[APP]Add EB QA Bot (#263)

* add qa bot

* add code force

* mypy

* fix md load bug

* fix md load bug

* add summary

* rm origin file

* add des

* add init info

* add notes

* reformat and rename

* add readme

* add readme

* fix des

* add note

* add faiss vectorbase

* revise

* revise

* rm faiss

* Update README.md

修改ERNIE Bot Agent

* rm db file

* rm db file

* update EB

erniebot-agent/applications/eb-agent-qa-bot/README.md

@@ -0,0 +1,43 @@
+# ERNIE Bot Agent QA Bot
+
+ERNIE Bot Agent QA Bot是一个ERNIE Bot Agent使用教学机器人，基于 `FunctionAgentWithRetrieval`，该工具旨在辅助用户解决与EB-Agent相关的问题，帮助用户更快的使用 `erniebot_agent`库，搭建属于自己的Agent。
+
+## 架构
+
+此应用基于 `FunctionAgentWithRetrieval`（后续 `RetrievalAgent`上线后将同步更换），将此仓库中相关模块的markdown文件以及ipynb的示例代码文件向量化并通过自定义检索工具检索，实现EB-Agent教学机器人。
+
+### 自定义检索工具
+
+此应用中的检索工具基于 `langchain`的 `faiss`本地向量库，同时基于此应用特性，用户可能需要了解具体的代码实现。因此在实现时同时检索召回说明文档的内容(存储于db)以及相关的代码内容(存储于module_code_db)。
+
+```python
+class FaissSearch:
+    def __init__(self, db, embeddings, module_code_db):
+        self.db = db
+        self.module_code_db = module_code_db
+        self.embeddings = embeddings
+```
+
+## 如何开始
+
+**注意：** 建库的过程比较缓慢，请耐心等待。
+
+> 第一步：下载项目源代码，请确保您已经安装了erniebot_agent以及erniebot
+
+```bash
+git clone https://github.com/PaddlePaddle/ERNIE-Bot-SDK.git
+cd ERNIE-Bot-SDK
+pip install ernie_agent
+```
+
+> 第二步：如果是第一次运行，请先初始化向量库(应用中同时上传了向量库也可以)
+
+```bash
+python question_bot.py --init=True --access-token <aistudio-access-token>
+```
+
+> 如果已经初始化过向量库，直接运行即可
+
+```bash
+python question_bot.py --access-token <aistudio-access-token>
+```

erniebot-agent/applications/eb-agent-qa-bot/init_vector_db.py

@@ -0,0 +1,119 @@
+from typing import List, Union
+
+import erniebot
+import nbformat
+from langchain.text_splitter import (
+    MarkdownHeaderTextSplitter,
+    RecursiveCharacterTextSplitter,
+)
+from langchain.vectorstores import FAISS
+from langchain_core.documents import Document
+from tqdm import tqdm
+
+headers_to_split_on = [
+    ("#", "Header 1"),
+    ("##", "Header 2"),
+    # ("###", "Header 3"),
+    # ("####", "Header 4"),
+]
+
+
+def get_summary(content: str) -> Union[str, None]:
+    """Get summary of md files, you can also change another llm model."""
+    chat_message = {"role": "user", "content": f"请帮我给以下markdown文件生成摘要用于用户问文档内容时的检索匹配，不要超过400个字：\n{content}"}
+    summary = erniebot.ChatCompletion.create(
+        model="ernie-longtext",
+        messages=[chat_message],
+    ).get_result()
+    return summary
+
+
+def open_and_concatenate_ipynb(ipynb_path: str, encoding: str) -> str:
+    """Get content of ipynb file."""
+    with open(ipynb_path, "r", encoding=encoding) as f:
+        notebook_content = nbformat.read(f, as_version=4)
+
+    # 按顺序拼接code单元
+    concatenated_content = ""
+    for cell in notebook_content["cells"]:
+        if cell["cell_type"] == "code":
+            concatenated_content += "```python\n" + cell["source"] + "```\n\n"
+
+    return concatenated_content
+
+
+def read_md_file(file_path: str) -> Union[str, None]:
+    try:
+        with open(file_path, "r", encoding="utf-8") as file:
+            md_content = file.read()
+        return md_content
+    except FileNotFoundError:
+        print(f"文件 '{file_path}' 不存在。")
+        return None
+    except Exception as e:
+        print(f"读取文件时出现错误： {e}")
+        return None
+
+
+def load_md_files_to_doc(
+    file_paths: List[str],
+    chunk_size: int = 1000,
+    chunk_overlap: int = 30,
+) -> List[Document]:
+    markdown_splitter = MarkdownHeaderTextSplitter(headers_to_split_on=headers_to_split_on)
+    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
+    output_document = []
+    for file in tqdm(file_paths):
+        content = read_md_file(file)
+        if content is None:
+            continue
+        md_header_splits = markdown_splitter.split_text(content)
+        splits = text_splitter.split_documents(md_header_splits)
+        for i in range(len(splits)):
+            # 生成summary用于检索
+            splits[i].metadata["raw_text"] = splits[i].page_content
+            splits[i].page_content = get_summary(splits[i].page_content)
+        output_document.extend(splits)
+    return output_document
+
+
+def init_db(faiss_name, faiss_name_module, embeddings):
+    md_file_path = [
+        "./docs/modules/file.md",
+        "./docs/modules/agents.md",
+        "./docs/modules/memory.md",
+        "./docs/modules/message.md",
+        "./docs/modules/chat_models.md",
+        "./docs/modules/tools.md",
+        "./docs/quickstart/agent.md",
+        "./docs/quickstart/use-tool.md",
+    ]
+    chunk_size = 1000
+    chunk_overlap = 30
+    content_doc = load_md_files_to_doc(md_file_path, chunk_size, chunk_overlap)
+
+    db = FAISS.from_documents(content_doc, embeddings)
+    db.save_local(faiss_name)
+
+    ipynb_path = [
+        "./docs/cookbooks/agent/function_agent.ipynb",
+        "./docs/cookbooks/agent/chat_models.ipynb",
+        "./docs/cookbooks/agent/memory.ipynb",
+        "./docs/cookbooks/agent/message.ipynb",
+        "./docs/cookbooks/agent/local_tool.ipynb",
+        "./docs/cookbooks/agent/tools_intro.ipynb",
+        "./docs/cookbooks/agent/remote-tool/remote_tool.ipynb",
+    ]
+    modules = [item[item.rfind("/") + 1 : item.rfind(".ipynb")] for item in ipynb_path]
+    module_doc = []
+
+    for i in range(len(modules)):
+        module_doc.append(
+            Document(
+                page_content=modules[i],
+                metadata={"ipynb": open_and_concatenate_ipynb(ipynb_path[i], "utf-8")},
+            )
+        )
+
+    module_code_db = FAISS.from_documents(module_doc, embeddings)
+    module_code_db.save_local(faiss_name_module)

