Local Researcher

完全本地的深度研究助手 Skill。使用 Ollama 或 LMStudio 本地 LLM 进行迭代式网络研究，生成带引用来源的 Markdown 报告。当用户需要进行隐私优先的研究、本地文档分析或生成结构化研究报告时触发。

Security Scan

The skill's metadata claims no requirements, but the runtime instructions expect local LLM runtimes (Ollama/LMStudio), environment variables, and optional third‑party search API keys — these mismatches and privacy claims warrant caution.

Purpose & Capability

The skill claims to be a fully local research assistant, but the package metadata lists no required binaries or env vars while the SKILL.md and quickstart script explicitly assume Ollama or LMStudio installed and configured. Requiring a local LLM runtime and optional external search providers is coherent with the stated purpose, but the registry metadata omits those requirements, which is an inconsistency.

Instruction Scope

SKILL.md instructs cloning a GitHub repo, installing dependencies, configuring .env, running local LLM endpoints, and performing web searches. It also recommends optional external search services (Tavily, Perplexity) that would contact third‑party servers. The 'all data stays local' privacy claim is therefore conditional and misleading unless the user explicitly avoids external search APIs.

Install Mechanism

There is no formal install spec in the registry (instruction-only). The README suggests installing Ollama via a curl | sh installer and pip install -e ., which will pull packages from the network. No embedded or obfuscated installers are present in the skill bundle itself, but the user will run network installs at setup time.

Credentials

Registry metadata declares no required environment variables, yet SKILL.md expects many optional/required env settings (LLM_PROVIDER, OLLAMA_BASE_URL, LOCAL_LLM, LMSTUDIO_BASE_URL, SEARCH_API and API keys such as TAVILY_API_KEY/PERPLEXITY_API_KEY). This mismatch is problematic: secrets may be needed depending on chosen search provider, and the metadata does not surface that.

Persistence & Privilege

The skill is not always-enabled and does not request elevated or persistent platform privileges. The included quickstart script only inspects local repo state and calls the 'ollama' binary; it does not attempt to modify other skills or global agent configuration.

What to consider before installing

This skill appears to implement what it claims (local research with local LLMs) but the registry metadata understates its requirements. Before installing: 1) Verify the source repository (the SKILL.md points to https://github.com/langchain-ai/local-deep-researcher.git) and review its code and setup files; 2) Expect to install Ollama or LMStudio and to run models locally — verify you trust those installers (curl | sh); 3) Treat any .env or API key values as sensitive; do not supply third‑party search API keys (Tavily/Perplexity) if you require strict privacy — use DuckDuckGo or a self‑hosted SearXNG; 4) Run initial setup in an isolated environment (VM/container) and inspect network traffic if you want to confirm data remains local; 5) If you rely on the registry metadata to decide safety, ask the publisher/maintainer to correct the declared requirements before proceeding.

Local Researcher Skill

完全在本地运行的深度研究助手，无需将数据发送到云端 LLM 服务。支持 Ollama 和 LMStudio，迭代式网络研究，输出带引用的专业报告。

前置要求

安装 Ollama（推荐）

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# 拉取模型
ollama pull deepseek-r1:8b
ollama pull llama3.2
ollama pull qwen:14b

或使用 LMStudio

1. 下载 LMStudio
2. 下载并加载模型（如 qwen_qwq-32b）
3. 进入 "Local Server" 标签页
4. 启动 OpenAI 兼容 API 服务
5. 记下服务地址（默认: http://localhost:1234/v1）

安装本 Skill

# 克隆仓库
git clone https://github.com/langchain-ai/local-deep-researcher.git
cd local-deep-researcher

# 创建虚拟环境
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

# 安装依赖
pip install -e .

配置

复制环境变量模板并编辑：

cp .env.example .env

编辑 .env 文件：

# LLM 提供商选择
LLM_PROVIDER=ollama
# LLM_PROVIDER=lmstudio

# Ollama 配置
OLLAMA_BASE_URL=http://localhost:11434
LOCAL_LLM=deepseek-r1:8b

# LMStudio 配置
# LMSTUDIO_BASE_URL=http://localhost:1234/v1
# LOCAL_LLM=qwen_qwq-32b

# 搜索工具配置
SEARCH_API=duckduckgo  # 默认，无需 API key
# SEARCH_API=tavily
# TAVILY_API_KEY=tvly-xxx
# SEARCH_API=perplexity
# PERPLEXITY_API_KEY=pplx-xxx

# 研究循环次数
MAX_WEB_RESEARCH_LOOPS=3

# 是否获取完整页面内容
FETCH_FULL_PAGE=true

使用方法

快速开始

# 启动研究（交互模式）
python src/ollama_deep_researcher/main.py

# 或使用 LangGraph CLI
langgraph dev

程序化使用

from langgraph.graph import StateGraph
from ollama_deep_researcher.graph import graph

# 定义研究主题
topic = "量子计算在药物发现中的应用"

# 配置参数
config = {
    "llm_provider": "ollama",
    "local_llm": "deepseek-r1:8b",
    "search_api": "duckduckgo",
    "max_web_research_loops": 3,
    "fetch_full_page": True
}

