开发 LlamaIndex 查询流水线代理

使用集合让一切井井有条 根据您的偏好保存内容并对其进行分类。

本页介绍了如何使用 LlamaIndex Query Pipelines 模板（Vertex AI SDK for Python 中的 LlamaIndexQueryPipelineAgent 类）开发代理。此代理旨在使用检索增强生成 (RAG) 技术回答问题，例如以下查询：“Paul Graham 的大学生活是怎样的？”

如需使用 LlamaIndex 查询流水线开发代理，请按以下步骤操作：

1. 定义和配置模型
2. 定义和使用检索器
3. 定义和使用响应合成器
4. （可选）自定义提示模板
5. （可选）自定义编排

准备工作

请按照设置环境中的步骤设置您的环境。

定义和配置模型

为 LlamaIndex 查询流水线代理定义和配置模型以供其使用。

1. 定义模型版本：

   model = "gemini-2.0-flash"
   

2. （可选）指定模型参数：

   model_kwargs = {
       # vertexai_config (dict): By providing the region and project_id parameters,
       # you can enable model usage through Vertex AI.
       "vertexai_config": {
           "project": "PROJECT_ID",
           "location": "LOCATION"
       },
       # temperature (float): The sampling temperature controls the degree of
       # randomness in token selection.
       "temperature": 0.28,
       # context_window (int): The context window of the model.
       # If not provided, the default context window is 200000.
       "context_window": 200000,
       # max_tokens (int): Token limit determines the maximum
       # amount of text output from one prompt. If not provided,
       # the default max_tokens is 256.
       "max_tokens": 256,
   }
   

3. 使用以下模型配置创建 LlamaIndexQueryPipelineAgent：

   from vertexai.preview import reasoning_engines
   
   agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
       model=model,                # Required.
       model_kwargs=model_kwargs,  # Optional.
   )
   

4. 如果您是在交互式环境（例如终端或 Colab 笔记本）中运行，则可以查询代理：

   response = agent.query(input="What is Paul Graham's life in college?")
   
   print(response)
   

   您应该会收到类似如下所示的响应：

   {'message': {'role': 'assistant',
     'additional_kwargs': {},
     'blocks': [{'block_type': 'text',
       'text': "Unfortunately, there's not a lot of publicly available information about Paul Graham's personal life in college. ..."}]},
     'raw': {'content': {'parts': [{'video_metadata': None,
         'thought': None,
         'code_execution_result': None,
         'executable_code': None,
         'file_data': None,
         'function_call': None,
         'function_response': None,
         'inline_data': None,
         'text': "Unfortunately, there's not a lot of publicly available information about Paul Graham's personal life in college. ..."}],
       'role': 'model'},
       'citation_metadata': None,
       'finish_message': None,
       'token_count': None,
       'avg_logprobs': -0.1468650027438327,
       'finish_reason': 'STOP',
       'grounding_metadata': None,
       'index': None,
       'logprobs_result': None,
       'safety_ratings': [{'blocked': None,
         'category': 'HARM_CATEGORY_HATE_SPEECH',
         'probability': 'NEGLIGIBLE',
         'probability_score': 0.022949219,
         'severity': 'HARM_SEVERITY_NEGLIGIBLE',
         'severity_score': 0.014038086},
       {'blocked': None,
         'category': 'HARM_CATEGORY_DANGEROUS_CONTENT',
         'probability': 'NEGLIGIBLE',
         'probability_score': 0.056640625,
         'severity': 'HARM_SEVERITY_NEGLIGIBLE',
         'severity_score': 0.029296875},
       {'blocked': None,
         'category': 'HARM_CATEGORY_HARASSMENT',
         'probability': 'NEGLIGIBLE',
         'probability_score': 0.071777344,
         'severity': 'HARM_SEVERITY_NEGLIGIBLE',
         'severity_score': 0.024047852},
       {'blocked': None,
         'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
         'probability': 'NEGLIGIBLE',
         'probability_score': 0.103515625,
         'severity': 'HARM_SEVERITY_NEGLIGIBLE',
         'severity_score': 0.05102539}],
       'usage_metadata': {'cached_content_token_count': None,
       'candidates_token_count': 222,
       'prompt_token_count': 10,
       'total_token_count': 232}},
     'delta': None,
     'logprobs': None,
     'additional_kwargs': {}}
   

（可选）自定义模型

LlamaIndexQueryPipelineAgent 模板默认使用 Google GenAI 来提供对 Google Cloud中所有基础模型的访问权限。如需使用无法通过 Google GenAI 获取的模型，请按如下方式定义 model_builder=：

from typing import Optional

def model_builder(
    *,
    model_name: str,                      # Required. The name of the model
    model_kwargs: Optional[dict] = None,  # Optional. The model keyword arguments.
    **kwargs,                             # Optional. The remaining keyword arguments to be ignored.
):

