6b", help="The chatbot models ")
+parser.add_argument("--api_key", default=None, type=str, help="The API Key.")
+parser.add_argument("--secret_key", default=None, type=str, help="The secret key.")
+args = parser.parse_args()
+# yapf: enable
+
+
+def indexing_files(retriever, document_store, filepaths, chunk_size):
+    try:
+        text_converter = TextConverter()
+        pdf_converter = PDFToTextConverter()
+        doc_converter = DocxToTextConverter()
+
+        text_splitter = CharacterTextSplitter(separator="\f", chunk_size=chunk_size, chunk_overlap=0, filters=["\n"])
+        pdf_splitter = CharacterTextSplitter(
+            separator="\f",
+            chunk_size=chunk_size,
+            chunk_overlap=0,
+            filters=['([﹒﹔﹖﹗．。！？]["’”」』]{0,2}|(?=["‘“「『]{1,2}|$))'],
+        )
+        file_classifier = FileTypeClassifier()
+        indexing_pipeline = Pipeline()
+        indexing_pipeline.add_node(component=file_classifier, name="file_classifier", inputs=["File"])
+        indexing_pipeline.add_node(component=doc_converter, name="DocConverter", inputs=["file_classifier.output_4"])
+        indexing_pipeline.add_node(component=text_converter, name="TextConverter", inputs=["file_classifier.output_1"])
+        indexing_pipeline.add_node(component=pdf_converter, name="PDFConverter", inputs=["file_classifier.output_2"])
+
+        indexing_pipeline.add_node(
+            component=text_splitter, name="TextSplitter", inputs=["TextConverter", "DocConverter"]
+        )
+        indexing_pipeline.add_node(component=pdf_splitter, name="PDFSplitter", inputs=["PDFConverter"])
+        indexing_pipeline.add_node(component=retriever, name="Retriever", inputs=["TextSplitter", "PDFSplitter"])
+        indexing_pipeline.add_node(component=document_store, name="DocumentStore", inputs=["Retriever"])
+        files = glob.glob(filepaths + "/*.*", recursive=True)
+        indexing_pipeline.run(file_paths=files)
+    except Exception as e:
+        print(e)
+        pass
+
+
+def get_faiss_retriever(use_gpu):
+    faiss_document_store = "faiss_document_store.db"
+    if os.path.exists(args.index_name) and os.path.exists(faiss_document_store):
+        # connect to existed FAISS Index
+        document_store = FAISSDocumentStore.load(args.index_name)
+        retriever = DensePassageRetriever(
+            document_store=document_store,
+            query_embedding_model=args.query_embedding_model,
+            passage_embedding_model=args.passage_embedding_model,
+            params_path=args.params_path,
+            output_emb_size=args.embedding_dim if args.model_type in ["ernie_search", "neural_search"] else None,
+            max_seq_len_query=args.max_seq_len_query,
+            max_seq_len_passage=args.max_seq_len_passage,
+            batch_size=args.retriever_batch_size,
+            use_gpu=use_gpu,
+            embed_title=args.embed_title,
+        )
+    else:
+        dureader_data = "https://paddlenlp.bj.bcebos.com/applications/dureader_dev.zip"
+        zip_dir = "data/dureader_dev"
+        fetch_archive_from_http(url=dureader_data, output_dir=zip_dir)
+
+        document_store = FAISSDocumentStore(embedding_dim=args.embedding_dim, faiss_index_factory_str="Flat")
+        retriever = DensePassageRetriever(
+            document_store=document_store,
+            query_embedding_model=args.query_embedding_model,
+            passage_embedding_model=args.passage_embedding_model,
+            params_path=args.params_path,
+            output_emb_size=args.embedding_dim if args.model_type in ["ernie_search", "neural_search"] else None,
+            max_seq_len_query=args.max_seq_len_query,
+            max_seq_len_passage=args.max_seq_len_passage,
+            batch_size=args.retriever_batch_size,
+            use_gpu=use_gpu,
+            embed_title=args.embed_title,
+            top_k=5,
+        )
+        filepaths = "data/dureader_dev/dureader_dev"
+        indexing_files(retriever, document_store, filepaths, chunk_size=500)
+        document_store.save(args.index_name)
+    return retriever
+
+
+def search_and_action_example(web_retriever):
+
+    qa_template = PromptTemplate(
+        name="文档问答",
+        prompt_text="使用以下段落作为来源回答以下问题。"
+        "答案应该简短，最多几个字。\n"
+        "段落:\n{documents}\n"
+        "问题: {query}\n\n"
+        "说明: 考虑以上所有段落及其相应的分数，得出答案。 "
+        "虽然一个段落可能得分很高， "
+        "但重要的是要考虑同一候选答案的所有段落，以便准确回答。\n\n"
+        "在考虑了所有的可能性之后，最终答案是:\n",
+    )
+    pn = PromptNode(
+        args.llm_name,
+        max_length=512,
+        default_prompt_template=qa_template,
+        api_key=args.api_key,
+        secret_key=args.secret_key,
+    )
+
+    pipeline = WebQAPipeline(retriever=web_retriever, prompt_node=pn)
+
+    prompt_node = PromptNode(
+        args.llm_name, max_length=512, api_key=args.api_key, secret_key=args.secret_key, stop_words=["观察: "]
+    )
+
+    web_qa_tool = Tool(
+        name="搜索",
+        pipeline_or_node=pipeline,
+        description="当你需要用谷歌搜索问题时很有用。",
+        output_variable="results",
+    )
+    few_shot_agent_template = PromptTemplate("few-shot-react", prompt_text=few_shot_prompt)
+    # Time to initialize the Agent specifying the PromptNode to use and the Tools
+    agent = Agent(
+        prompt_node=prompt_node,
+        prompt_template=few_shot_agent_template,
+        tools_manager=ToolsManager(
+            tools=[web_qa_tool],
+            tool_pattern=r"工具:\s*(\w+)\s*工具输入:\s*(?:\"([\s\S]*?)\"|((?:.|\n)*))\s*",
+            observation_prefix="观察: ",
+            llm_prefix="思考: ",
+        ),
+        max_steps=8,
+        final_answer_pattern=r"最终答案\s*:\s*(.*)",
+        observation_prefix="观察: ",
+        llm_prefix="思考: ",
+    )
+    hotpot_questions = ["范冰冰的身高是多少?", "武则天传位给了谁？"]
+    for question in hotpot_questions:
+        result = agent.run(query=question)
+        print(f"\n{result['transcript']}")
+
+
+if __name__ == "__main__":
+    if args.retriever == "dense":
+        use_gpu = True if args.device == "gpu" else False
+        web_retriever = get_faiss_retriever(use_gpu)
+    else:
+        # https://serpapi.com/dashboard
+        web_retriever = WebRetriever(api_key=args.search_api_key, engine="bing", top_search_results=2)
+    search_and_action_example(web_retriever)