# 运行研究
result = graph.invoke(
    {"topic": topic},
    config=config
)

# 输出报告
print(result["final_summary"])

核心功能

1. 迭代式深度研究

系统自动执行以下循环：

1. 根据主题生成搜索查询
2. 执行网络搜索
3. 总结搜索结果
4. 反思总结，识别知识缺口
5. 生成新查询填补缺口
6. 重复直到达到最大循环次数

2. 多搜索源支持

搜索源	需要 API Key	特点
DuckDuckGo	❌ 不需要	默认选项，隐私友好
Tavily	✅ 需要	高质量搜索结果
Perplexity	✅ 需要	AI 增强搜索
SearXNG	❌ 不需要	自托管选项

3. 输出格式

最终输出为 Markdown 格式报告，包含：

• 执行摘要
• 详细研究发现
• 所有引用的来源链接
• 研究过程元数据

4. LangGraph Studio 可视化

# 安装 LangGraph CLI
pip install "langgraph-cli[inmem]"

# 启动开发服务器
langgraph dev

打开浏览器访问 Studio UI，可实时观察研究流程：

• 搜索查询生成
• 来源收集
• 总结迭代
• 最终报告生成

完整工作流示例

学术论文预研

import asyncio
from ollama_deep_researcher.graph import graph

async def research_paper_prep():
    topic = "Transformer 架构在生物信息学中的最新应用"
    
    config = {
        "llm_provider": "ollama",
        "local_llm": "deepseek-r1:14b",  # 使用更大模型获得更好结果
        "search_api": "duckduckgo",
        "max_web_research_loops": 5,  # 更多轮次深入挖掘
        "fetch_full_page": True
    }
    
    result = await graph.ainvoke(
        {"topic": topic},
        config=config
    )
    
    # 保存报告
    with open("literature_review.md", "w") as f:
        f.write(result["final_summary"])
    
    # 打印引用的来源
    print("参考来源:")
    for source in result.get("sources", []):
        print(f"- {source}")
    
    return result

# 运行
result = asyncio.run(research_paper_prep())

市场调研报告

def market_research(product_category: str):
    """生成市场调研报告"""
    
    topic = f"{product_category} 市场规模、主要竞争者和发展趋势 2024"
    
    config = {
        "llm_provider": "ollama",
        "local_llm": "qwen:14b",
        "search_api": "tavily",  # 使用 Tavily 获得更商业化的结果
        "tavily_api_key": "tvly-xxx",
        "max_web_research_loops": 4,
        "fetch_full_page": True
    }
    
    result = graph.invoke(
        {"topic": topic},
        config=config
    )
    
    return result["final_summary"]

# 生成报告
report = market_research("新能源汽车")
print(report)

技术趋势追踪

# 使用命令行快速研究
cd local-deep-researcher
source .venv/bin/activate

# 创建研究脚本
python -c "
from ollama_deep_researcher.graph import graph

result = graph.invoke(
    {'topic': 'Rust 语言在系统编程领域的最新发展'},
    config={
        'llm_provider': 'ollama',
        'local_llm': 'llama3.2',
        'search_api': 'duckduckgo',
        'max_web_research_loops': 3
    }
)

print(result['final_summary'])
"

Docker 部署

# 构建镜像
docker build -t local-researcher .

# 运行容器
docker run --rm -it -p 2024:2024 \
  -e SEARCH_API=tavily \
  -e TAVILY_API_KEY=tvly-xxx \
  -e LLM_PROVIDER=ollama \
  -e OLLAMA_BASE_URL=http://host.docker.internal:11434/ \
  -e LOCAL_LLM=llama3.2 \
  local-researcher

注意：Ollama 需要在宿主机单独运行，容器通过 host.docker.internal 访问。

模型兼容性说明

模型	JSON 模式支持	备注
llama3.2	✅	推荐，轻量快速
deepseek-r1:8b	✅	推理能力强
qwen:14b	✅	中文表现好
gpt-oss	⚠️	需要启用 tool calling

gpt-oss 模型特殊配置：

# gpt-oss 不支持 JSON 模式，需要启用 tool calling
USE_TOOL_CALLING=true

故障排查

Ollama 连接问题

# 检查 Ollama 服务状态
curl http://localhost:11434/api/tags

# 确保模型已下载
ollama list

# 测试模型
ollama run llama3.2 "Hello"

搜索结果为空

• 检查网络连接
• 尝试更换搜索 API
• 调整搜索查询语言

生成质量不佳

• 使用更大的模型（如 14B 参数以上）
• 增加 MAX_WEB_RESEARCH_LOOPS
• 启用 FETCH_FULL_PAGE 获取更完整内容

隐私与安全

• ✅ 所有数据留在本地
• ✅ 无需联网到 OpenAI/Claude
• ✅ 搜索查询不关联个人身份
• ✅ 适合处理敏感商业/研究数据

高级用法

自定义研究流程

from langgraph.graph import StateGraph
from ollama_deep_researcher.configuration import Configuration
from ollama_deep_researcher import research_node, reflect_node

# 创建自定义流程
builder = StateGraph(ResearchState)
builder.add_node("research", research_node)
builder.add_node("reflect", reflect_node)
# ... 添加更多节点

# 编译并运行
graph = builder.compile()

集成到其他应用

# FastAPI 示例
from fastapi import FastAPI
from ollama_deep_researcher.graph import graph

app = FastAPI()

@app.post("/research")
async def create_research(topic: str):
    result = await graph.ainvoke(
        {"topic": topic},
        config={"max_web_research_loops": 3}
    )
    return {
        "topic": topic,
        "report": result["final_summary"],
        "sources": result.get("sources", [])
    }

相关资源

• LangChain Academy - 部署教程
• LangGraph 文档
• Ollama 模型库