如需查看 LlamaIndexQueryPipeline 中支持的聊天模型及其功能的列表，请参阅可用的 LLM 集成。 每个聊天模型都使用自己的一组支持的 model= 和 model_kwargs= 值。

from vertexai.preview import reasoning_engines

agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
    model=model,                # Required.
    model_kwargs=model_kwargs,  # Optional.
)

定义和使用检索器

定义模型后，定义模型用于推理的检索器。检索器可以基于索引构建，但也可以全面定义。您应在本地测试检索器。

1. 定义一个检索器，用于返回相关文档和相似度得分：

   def retriever_builder(model, retriever_kwargs=None):
       import os
       import requests
       from llama_index.core import (
           StorageContext,
           VectorStoreIndex,
           load_index_from_storage,
       )
       from llama_index.core import SimpleDirectoryReader
       from llama_index.embeddings.vertex import VertexTextEmbedding
       import google.auth
   
       credentials, _ = google.auth.default()
       embed_model = VertexTextEmbedding(
           model_name="textembedding-gecko@003", project="PROJECT_ID", credentials=credentials
       )
   
       data_dir = "data/paul_graham"
       essay_file = os.path.join(data_dir, "paul_graham_essay.txt")
       storage_dir = "storage"
   
       # --- Simple Download (if needed) ---
       if not os.path.exists(essay_file):
           os.makedirs(data_dir, exist_ok=True)  # Make sure the directory exists
           essay_url = "https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/paul_graham/paul_graham_essay.txt"
           try:
               response = requests.get(essay_url)
               response.raise_for_status()  # Check for download errors
               with open(essay_file, "wb") as f:
                   f.write(response.content)
               print("Essay downloaded.")
           except requests.exceptions.RequestException as e:
               print(f"Download failed: {e}")
   
       # --- Build/Load Index ---
       if not os.path.exists(storage_dir):
           print("Creating new index...")
           # --- Load Data ---
           reader = SimpleDirectoryReader(data_dir)
           docs = reader.load_data()
   
           index = VectorStoreIndex.from_documents(docs, model=model, embed_model=embed_model)
           index.storage_context.persist(persist_dir=storage_dir)
       else:
           print("Loading existing index...")
           storage_context = StorageContext.from_defaults(persist_dir=storage_dir)
           index = load_index_from_storage(storage_context, embed_model=embed_model)
   
       return index.as_retriever()
   

2. 测试检索器：

   from llama_index.llms.google_genai import GoogleGenAI
   
   model = GoogleGenAI(
       model=model,
       **model_kwargs
   )
   retriever = retriever_builder(model)
   retrieved_response = retriever.retrieve("What is Paul Graham's life in College?")
   

   检索到的响应应类似如下所示：

   [
     NodeWithScore(
       node=TextNode(
         id_='692a5d5c-cd56-4ed0-8e29-ecadf6eb9933',
         embedding=None,
         metadata={'file_path': '/content/data/paul_graham/paul_graham_essay.txt', 'file_name': 'paul_graham_essay.txt', 'file_type': 'text/plain', 'file_size': 75042, 'creation_date': '2025-03-24', 'last_modified_date': '2025-03-24'},
         excluded_embed_metadata_keys=['file_name', 'file_type', 'file_size', 'creation_date', 'last_modified_date', 'last_accessed_date'],
         excluded_llm_metadata_keys=['file_name', 'file_type', 'file_size', 'creation_date', 'last_modified_date', 'last_accessed_date'],
         relationships={
           <NodeRelationship.SOURCE: '1'>: RelatedNodeInfo(node_id='3e1c4d73-1e1d-4e83-bd16-2dae24abb231', node_type='4', metadata={'file_path': '/content/data/paul_graham/paul_graham_essay.txt', 'file_name': 'paul_graham_essay.txt', 'file_type': 'text/plain', 'file_size': 75042, 'creation_date': '2025-03-24', 'last_modified_date': '2025-03-24'}, hash='0c3c3f46cac874b495d944dfc4b920f6b68817dbbb1699ecc955d1fafb2bf87b'), <NodeRelationship.PREVIOUS: '2'>: RelatedNodeInfo(node_id='782c5787-8753-4f65-85ed-c2833ea6d4d8', node_type='1', metadata={'file_path': '/content/data/paul_graham/paul_graham_essay.txt', 'file_name': 'paul_graham_essay.txt', 'file_type': 'text/plain', 'file_size': 75042, 'creation_date': '2025-03-24', 'last_modified_date': '2025-03-24'}, hash='b8e6463833887a8a2b13f1b5a623672819faedc1b725d9565ba003223628db0e'), <NodeRelationship.NEXT: '3'>: RelatedNodeInfo(node_id='f7d2cb7e-fa0c-40bf-b8e7-b888e36b87f9', node_type='1', metadata={}, hash='db7cc1a67fa3afd1e5f24c8c61583781ce6a00c444da8f25a5374468c17b7de0')
         },
         metadata_template='{key}: {value}',
         metadata_separator='\n',
         text='So I looked around to see what I could salvage from the wreckage of my plans, and there was Lisp...',
         mimetype='text/plain',
         start_char_idx=7166,
         end_char_idx=11549,
         metadata_separator='\n',
         text_template='{metadata_str}\n\n{content}'
       ),
       score=0.7403571819090398
     )
   ]
   

