llm-securescan 2025.12.21142626

A new package is designed to facilitate secure and structured user interactions with language models by analyzing and extracting specific information from user-provided text. It uses pattern matching

llm-securescan

llm-securescan is a tiny Python package that helps you securely extract structured information from user‑provided text.
It uses pattern matching and automatic retries to ensure the language model returns data that conforms to a predefined regex. This makes it useful for detecting potential data‑exfiltration patterns, sensitive data leaks, or any other custom signatures you define.

---

Features

• Zero‑setup default LLM – uses ChatLLM7 (from the langchain_llm7 package) automatically.
• Pluggable LLM – pass any LangChain‑compatible chat model (OpenAI, Anthropic, Google, etc.).
• Regex‑based extraction – the LLM is forced to obey a regex pattern, guaranteeing consistent output.
• Simple API – one function call returns a list of extracted strings or raises an informative error.

---

Installation

pip install llm_securescan

---

Quick Start

from llm_securescan import llm_securescan

user_input = """
John Doe's credit card number is 4111 1111 1111 1111.
Please send the PDF to [e-mail].
"""

# Use the default ChatLLM7 (API key taken from env var LLM7_API_KEY or default)
extracted = llm_securescan(user_input)

print(extracted)   # → ['4111 1111 1111 1111', '[e-mail]']

---

API Reference

llm_securescan(user_input: str, llm: Optional[BaseChatModel] = None, api_key: Optional[str] = None) -> List[str]

Parameter	Type	Description
user_input	str	The raw text you want to scan.
llm	Optional[BaseChatModel]	A LangChain chat model instance. If omitted, the function creates a ChatLLM7 instance automatically.
api_key	Optional[str]	API key for ChatLLM7. If not provided, the function looks for the environment variable LLM7_API_KEY. If that also isn’t set, a placeholder value "None" is used (which will cause the LLM call to fail with a clear error).

Returns: A list of strings that match the configured regex pattern. If the LLM call fails, a RuntimeError is raised with the underlying error message.

---

Using a Custom LLM

You can pass any LangChain‑compatible chat model. Below are examples for the most common providers.

OpenAI

from langchain_openai import ChatOpenAI
from llm_securescan import llm_securescan

my_llm = ChatOpenAI(model="gpt-4o-mini")
result = llm_securescan(user_input, llm=my_llm)

Anthropic

from langchain_anthropic import ChatAnthropic
from llm_securescan import llm_securescan

my_llm = ChatAnthropic(model="claude-3-haiku-20240307")
result = llm_securescan(user_input, llm=my_llm)

Google Generative AI

from langchain_google_genai import ChatGoogleGenerativeAI
from llm_securescan import llm_securescan

my_llm = ChatGoogleGenerativeAI(model="gemini-1.5-flash")
result = llm_securescan(user_input, llm=my_llm)

---

Configuration Details

• Default LLM: ChatLLM7 from the langchain_llm7 package (see https://pypi.org/project/langchain-llm7/).

• Rate limits: The free tier of LLM7 provides sufficient calls for typical scanning workloads.

• Custom API key: Provide an API key either via the LLM7_API_KEY environment variable or directly:

  result = llm_securescan(user_input, api_key="YOUR_LLM7_API_KEY")
  

• Getting a free API key: Register at https://token.llm7.io/.

---

How It Works Internally

1. The function builds a system prompt and a human prompt based on the supplied text.
2. A regular expression (pattern from llm_securescan.prompts) is compiled.
3. llmatch (from the internal llmatch_messages utility) sends the prompts to the LLM while enforcing that the response matches the regex.
4. If the LLM output satisfies the pattern, the captured groups are returned; otherwise, the call fails with a helpful error message.

This approach provides a deterministic extraction pipeline while still leveraging the LLM’s natural‑language understanding.

---

Contributing & Issues

Found a bug or have a feature request? Please open an issue on GitHub:

https://github.com/chigwell/llm_securescan/issues

Pull requests are welcome!

---

Author

Eugene Evstafev
✉️ Email: [e-mail]
🐙 GitHub: chigwell

---

License

This project is licensed under the MIT License. See the LICENSE file for details.