pipelines/pipelines/agents/agent_step.py

@@ -36,6 +36,8 @@ def __init__(
         final_answer_pattern: Optional[str] = None,
         prompt_node_response: str = "",
         transcript: str = "",
+        observation_prefix: str = "Observation: ",
+        llm_prefix: str = "Thought: ",
     ):
         """
         :param current_step: The current step in the execution of the agent.
@@ -50,6 +52,8 @@ def __init__(
         self.final_answer_pattern = final_answer_pattern or r"^([\s\S]+)$"
         self.prompt_node_response = prompt_node_response
         self.transcript = transcript
+        self.observation_prefix = observation_prefix
+        self.llm_prefix = llm_prefix
 
     def create_next_step(self, prompt_node_response: Any, current_step: Optional[int] = None) -> AgentStep:
         """
@@ -119,7 +123,7 @@ def completed(self, observation: Optional[str]) -> None:
         :param observation: received observation from the Agent environment.
         """
         self.transcript += (
-            f"{self.prompt_node_response}\nObservation: {observation}\nThought:"
+            f"{self.prompt_node_response}\n{self.observation_prefix} {observation}\n{self.llm_prefix} "
             if observation
             else self.prompt_node_response
         )
@@ -149,7 +153,6 @@ def parse_final_answer(self) -> Optional[str]:
         """
         # Search for a match with the final answer pattern in the prompt node response
         final_answer_match = re.search(self.final_answer_pattern, self.prompt_node_response)
-
         if final_answer_match:
             # If a match is found, get the first group (i.e., the content inside the parentheses of the regex pattern)
             final_answer = final_answer_match.group(1)