3. 如需在 LlamaIndexQueryPipelineAgent 模板内使用该检索工具，请将其添加到 retriever_builder= 参数下：

   from vertexai.preview import reasoning_engines
   
   agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
       model=model,                          # Required.
       model_kwargs=model_kwargs,            # Optional.
       retriever_builder=retriever_builder,  # Optional.
   )
   

4. 通过发出测试查询在本地测试代理：

   response = agent.query(
       input="What is Paul Graham's life in College?"
   )
   

   响应是一个 JSON 可序列化的节点列表，其中包含得分。

   [{'node': {'id_': '692a5d5c-cd56-4ed0-8e29-ecadf6eb9933',
     'embedding': None,
     'metadata': {'file_path': '/content/data/paul_graham/paul_graham_essay.txt',
       'file_name': 'paul_graham_essay.txt',
       'file_type': 'text/plain',
       'file_size': 75042,
       'creation_date': '2025-03-12',
       'last_modified_date': '2025-03-12'},
     'excluded_embed_metadata_keys': ['file_name',
       'file_type',
       'file_size',
       'creation_date',
       'last_modified_date',
       'last_accessed_date'],
     'excluded_llm_metadata_keys': ['file_name',
       'file_type',
       'file_size',
       'creation_date',
       'last_modified_date',
       'last_accessed_date'],
     'relationships': {'1': {'node_id': '07ee9574-04c8-46c7-b023-b22ba9558a1f',
       'node_type': '1',
       'metadata': {},
       'hash': '44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a',
       'class_name': 'RelatedNodeInfo'},
       '2': {'node_id': 'ac7e54aa-6fff-40b5-a15e-89c5eb234936',
       'node_type': '1',
       'metadata': {'file_path': '/content/data/paul_graham/paul_graham_essay.txt',
         'file_name': 'paul_graham_essay.txt',
         'file_type': 'text/plain',
         'file_size': 75042,
         'creation_date': '2025-03-12',
         'last_modified_date': '2025-03-12'},
       'hash': '755327a01efe7104db771e4e6f9683417884ea6895d878da882d2b21a6b66442',
       'class_name': 'RelatedNodeInfo'},
       '3': {'node_id': '3a04be27-ac46-4acd-a8c6-031689508982',
       'node_type': '1',
       'metadata': {},
       'hash': 'db7cc1a67fa3afd1e5f24c8c61583781ce6a00c444da8f25a5374468c17b7de0',
       'class_name': 'RelatedNodeInfo'}},
     'metadata_template': '{key}: {value}',
     'metadata_separator': '\n',
     'text': 'So I looked around to see what I could salvage from the wreckage of my plans, and there was Lisp...',
     'mimetype': 'text/plain',
     'start_char_idx': 7164,
     'end_char_idx': 11547,
     'metadata_separator': '\n',
     'text_template': '{metadata_str}\n\n{content}',
     'class_name': 'TextNode'},
     'score': 0.25325886336265013,
     'class_name': 'NodeWithScore'}
   ]
   

定义和使用响应合成器

定义模型和检索器后，定义响应合成器，该合成器使用用户询问和给定的一组文本块从 LLM 生成回答。您可以使用默认的 get_response_synthesizer，也可以配置响应模式。

1. 定义用于返回回答的回答合成器：

   def response_synthesizer_builder(model, response_synthesizer_kwargs=None):
       from llama_index.core.response_synthesizers import SimpleSummarize
   
       return SimpleSummarize(llm=model)
   

2. 测试函数：

   response_synthesizer = response_synthesizer_builder(model=model)
   response = response_synthesizer.get_response(
       "What is Paul Graham's life in College?",
       [node.model_dump_json() for node in retrieved_response],
   )
   

   响应应该类似以下内容：

   "While in a PhD program for computer science, he took art classes and worked on a book about Lisp hacking. He applied to art schools, got accepted to RISD, and later got an invitation to take the entrance exam at the Accademia di Belli Arti in Florence. He was accepted to both. He attended the Accademia, but was disappointed by the lack of instruction."
   

