Jina AI × DeepAgents
网络搜索工具 & 文章撰写子 Agent

@tool
def jina_search(
    query: Annotated[str, "Search query — describe what you want to find"],
    num_results: Annotated[int, "Number of results (1-10, default 5)"] = 5,
) -> str:
    """Search the web using Jina AI (s.jina.ai).
    Returns top results as LLM-friendly markdown with title, URL, and content.
    Use for: current events, factual lookup, finding primary sources.
    """
    headers = {
        "Authorization": f"Bearer {JINA_API_KEY}",
        "Accept":         "application/json",
        "X-With-Links-Summary": "true",
        "X-Return-Format":      "markdown",
    }
    payload = {"q": query, "num": max(1, min(10, num_results))}

    resp = requests.post(_SEARCH_URL, headers=headers, json=payload, timeout=30)
    resp.raise_for_status()
    items = resp.json().get("data", [])

    parts = []
    for i, item in enumerate(items, 1):
        content = (item.get("content") or "").strip()
        content = content[:2000] + ("…" if len(content) > 2000 else "")
        parts.append(f"[{i}] **{item.get('title')}**\nURL: {item.get('url')}\n\n{content}")

    return "\n\n---\n\n".join(parts) if parts else "No results found."

@tool
def jina_reader(
    url: Annotated[str, "Full URL to read (must start with http/https)"],
    max_chars: Annotated[int, "Max chars to return (default 6000)"] = 6000,
) -> str:
    """Read and extract clean Markdown from a URL using Jina AI (r.jina.ai).
    Use for: reading a specific article in full; following up on search results.
    """
    if not url.startswith("http"):
        return "[jina_reader error] URL must start with http:// or https://"

    headers = {
        "Authorization": f"Bearer {JINA_API_KEY}",
        "Accept":          "application/json",
        "X-Return-Format": "markdown",
    }
    resp = requests.get(f"{_READER_URL}{url}", headers=headers, timeout=30)
    resp.raise_for_status()

    payload = resp.json().get("data", {})
    title   = payload.get("title", url)
    content = (payload.get("content") or "").strip()

    truncated = content[:max_chars] + ("…[truncated]" if len(content) > max_chars else "")
    return f"# {title}\nSource: {url}\n\n{truncated}"

with requests.post(_DEEPSEARCH_URL, headers=headers, json=payload,
                    stream=True, timeout=300) as resp:
    for raw_line in resp.iter_lines():
        line = raw_line.decode("utf-8") if isinstance(raw_line, bytes) else raw_line
        if not line.startswith("data: "): continue

        data_str = line[6:].strip()
        if data_str == "[DONE]": break

        chunk  = json.loads(data_str)
        delta  = chunk["choices"][0]["delta"]

        # 1. 累积正文（含 <think> 推理块）
        full_content += delta.get("content") or ""

        # 2. 收集 inline citations
        for ann in delta.get("annotations") or []:
            if ann.get("type") == "url_citation":
                citations.append(ann.get("url_citation", {}))

# 后处理
answer = _strip_thinking(full_content)   # 移除 <think>...</think>
answer += _format_sources(citations)     # 去重 + 追加 Sources 节

jina_writer_subagent = {
    "name": "jina-writer",
    "description": (
        "Writes comprehensive, citation-backed articles and research reports "
        "using Jina AI DeepSearch for up-to-date web research. "
        "Use for: long-form articles, market research, topic overviews, "
        "technical summaries requiring current sources. "
        "Pass: article topic, desired length/sections, target audience."
    ),
    "system_prompt": JINA_WRITER_INSTRUCTIONS,   # 见下方
    "tools": [jina_deepsearch, jina_search, jina_reader],
    # "model": "anthropic:claude-haiku-4-5-20251001",  # 可选降本
}

## Role
Given a writing request, you:
1. Research the topic thoroughly using jina_deepsearch
2. Optionally deepen specific angles with jina_search + jina_reader
3. Synthesize all findings into a well-structured, citation-backed piece

## Tools Available
- jina_deepsearch(query, reasoning_effort)  ← primary: deep research
- jina_search(query, num_results)           ← supplement: targeted search
- jina_reader(url)                          ← supplement: read specific source

## Writing Process
Step 1: Understand scope, sections, audience, tone
Step 2: Call jina_deepsearch (high for >1000 words, medium for standard)
Step 3: jina_search + jina_reader for sub-topic depth if needed
Step 4: Write article — synthesize, do NOT dump raw DeepSearch output

## Output Structure (Markdown)
# [Title]
## Introduction / Executive Summary
## [Section 1…N]  ← with inline citations [1][2]
## Conclusion
## References
[1] Title — URL
...

## Constraints
- Every fact must carry [x] citation
- Consolidate all URLs in References
- No fabrication — only cite tool-returned sources

### F. Report Writing and Structured Analysis
...
  1. Pre-collect context: retrieve_from_kb + list_business_contexts
  2. Route by evidence availability:
     - KB evidence sufficient → task(subagent_type="research-agent")
       # 返回 Evidence Package，主 agent 合成报告
     - Needs current web data → task(subagent_type="jina-writer")
       # 返回完整文章，主 agent 直接整合或转述
     - Both needed → sequential: research-agent first, then jina-writer
       # 内部 KB 证据 + 网络信息双来源合并

from jina_tools import (
    jina_search, jina_reader, jina_deepsearch,
    jina_writer_subagent,
)

agent = create_deep_agent(
    model=init_chat_model("anthropic:claude-sonnet-4-6", temperature=0.0),
    tools=[
        retrieve_from_kb,
        list_business_contexts,
        get_doc_context_by_id,
        think_tool,
        jina_search,    # ← 新增：主 agent 也可直接搜索
        jina_reader,    # ← 新增：主 agent 可读取指定 URL
    ],
    system_prompt=KB_WORKFLOW_INSTRUCTIONS_RESEARCH,
    subagents=[
        research_sub_agent,     # KB 内部证据采集
        jina_writer_subagent,   # ← 新增：网络研究 + 文章撰写
        mermaid_expert_subagent,
        file_catalog_subagent,
    ],
)

# .env
JINA_API_KEY=jina_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   # 从 jina.ai 获取

# 获取免费 API Key（10M tokens 免费额度）
# https://jina.ai/?sui=apikey