pipelines/pipelines/agents/base.py

@@ -134,6 +134,8 @@ def __init__(
         self,
         tools: Optional[List[Tool]] = None,
         tool_pattern: str = r"Tool:\s*(\w+)\s*Tool Input:\s*(?:\"([\s\S]*?)\"|((?:.|\n)*))\s*",
+        observation_prefix: str = "Observation: ",
+        llm_prefix: str = "Thought: ",
     ):
         """
         :param tools: A list of tools to add to the ToolManager. Each tool must have a unique name.
@@ -143,6 +145,8 @@ def __init__(
         self._tools: Dict[str, Tool] = {tool.name: tool for tool in tools} if tools else {}
         self.tool_pattern = tool_pattern
         self.callback_manager = Events(("on_tool_start", "on_tool_finish", "on_tool_error"))
+        self.observation_prefix = observation_prefix
+        self.llm_prefix = llm_prefix
 
     def add_tool(self, tool: Tool):
         """
@@ -193,8 +197,8 @@ def run_tool(self, llm_response: str, params: Optional[Dict[str, Any]] = None) -
                     tool_result = tool.run(tool_input, params)
                     self.callback_manager.on_tool_finish(
                         tool_result,
-                        observation_prefix="Observation: ",
-                        llm_prefix="Thought: ",
+                        observation_prefix=f"{self.observation_prefix}",
+                        llm_prefix="{self.llm_prefix}",
                         color=tool.logging_color,
                     )
                 except Exception as e:
@@ -241,6 +245,8 @@ def __init__(
         prompt_parameters_resolver: Optional[Callable] = None,
         max_steps: int = 8,
         final_answer_pattern: str = r"Final Answer\s*:\s*(.*)",
+        observation_prefix: str = "Observation: ",
+        llm_prefix: str = "Thought: ",
     ):
         """
          Creates an Agent instance.
@@ -269,6 +275,8 @@ def __init__(
         self.prompt_node = prompt_node
         prompt_template = prompt_template or "zero-shot-react"
         resolved_prompt_template = prompt_node.get_prompt_template(prompt_template)
+        self.observation_prefix = observation_prefix
+        self.llm_prefix = llm_prefix
         if not resolved_prompt_template:
             raise ValueError(
                 f"Prompt template '{prompt_template}' not found. Please check the spelling of the template name."
@@ -394,18 +402,15 @@ def run(
     def _step(self, query: str, current_step: AgentStep, params: Optional[dict] = None):
         # plan next step using the LLM
         prompt_node_response = self._plan(query, current_step)
-
         # from the LLM response, create the next step
         next_step = current_step.create_next_step(prompt_node_response)
         self.callback_manager.on_agent_step(next_step)
-
         # run the tool selected by the LLM
         observation = self.tm.run_tool(next_step.prompt_node_response, params) if not next_step.is_last() else None
 
         # save the input, output and observation to memory (if memory is enabled)
         memory_data = self.prepare_data_for_memory(input=query, output=prompt_node_response, observation=observation)
         self.memory.save(data=memory_data)
-
         # update the next step with the observation
         next_step.completed(observation)
         return next_step

pipelines/pipelines/nodes/llm/chatglm.py

@@ -21,7 +21,14 @@
 
 
 class ChatGLMBot(BaseComponent):
-    def __init__(self, batch_size: int = 2, max_seq_length: int = 1024, tgt_length: int = 256, **kwargs):
+    def __init__(
+        self,
+        model_name_or_path="THUDM/chatglm-6b-v1.1",
+        batch_size: int = 2,
+        max_seq_length: int = 2048,
+        tgt_length: int = 2048,
+        **kwargs
+    ):
         """
         Initialize the ChatGLMBot instance.
 
@@ -32,18 +39,23 @@ def __init__(self, batch_size: int = 2, max_seq_length: int = 1024, tgt_length:
         self.kwargs = kwargs
         self.chatglm = Taskflow(
             "text2text_generation",
+            model=model_name_or_path,
             batch_size=batch_size,
             max_seq_length=max_seq_length,
             tgt_length=tgt_length,
             **self.kwargs,
         )
 
+    def predict(self, query, stream=False):
+        result = self.chatglm(query)
+        return result
+
     def run(self, query, stream=False):
         """
         Using the chatbot to generate the answers
         :param query: The user's input/query to be sent to the chatGLM.
         :param stream: Whether to use streaming mode when making the request. Currently not in use. Defaults to False.
         """
         logger.info(query)
-        result = self.chatglm(query)
+        result = self.predict(query=query, stream=stream)
         return result, "output_1"