从零开始构建Deep Research Agent

report_structure = json.loads(clean_json_tags(remove_reasoning_from_output(response.choices[0].message.content)))

STATE = State()
for paragraph in report_structure:
STATE.paragraphs.append(Paragraph(title=paragraph["title"], content=paragraph["content"]))

实现步骤三：网络搜索工具
我们使用Tavily执行网络搜索。您可以在这里获取您的令牌。
搜索工具的实现非常简单：

def tavily_search(query, include_raw_content=True, max_results=5):
tavily_client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
return tavily_client.search(query,
include_raw_content=include_raw_content,
max_results=max_results)

调用该函数将返回最多max_results个搜索结果，每个结果包含标题、URL、内容摘要以及可能的完整页面内容（我们需要完整内容以获得最佳结果）。
获取搜索结果后，我们可以直接将其添加到全局状态中，无需再次调用LLM，但需要确保更新正确的段落索引。
以下函数可以根据Tavily搜索结果、段落索引和状态对象正确更新状态：

我们将追加到搜索历史中。latest_summary和reflection_iteration字段需要LLM进一步处理，将在后续“反思”部分讨论。这一工具的目的是为后续研究提供丰富的数据支持。
实现步骤四：规划搜索
以下提示词能持续产生较好的结果，用于规划首次搜索：

output_schema_first_search = {
"type": "object",
"properties": {
"search_query": {"type": "string"},
"reasoning": {"type": "string"}
}
}

SYSTEM_PROMPT_FIRST_SEARCH = f"""
You are a Deep Research assistant. You will be given a paragraph in a report, its title and expected content in the following json schema definition:

{json.dumps(input_schema_first_search, indent=2)}

You can use a web search tool that takes a 'search_query' as parameter.
Your job is to reflect on the topic and provide the most optimal web search query to enrich your current knowledge.
Format the output in json with the following json schema definition:

{json.dumps(output_schema_first_search, indent=2)}

Make sure that the output is a json object with an output json schema defined above.
Only return the json object, no explanation or additional text.
"""

在输出模式中要求推理内容，以促使模型对查询进行更多思考。虽然对于推理模型来说这可能有些多余，但对于普通LLM可能是个好主意。
现在，我们已经有了规划好的段落列表及其内容和描述，可以直接将步骤三的输出输入到提示词中：

print(response.choices[0].message.content)

首次搜索计划如下：

我们可以直接将该查询输入搜索工具：

实现步骤五：首次总结
首次总结与后续的反思步骤不同，因为此时尚无内容可供反思，本步骤将生成可供反思的内容。以下提示词效果较好：

output_schema_first_summary = {
"type": "object",
"properties": {
"paragraph_latest_state": {"type": "string"}
}
}

SYSTEM_PROMPT_FIRST_SUMMARY = f"""
You are a Deep Research assistant. You will be given a search query, search results and the paragraph of a report that you are researching following json schema definition:

{json.dumps(input_schema_first_summary, indent=2)}

Your job is to write the paragraph as a researcher using the search results to align with the paragraph topic and structure it properly to be included in the report.
Format the output in json with the following json schema definition:

{json.dumps(output_schema_first_summary, indent=2)}

Make sure that the output is a json object with an output json schema defined above.
Only return the json object, no explanation or additional text.
"""

应以以下格式为LLM提供数据：

我们可以从已有数据构建该JSON：

input_data = {
"title": "Introduction to Human Adaptability",
"content": "Humans possess a unique capacity for adaptability, which has been crucial in their survival and dominance across various environments. This introduction sets the stage for exploring the different facets of human adaptability.",
"search_query": "Homo sapiens characteristics basic biological traits cognitive abilities behavioral traits",
"search_results": [result["raw_content"][0:20000] for result in search_results["results"] if result["raw_content"]]
}

response = client.chat.completions.create(
model="DeepSeek-R1",
messages=[{"role":"system","content": SYSTEM_PROMPT_FIRST_SUMMARY},\
{"role":"user","content":json.dumps(input_data)}],
temperature=1
)

print(remove_reasoning_from_output(response.choices[0].message.content))

您将得到如下输出：

我们将用此内容更新STATE.paragraphs[0].research.latest_summary字段，并在后续“反思”部分继续基于段落总结的最新状态进行改进。这一总结的目的是为报告奠定基础内容。
实现步骤六：反思
有了报告段落内容的最新状态后，通过提示LLM反思文本，寻找可能遗漏的角度，以改进内容。以下提示词效果很好：

output_schema_reflection = {
"type": "object",
"properties": {
"search_query": {"type": "string"},
"reasoning": {"type": "string"}
}
}

SYSTEM_PROMPT_REFLECTION = f"""
You are a Deep Research assistant. You are responsible for constructing comprehensive paragraphs for a research report. You will be provided paragraph title and planned content summary, as well as the latest state of the paragraph that you have already created, all in the following json schema definition:

{json.dumps(input_schema_reflection, indent=2)}

You can use a web search tool that takes a 'search_query' as a parameter.
Your job is to reflect on the current state of the paragraph text and think if you haven't missed some critical aspect of the topic and provide the most optimal web search query to enrich the latest state.
Format the output in json with the following json schema definition:

{json.dumps(output_schema_reflection, indent=2)}

Make sure that the output is a json object with an output json schema defined above.
Only return the json object, no explanation or additional text.
"""

输入数据样例：

response = client.chat.completions.create(
model="DeepSeek-R1",
messages=[{"role":"system","content": SYSTEM_PROMPT_REFLECTION},\
{"role":"user","content":json.dumps(input_data)}],
temperature=1
)

print(remove_reasoning_from_output(response.choices[0].message.content))

输出样例：

现在我们可以运行查询，将新结果追加到段落状态中，并结合新结果与段落的最新状态。反思步骤通过重新评估已有信息，帮助我们发现可能遗漏的关键点，从而提升研究的全面性和准确性。
实现步骤七：用反思搜索结果丰富段落状态
执行反思步骤的搜索查询后：

我们可以用以下方式更新段落的搜索状态：

我们将重复执行步骤六和步骤七指定次数，以确保内容的深度和广度。
实现步骤八：总结并生成报告
对步骤二中规划的每个段落执行步骤四到步骤七。一旦所有段落的最终状态准备就绪，我们就可以将所有内容整合在一起，使用LLM生成格式优美的Markdown文档。如果结论段落不存在，则从其他段落的最新状态中添加到报告的末尾。以下是提示词：

SYSTEM_PROMPT_REPORT_FORMATTING = f"""
You are a Deep Research assistant. You have already performed the research and constructed final versions of all paragraphs in the report.
You will get the data in the following json format:

{json.dumps(input_schema_report_formatting, indent=2)}

Your job is to format the Report nicely and return it in Markdown.
If a Conclusion paragraph is not present, add it to the end of the report from the latest state of the other paragraphs.
"""

执行以下代码：

response = client.chat.completions.create(
model="DeepSeek-R1",
messages=[{"role":"system","content": SYSTEM_PROMPT_REPORT_FORMATTING},\
{"role":"user","content":json.dumps(report_data)}],
temperature=1
)