llmatch-messages

llmatch-messages is a robust utility for pattern-verified LLM interactions built on top of langchain-llm7 and langchain-core. It enables structured, retryable conversations with LLMs by matching regex patterns in their responses, ideal for use cases where consistency and format are crucial — such as parsing XML/JSON-style tags, code blocks, or key phrases.

This package is production-oriented, supports retries with exponential backoff, and provides rich diagnostics when responses do not match expectations.

🔧 Installation

pip install llmatch-messages

This will also install:

• langchain-llm7==2025.05.91116
• langchain-core==0.3.51

✨ Example

from llmatch_messages import llmatch
from langchain_llm7 import ChatLLM7
from langchain_core.messages import HumanMessage, SystemMessage

llm = ChatLLM7()

response = llmatch(
    llm=llm,
    messages=[
        SystemMessage(content="You are a helpful assistant. Write the output in the format: <image_desc>...</image_desc>"),
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image:"},
                {"type": "image_url", "image_url": {"url": "https://llm7.io/logo.png"}},
            ],
        )
    ],
    pattern=r"<image_desc>\s*(.*?)\s*</image_desc>",
    verbose=True,
)

if response["success"]:
    print("Extracted:", response["extracted_data"])
else:
    print("Error:", response["error_message"])

🧠 Key Features

• Pattern Matching: Use regex to validate and extract structured parts of LLM responses.
• Retry Logic: Automatically retries if output does not conform, using exponential backoff.
• LangChain Native: Works seamlessly with langchain_core.messages and langchain-llm7.
• Message-Aware: Operates on BaseMessage list (e.g. HumanMessage, SystemMessage).
• Detailed Diagnostics: Verbose mode traces all steps and decision points.
• Fail-Safe: Handles malformed LLM responses gracefully and provides fallback messaging.

🧪 Return Format

The function returns a dictionary:

{
    "success": True | False,
    "extracted_data": Optional[List[str]],
    "final_content": Optional[str],
    "retries_attempted": int,
    "error_message": Optional[str],
    "raw_response": Optional[Any],
}

🪄 Common Use Cases

• Validate XML-like or markdown-formatted LLM outputs.
• Parse code blocks (```...```), JSON sections, or tags.
• Build structured LLM workflows that require machine-readable responses.

📄 License

Licensed under the Apache License 2.0.

🙋‍♂️ Author

Developed by Eugene Evstafev, software developer at the University of Cambridge, creator of llm7.io.

For feedback or contributions, feel free to open an issue or PR on GitHub.