调试、优化与强化我们的代理

代码示例

1Thought: 用户在询问“什么是...”某事物，这听起来像是在请求定义。我应该使用 `definition_tool` 来定义“四层查询”。
2Action: definition_tool
3Action Input: {"term": "four levels of queries"}
4Observation: Error: Could not find a Wikipedia page for 'four levels of queries'. Please try a different term.
5Thought: 该工具未能找到定义。我无法回答用户的问题。

代码示例

1description=(
2    "包含用户提供的特定研究论文的详细信息。"
3    "使用此工具来回答关于这些论文的深入、上下文相关问题。"
4)

代码示例

1description=(
2    "此工具提供对一系列关于 RAG 的特定研究论文的访问。"
3    "它能够回答这些论文内容的深入、上下文相关问题，"
4    "包括其作者、如 eRAG 的方法，以及查询层级（L1-L4）。"
5)

代码示例

1except wikipedia.exceptions.DisambiguationError as e:
2    return f"Error: '{term}' is ambiguous. Please be more specific."

代码示例

1except wikipedia.exceptions.DisambiguationError as e:
2    # 给代理提供可用来恢复的选项！
3    return f"Error: '{term}' is ambiguous. Please be more specific. Did you mean one of these: {e.options[:3]}?"

代码示例

1system_prompt = """You are a helpful assistant...
2Only use a tool if the user's question directly relates to the tool's description."""

代码示例

1system_prompt = """You are a helpful assistant. Your primary goal is to answer user questions accurately.
2You have access to a set of tools. Here are the descriptions of your tools:
3---
4{tool_desc}
5---
6Carefully analyze the user's query and the tool descriptions to decide if a tool is relevant.
7If a tool returns an error or if the information is insufficient,
8do not give up. Re-evaluate the query and decide if you should try another tool
9or if you can answer from your own knowledge.
10"""

代码示例

1Thought: The user is asking about a person and a topic related to my documents ("Transformers"). I should use the `rag_research_paper_tool` to find this information.
2Action: rag_research_paper_tool
3Action Input: {"query": "Information on Yann LeCun and what he wrote about Transformers"}
4Observation: I'm sorry, but the provided documents do not contain specific information about Yann LeCun's writings on Transformers. The papers focus on the work of other authors.
5Thought: The tool could not find the information. I am unable to answer the user's question.

代码示例

1description=(
2    "This tool provides access to a collection of specific research papers about RAG..."
3),

代码示例

1description=(
2    "Contains detailed information from specific research papers by authors like Vaswani and Salemi. "
3    "It is an expert on the 'RAG and Beyond' and 'eRAG' papers. Use this for deep, contextual questions about THEIR content."
4),

代码示例

1Thought: The user is asking about "Yann LeCun". I will check my tool descriptions. The `rag_research_paper_tool` description mentions Vaswani and Salemi, not LeCun. Therefore, that tool is not relevant. I should first find out who Yann LeCun is. The `definition_tool` is best for this.
2Action: definition_tool
3Action Input: {"term": "Yann LeCun"}
4Observation: Yann LeCun is a French computer scientist working primarily in the fields of machine learning, computer vision, mobile robotics, and computational neuroscience...
5Thought: Now that I know who Yann LeCun is, I need to find what he wrote about Transformers. The `arxiv_search_tool` is the correct tool for finding papers by a specific author.
6Action: arxiv_search_tool
7Action Input: {"query": "Transformers", "author": "Yann LeCun"}
8Observation: [List of papers by Yann LeCun related to Transformers]
9Thought: I have successfully found who Yann LeCun is and a list of his relevant papers. I can now synthesize this information into a final answer.

一个示例追踪

1Running step init_run
2Step init_run → AgentInput
3Running step setup_agent
4Step setup_agent → AgentSetup
5Running step run_agent_step
6LLM Thought: "User is asking for a definition."
7Parsed Action: CALL_TOOL name="definition_tool"
8Parsed Action Input: {"query": "hyperparameter"}   # <-- note the key
9Running step call_tool
10Tool definition_tool invoked with kwargs={'query': 'hyperparameter'}
11ToolError: TypeError: define_term() missing 1 required positional argument: 'term'
12Step call_tool → ToolExecutionError
13Running step aggregate_tool_results
14Agent Observation: "definition_tool failed: missing required arg 'term'"
15Running step run_agent_step
16LLM Thought: "Tool failed. Returning general answer."
17Agent Output: "A hyperparameter is..."