3. 如需在 LlamaIndexQueryPipeline 模板内使用响应合成器，请将其添加到 response_synthesizer_builder= 参数下：

   from vertexai.preview import reasoning_engines
   
   agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
       model=model,                                                    # Required.
       model_kwargs=model_kwargs,                                      # Optional.
       retriever_builder=retriever_builder,                            # Optional.
       response_synthesizer_builder=response_synthesizer_builder,      # Optional.
   )
   

4. 通过运行测试查询在本地测试完整的 RAG 查询流水线：

   response = agent.query(
       input="What is Paul Graham's life in College?"
   )
   

   响应是一个类似于以下内容的字典：

   {
     'response': "While in college, he was drawn to McCarthy's 1960 Lisp, although he didn't fully grasp the reasons for his interest at the time. He also had a brief encounter with surplus Xerox Dandelions in the computer lab but found them too slow for his liking. \n",
     'source_nodes': [
       '{"node":{"id_":"95889c30-53c7-43d0-bf91-930dbb23bde6"...,"score":0.7077213268404997,"class_name":"NodeWithScore"}'
     ],
     'metadata': {
       '95889c30-53c7-43d0-bf91-930dbb23bde6': {
         'file_path': '/content/data/paul_graham/paul_graham_essay.txt',
         'file_name': 'paul_graham_essay.txt',
         'file_type': 'text/plain',
         'file_size': 75042,
         'creation_date': '2025-03-25',
         'last_modified_date': '2025-03-25'
       }
     }
   }
   

（可选）自定义提示模板

提示模板可将用户输入转换为模型指令，引导回答，以生成与上下文相关且连贯的输出。如需了解详情，请参阅提示。

默认的提示模板会按顺序分为以下几个部分：

如果您在创建代理时未指定自己的提示模板，系统会生成默认提示模板，该模板如下所示：

from llama_index.core import prompts
from llama_index.core.base.llms import types

message_templates = [
  types.ChatMessage(role=types.MessageRole.SYSTEM, content=system_instruction),
  types.ChatMessage(role=types.MessageRole.USER, content="{input}"),
]
prompts.ChatPromptTemplate(message_templates=message_templates)

在以下示例中，您可以在实例化代理时使用完整的提示模板：

  from vertexai.preview import reasoning_engines

  system_instruction = "I help to find what is Paul Graham's life in College"

  agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
      model=model,
      system_instruction=system_instruction,
  )

您可以使用自己的提示模板替换默认提示模板，并在构建代理时使用该模板：

prompt_str = "Please answer {question} about {name}"
prompt_tmpl = PromptTemplate(prompt_str)

from vertexai.preview import reasoning_engines
agent = reasoning_engines.LlamaIndexQueryPipelineAgent(
    model = model,
    prompt = prompt_tmpl,
)

agent.query(
    input={
        "name": "Paul Graham",
        "question": "What is the life in college?",
    }
)

（可选）自定义编排

所有 LlamaIndexQueryPipeline 组件都实现了查询组件接口，该接口为编排提供了输入和输出架构。LlamaIndexQueryPipelineAgent 需要构建一个可响应查询的可运行对象。默认情况下，LlamaIndexQueryPipelineAgent 使用 Query Pipeline 构建顺序链或有向无环图 (DAG)。

如果您打算执行以下任一操作，则可能需要自定义编排：

• 实现用于扩展 RAG 流水线的代理（例如，将现有的提示、模型、检索器、回答合成器模块扩展到查询引擎、查询转换器、输出解析器、后处理器/重新排名器或自定义查询组件）。

• 使用 ReAct 提示代理执行工具，并为每个步骤添加注解，说明执行该步骤的原因。为此，请在创建 LlamaIndexQueryPipelineAgent 时替换默认可运行对象，方法是指定 runnable_builder= 参数：

  from typing import Optional
  from llama_index.core.llms import function_calling
  
  def runnable_builder(
      model: function_calling.FunctionCallingLLM,
      *,
      system_instruction: Optional[str] = None,
      prompt: Optional[query.QUERY_COMPONENT_TYPE] = None,
      retriever: Optional[query.QUERY_COMPONENT_TYPE] = None,
      response_synthesizer: Optional[query.QUERY_COMPONENT_TYPE] = None,
      runnable_kwargs: Optional[Mapping[str, Any]] = None,
  ):
  

  其中：

  • model 对应于从 model_builder 返回的聊天模型（请参阅定义和配置模型）。
  • retriever 和 retriever_kwargs 对应于要使用的检索器和配置（请参阅定义检索器）。
  • response_synthesizer 和 response_synthesizer_kwargs 对应于要使用的响应合成器和配置（请参阅定义响应合成器）。
  • system_instruction 和 prompt 对应于提示配置（请参阅自定义提示模板）。
  • agent_executor_kwargs 和 runnable_kwargs 是您可以用于自定义可运行对象的关键字参数。

您可以使用自定义流水线或 ReAct 自定义编排逻辑：

后续步骤

• 评估代理。
• 部署代理。
• 排查代理开发问题。
• 获取支持。