erniebot-agent/applications/eb-agent-qa-bot/question_bot.py

@@ -0,0 +1,101 @@
+import argparse
+import os
+
+os.environ["EB_AGENT_LOGGING_LEVEL"] = "INFO"
+
+from init_vector_db import init_db
+from langchain.vectorstores import FAISS
+from sklearn.metrics.pairwise import cosine_similarity
+
+from erniebot_agent.agents.function_agent_with_retrieval import (
+    FunctionAgentWithRetrieval,
+)
+from erniebot_agent.chat_models import ERNIEBot
+from erniebot_agent.extensions.langchain.embeddings import ErnieEmbeddings
+from erniebot_agent.memory import SystemMessage
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--init", type=bool, default=False)
+parser.add_argument("--access-token", type=str, help="access token for erniebot-agent")
+args = parser.parse_args()
+
+if args.access_token:
+    os.environ["EB_AGENT_ACCESS_TOKEN"] = args.access_token
+
+embeddings = ErnieEmbeddings(aistudio_access_token=os.environ["EB_AGENT_ACCESS_TOKEN"], chunk_size=16)
+
+
+class FaissSearch:
+    def __init__(self, db, embeddings, module_code_db):
+        self.db = db
+        self.module_code_db = module_code_db
+        self.embeddings = embeddings
+
+    def search(self, query: str, top_k: int = 2):
+        # 搜索时，同时召回最相关的两个文档片段以及最相关的一个代码示例
+        docs = self.db.similarity_search(query, top_k)
+        para_result = self.embeddings.embed_documents([i.page_content for i in docs])
+        query_result = self.embeddings.embed_query(query)
+        similarities = cosine_similarity([query_result], para_result).reshape((-1,))
+        retrieval_results = []
+        # make sure 'raw_text' in doc.metadata
+        for index, doc in enumerate(docs):
+            if "Header 1" in doc.metadata:
+                retrieval_results.append(
+                    {
+                        "content": doc.metadata["raw_text"],
+                        "score": similarities[index],
+                        "title": doc.metadata["Header 1"],
+                    }
+                )
+            else:
+                retrieval_results.append(
+                    {"content": doc.metadata["raw_text"], "score": similarities[index], "title": ""}
+                )
+        # module_code_db 用于相关代码的召回
+        code = self.module_code_db.similarity_search(query, 1)[0]
+        # make sure 'ipynb' in code.metadata
+        retrieval_results.append({"content": code.metadata["ipynb"], "score": 1, "title": code.page_content})
+
+        return retrieval_results
+
+
+def load_agent():
+    faiss_name = os.path.join(os.path.dirname(os.path.abspath(__file__)), "faiss_index")
+    faiss_name_module = os.path.join(os.path.dirname(os.path.abspath(__file__)), "faiss_index_module")
+    if args.init:
+        init_db(faiss_name, faiss_name_module, embeddings)
+    try:
+        db = FAISS.load_local(faiss_name, embeddings)
+        module_code_db = FAISS.load_local(faiss_name_module, embeddings)
+    except RuntimeError as e:
+        raise RuntimeError(f"Make sure you have initialized the database first.\n {e}")
+
+    llm = ERNIEBot(model="ernie-3.5")
+    faiss_search = FaissSearch(db=db, embeddings=embeddings, module_code_db=module_code_db)
+    agent = FunctionAgentWithRetrieval(
+        llm=llm,
+        tools=[],
+        knowledge_base=faiss_search,
+        threshold=0,
+        system_message=SystemMessage(
+            "你是ERNIEBot Agent的小助手，用于解决用户关于EB-Agent的问题，涉及File, Memory, Message, Agent, ChatModels等模块。"
+            "请你严格按照搜索到的内容回答，不要自己生成相关代码。如果询问与ERNIEBot Agent无关的问题，请直接回答“我只能回答EB—Agent相关问题”"
+        ),
+        top_k=2,
+        token_limit=5000,
+    )
+    return agent
+
+
+async def main(agent):
+    # response = await agent.run('怎么从aistudio创建远程tool？')
+    # response = await agent.run('如何创建一个LocalTool？')
+    response = await agent.run("如何创建一个agent")
+    print(response.text)
+
+
+if __name__ == "__main__":
+    agent = load_agent()
+    # asyncio.run(main(agent))
+    agent.launch_gradio_demo()