backend/open_webui/utils/misc.py

from __future__ import annotations

import collections.abc
import hashlib
import json
import logging
import re
import threading
import time
import uuid
from datetime import timedelta
from pathlib import Path
from typing import Callable, Optional, Sequence, Union

import aiohttp
import mimeparse
from open_webui.env import CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE

log = logging.getLogger(__name__)


def deep_update(d, u):
    for k, v in u.items():
        if isinstance(v, collections.abc.Mapping):
            d[k] = deep_update(d.get(k, {}), v)
        else:
            d[k] = v
    return d


def get_allow_block_lists(filter_list):
    allow_list = []
    block_list = []

    if filter_list:
        for d in filter_list:
            if d.startswith('!'):
                # Domains starting with "!" → blocked
                block_list.append(d[1:].strip())
            else:
                # Domains starting without "!" → allowed
                allow_list.append(d.strip())

    return allow_list, block_list


def is_string_allowed(string: Union[str, Sequence[str]], filter_list: list[str | None] = None) -> bool:
    """
    Checks if a string is allowed based on the provided filter list.
    :param string: The string or sequence of strings to check (e.g., domain or hostname).
    :param filter_list: List of allowed/blocked strings. Strings starting with "!" are blocked.
    :return: True if the string or sequence of strings is allowed, False otherwise.
    """
    if not filter_list:
        return True

    allow_list, block_list = get_allow_block_lists(filter_list)
    strings = [string] if isinstance(string, str) else list(string)

    # If allow list is non-empty, require domain to match one of them
    if allow_list:
        if not any(s.endswith(allowed) for s in strings for allowed in allow_list):
            return False

    # Block list always removes matches
    if any(s.endswith(blocked) for s in strings for blocked in block_list):
        return False

    return True


def get_message_list(messages_map, message_id):
    """
    Reconstructs a list of messages in order up to the specified message_id.

    :param message_id: ID of the message to reconstruct the chain
    :param messages: Message history dict containing all messages
    :return: List of ordered messages starting from the root to the given message
    """

    # Handle case where messages is None
    if not messages_map:
        return []  # Return empty list instead of None to prevent iteration errors

    # Find the message by its id
    current_message = messages_map.get(message_id)

    if not current_message:
        return []  # Return empty list instead of None to prevent iteration errors

    # Reconstruct the chain by following the parentId links
    message_list = []
    visited_message_ids = set()

    while current_message:
        message_id = current_message.get('id')
        if message_id in visited_message_ids:
            # Cycle detected, break to prevent infinite loop
            break

        if message_id is not None:
            visited_message_ids.add(message_id)

        message_list.append(current_message)
        parent_id = current_message.get('parentId')  # Use .get() for safety
        current_message = messages_map.get(parent_id) if parent_id else None

    message_list.reverse()
    return message_list


def get_messages_content(messages: list[dict]) -> str:
    return '\n'.join([f'{message["role"].upper()}: {get_content_from_message(message)}' for message in messages])


def get_last_user_message_item(messages: list[dict]) -> dict | None:
    for message in reversed(messages):
        if message['role'] == 'user':
            return message
    return None


def get_content_from_message(message: dict) -> str | None:
    if isinstance(message.get('content'), list):
        for item in message['content']:
            if item['type'] == 'text':
                return item['text']
    else:
        return message.get('content')
    return None


def reconcile_tool_pairs(messages: list[dict]) -> list[dict]:
    """Drop unpaired tool_use / tool_result from a reconstructed conversation.

    Stored output can be incomplete — a tool result may be missing (e.g. the
    knowledge base was updated mid-chat, or the call was interrupted), or a
    tool call may be missing while its result survived.  Strict providers
    (Anthropic, AWS Bedrock Converse) reject either direction of mismatch.

    Well-formed output is unaffected: every id pairs, so nothing is stripped.
    """
    completed_tool_call_ids = {
        message['tool_call_id'] for message in messages if message.get('role') == 'tool' and message.get('tool_call_id')
    }
    requested_tool_call_ids = {
        tool_call['id']
        for message in messages
        for tool_call in message.get('tool_calls') or ()
        if message.get('role') == 'assistant' and tool_call.get('id')
    }

    reconciled_messages = []
    for message in messages:
        role = message.get('role')

        # Orphan tool result — no assistant ever claimed this call_id.
        if role == 'tool' and message.get('tool_call_id') not in requested_tool_call_ids:
            continue

        # Non-assistant or no tool_calls — pass through unchanged.
        if role != 'assistant' or not message.get('tool_calls'):
            reconciled_messages.append(message)
            continue

        # Keep only tool_calls whose id received a tool-role response.
        valid_tool_calls = [
            tool_call for tool_call in message['tool_calls'] if tool_call.get('id') in completed_tool_call_ids
        ]

        if valid_tool_calls:
            reconciled_messages.append({**message, 'tool_calls': valid_tool_calls})
            continue

        # All tool_calls were orphans — keep the message only if it
        # carries meaningful text or reasoning content.
        content = message.get('content', '')
        has_meaningful_content = content.strip() if isinstance(content, str) else content
        if has_meaningful_content or message.get('reasoning_content'):
            reconciled_messages.append({key: value for key, value in message.items() if key != 'tool_calls'})

    return reconciled_messages


def convert_output_to_messages(
    output: list,
    raw: bool = False,
    reasoning_format: str | None = None,
) -> list[dict]:
    """
    Convert OR-aligned output items to OpenAI Chat Completion-format messages.

