Skill

Google 出品的 Python 库，用 LLM 从非结构化文本中提取结构化信息，每个提取结果精确对应源文本位置，支持交互式可视化，适配 Gemini/OpenAI/Ollama 等多种模型

Security Scan

The skill's instructions are broadly coherent with a text-extraction library, but there are several mismatches and operational recommendations (API keys, remote install scripts, filesystem access) that the registry metadata does not declare — review before trusting or auto-running.

Purpose & Capability

The name/description (an LLM-driven extractor supporting Gemini/OpenAI/Ollama) matches the instructions and examples which show calling lx.extract(), saving JSONL, and generating HTML visualizations. However, the registry metadata declared no required environment variables or credentials while the SKILL.md and guides repeatedly instruct setting provider keys (GOOGLE_API_KEY, OPENAI_API_KEY) and configuring Ollama — a discrepancy between declared requirements and actual operational needs.

Instruction Scope

The SKILL.md and guides tell the agent to install the package, call lx.extract() on raw text, document URLs, and local files, and read/write JSONL/HTML files. Those actions are consistent with a text-extraction tool. Flagged items: instructions instruct the agent (or user) to configure API keys and to run remote installers (curl … | sh for Ollama) and to access local file paths (papers/*.txt). Accessing local files and downloading documents is expected for this purpose but is broad; if processing sensitive clinical/legal texts, the user should confirm data-handling policies and where model inference runs (remote provider vs local).

Install Mechanism

There is no formal install spec in the registry (instruction-only), which is lowest static risk. The docs recommend pip install langextract (standard) and, optionally, running a remote install script for Ollama via curl -fsSL https://ollama.com/install.sh | sh. The pip recommendation is expected; the curl | sh instruction pulls and runs a remote shell script (from an official-looking domain) — a common but higher-risk operation that the user/agent should not run blindly without verifying source and integrity.

Credentials

Although the registry lists no required env vars, the documentation explicitly requires/provider-key usage (GOOGLE_API_KEY for Gemini, OPENAI_API_KEY for OpenAI) and suggests writing them to .env or setting them in the environment. Requesting provider API keys is proportionate for a model-backed extractor, but the absence of these variables in the declared requirements is an incoherence to surface: an agent or installer might not prompt for, or be given, the expected keys. Also note guidance for Colab-specific userdata access which may tie into platform-specific secrets — ensure keys are handled securely and not auto-exfiltrated.

Persistence & Privilege

The skill does not request elevated privileges or permanent inclusion (always:false). It does not declare modifying other skills or system-wide agent settings. No persistence or privileged behavior is visible in the docs.

What to consider before installing

This skill appears to be a legitimate LLM-based text-extraction library, but review these before installing or auto-running it: - Verify the package/source: the docs cite a GitHub repo and PyPI pages; confirm they exist and are official (check repo owner, release tags, and package metadata) before pip installing. - Expect to provide API keys: the guides require GOOGLE_API_KEY and/or OPENAI_API_KEY (or a local Ollama install). The registry did not declare these env vars — supply keys only if you trust the provider and understand where inference runs (remote cloud vs local). - Do not run curl | sh blindly: the docs recommend installing Ollama via a remote installer script. If you plan to follow that path, inspect the script and verify its integrity first. - Be cautious with sensitive data: the skill targets clinical and legal texts. If you will process PHI or other sensitive content, prefer local models (Ollama) or verify contractual/data-protection arrangements with cloud model providers. - If you want higher confidence: ask the skill publisher for an explicit declared requires.env list, a reproducible install spec, and signed release artifacts (PyPI/GitHub release) or provide the exact package fingerprint you will install. Given the metadata/instruction mismatch (missing declared env vars) and the recommendations to run remote installers and access local files, treat this skill as requiring manual review before granting it full automation privileges.

