backend/core/utils.py

"""Utility functions for backend processing."""

import html
import re


def clean_model_output(content: str) -> str:
    """
    Clean LLM output by removing reasoning artifacts.

    Removes:
    1. <think>...</think> tags from DeepSeek models
    2. HTML entities (&lt;think&gt;)
    3. "Reasoning:" headers
    4. Malformed/cut-off tags

    Args:
        content: Raw LLM response string

    Returns:
        Cleaned response string
    """
    if not content:
        return ""

    # 1. Unescape HTML entities
    content = html.unescape(content)

    # 2. Remove <think> blocks (case-insensitive, DOTALL for multiline)
    content = re.sub(
        r'<think>.*?</think>',
        '',
        content,
        flags=re.DOTALL | re.IGNORECASE
    )

    # 3. Handle cut-off/malformed tags (content after </think>)
    if '</think>' in content:
        content = content.split('</think>')[-1]

    # 4. Remove "Reasoning:" headers
    if content.strip().startswith("Reasoning:"):
        parts = content.split("\n\n", 1)
        if len(parts) > 1:
            content = parts[1]

    return content.strip()