LangExtract — LLM 驱动的结构化信息提取

LangExtract 是 Google 出品的 Python 库，利用 LLM 从临床笔记、法律文书、新闻报道等非结构化文本中提取结构化信息。其核心优势是精确源文本定位：每个提取结果都映射到源文本的精确字符位置，支持可视化高亮审查，适合需要可追溯性的企业级数据提取场景。只需几个示例即可定义提取任务，无需微调模型。

核心使用场景

• 医疗文本提取：从临床笔记提取药物、剂量、诊断结果，含可追溯的源文本定位
• 法律文书解析：提取合同条款、当事方信息、关键日期等结构化字段
• 文学/学术分析：提取人物、情感、关系等实体（如 Romeo and Juliet 示例）
• 长文档批量处理：分块并行处理大型文档，支持多次 Pass 提高召回率
• 交互式审查：生成自包含 HTML 可视化文件，高亮显示提取结果与源文本的对应关系

AI 辅助使用流程

1. 安装依赖 — AI 执行 pip install langextract 并配置 API 密钥
2. 定义提取任务 — AI 根据需求编写提取描述（prompt）和高质量示例（examples）
3. 运行提取 — AI 调用 lx.extract() 处理文本或文档 URL
4. 保存结果 — AI 将结果存为 JSONL 格式，便于后续分析
5. 生成可视化 — AI 调用 lx.visualize() 生成交互式 HTML 审查界面
6. 过滤非定位提取 — AI 过滤掉 char_interval=None 的结果，保留有源文本依据的提取

关键章节导航

• 安装指南 — pip 安装、API 密钥配置、Ollama 本地模型
• 快速开始 — 定义提取任务、运行提取、结果过滤、可视化
• 高级用法 — 长文档处理、自定义模型、医疗/法律场景
• 故障排查 — API 认证、Prompt 对齐警告、召回率低

AI 助手能力

使用本技能时，AI 可以：

• ✅ 安装 LangExtract 并配置 Gemini/OpenAI API 密钥
• ✅ 根据用户描述的提取需求，自动编写 prompt_description 和 examples
• ✅ 调用 lx.extract() 处理文本字符串或文档 URL
• ✅ 过滤非定位（char_interval=None）的提取结果
• ✅ 保存提取结果为 JSONL 格式
• ✅ 生成交互式 HTML 可视化文件
• ✅ 配置 Ollama 本地模型，避免 API 费用
• ✅ 使用并行处理和分块策略处理长文档

核心功能

• ✅ 精确源文本定位 — 每个提取结果映射到源文本的 char_interval（字符偏移）
• ✅ 结构化输出 — 基于 Few-shot 示例强制输出一致 Schema
• ✅ 长文档优化 — 文本分块 + 并行处理 + 多次 Pass，提高大文档召回率
• ✅ 交互式可视化 — 自包含 HTML 文件，高亮显示数千个提取结果
• ✅ 多模型支持 — Gemini（推荐）、OpenAI、Ollama 本地模型
• ✅ 域无关 — 用 Few-shot 示例适配任意领域，无需微调
• ✅ 自动对齐检测 — 检测示例与文本不匹配的情况并发出警告
• ✅ URL 直接处理 — 可直接传入文档 URL（如 Gutenberg 文本）

快速示例

import langextract as lx

# 定义提取任务
result = lx.extract(
    text_or_documents="John took aspirin 500mg twice daily for headache.",
    prompt_description="Extract medications and dosages.",
    examples=[
        lx.data.ExampleData(
            text="Patient was prescribed ibuprofen 400mg.",
            extractions=[
                lx.data.Extraction(
                    extraction_class="medication",
                    extraction_text="ibuprofen 400mg",
                    attributes={"dosage": "400mg"}
                )
            ]
        )
    ],
    model_id="gemini-2.5-flash",
)

# 只保留有源文本依据的结果
grounded = [e for e in result.extractions if e.char_interval]
print(grounded)

# 生成可视化
lx.io.save_annotated_documents([result], output_name="out.jsonl", output_dir=".")
html = lx.visualize("out.jsonl")

安装要求

依赖	版本要求
Python	>= 3.9
Gemini API Key	推荐（或 OpenAI/Ollama）
pip	任意版本

项目链接

• GitHub：https://github.com/google/langextract
• PyPI：https://pypi.org/project/langextract/
• 论文 DOI：https://doi.org/10.5281/zenodo.17015089