    This reconstructs the full conversation from the stored Responses API-native
    output items, including assistant messages with tool_calls arrays and tool
    role messages.

    Args:
        output: List of OR-aligned output items (Responses API format).
        raw: If True, include code interpreter blocks for LLM re-processing
             follow-ups.
        reasoning_format: How to include reasoning blocks in the output:
            - None: skip reasoning (default, safe for strict providers).
            - ``'think_tags'``: wrap in ``<think>`` tags inside content
              (for Ollama, which expects reasoning as tagged content).
            - ``'reasoning_content'``: set as ``reasoning_content`` top-level field
              (for llama.cpp, which routes it via the chat template).
    """
    if not output or not isinstance(output, list):
        return []

    messages = []
    pending_tool_calls = []
    pending_content = []
    pending_reasoning = []  # Only populated when reasoning_format == 'reasoning_content'

    def flush_pending():
        nonlocal pending_content, pending_tool_calls, pending_reasoning
        if not pending_content and not pending_tool_calls and not pending_reasoning:
            return

        message = {
            'role': 'assistant',
            'content': '\n'.join(pending_content) if pending_content else '',
            **({'tool_calls': pending_tool_calls} if pending_tool_calls else {}),
        }

        if pending_reasoning:
            message['reasoning_content'] = '\n'.join(pending_reasoning)

        messages.append(message)
        pending_content = []
        pending_tool_calls = []
        pending_reasoning = []

    for item in output:
        item_type = item.get('type', '')

        if item_type == 'message':
            # Extract text from output_text content parts
            content_parts = item.get('content', [])
            text = ''
            for part in content_parts:
                if part.get('type') == 'output_text':
                    text += part.get('text', '')
            if text:
                pending_content.append(text)

        elif item_type == 'function_call':
            # Collect tool calls to batch into assistant message
            arguments = item.get('arguments', '{}')
            # Ensure arguments is always a JSON string
            if not isinstance(arguments, str):
                arguments = json.dumps(arguments)
            pending_tool_calls.append(
                {
                    'id': item.get('call_id', ''),
                    'type': 'function',
                    'function': {
                        'name': item.get('name', ''),
                        'arguments': arguments,
                    },
                }
            )

        elif item_type == 'function_call_output':
            # Flush any pending content/tool_calls before adding tool result
            flush_pending()

            # Extract text and images from output content parts
            output_parts = item.get('output', [])
            content = ''
            image_urls = []
            for part in output_parts:
                if part.get('type') == 'input_text':
                    output_text = part.get('text', '')
                    content += str(output_text) if not isinstance(output_text, str) else output_text
                elif part.get('type') == 'input_image':
                    url = part.get('image_url', '')
                    if url:
                        image_urls.append(url)

            if image_urls:
                # Multimodal tool content with image(s)
                messages.append(
                    {
                        'role': 'tool',
                        'tool_call_id': item.get('call_id', ''),
                        'content': [
                            {'type': 'input_text', 'text': content},
                            *[{'type': 'input_image', 'image_url': url} for url in image_urls],
                        ],
                    }
                )
            else:
                messages.append(
                    {
                        'role': 'tool',
                        'tool_call_id': item.get('call_id', ''),
                        'content': content,
                    }
                )

        elif item_type == 'reasoning':
            if not reasoning_format:
                continue

            reasoning_text = ''
            source_list = item.get('summary', []) or item.get('content', [])
            for part in source_list:
                if part.get('type') == 'output_text':
                    reasoning_text += part.get('text', '')
                elif 'text' in part:
                    reasoning_text += part.get('text', '')

            if reasoning_text:
                if reasoning_format == 'think_tags':
                    # Ollama: embed in content with the item's original tags
                    start_tag = item.get('start_tag', '<think>')
                    end_tag = item.get('end_tag', '</think>')
                    pending_content.append(f'{start_tag}{reasoning_text}{end_tag}')
                elif reasoning_format == 'reasoning_content':
                    # llama.cpp: collect for reasoning_content field
                    pending_reasoning.append(reasoning_text)

        elif item_type == 'open_webui:code_interpreter':
            # Always include code interpreter content so the LLM knows
            # the code was already executed and doesn't retry.
            code = item.get('code', '')
            code_output = item.get('output', '')

            if code:
                pending_content.append(f'<code_interpreter>\n{code}\n</code_interpreter>')

            if code_output:
                if isinstance(code_output, dict):
                    stdout = code_output.get('stdout', '')
                    result = code_output.get('result', '')
                    output_text = stdout or result
                else:
                    output_text = str(code_output)
                if output_text:
                    pending_content.append(f'<code_interpreter_output>\n{output_text}\n</code_interpreter_output>')

        elif item_type.startswith('open_webui:'):
            # Skip other extension types
            pass

    # Flush remaining content/tool_calls
    flush_pending()

    return reconcile_tool_pairs(messages)


def get_last_user_message(messages: list[dict]) -> str | None:
    message = get_last_user_message_item(messages)
    if message is None:
        return None
    return get_content_from_message(message)


def set_last_user_message_content(content: str, messages: list[dict]) -> list[dict]:
    """
    Replace the text content of the last user message in-place.
    Handles both plain-string and list-of-parts content formats.
    """
    for message in reversed(messages):
        if message.get('role') == 'user':
            if isinstance(message.get('content'), list):
                for item in message['content']:
                    if item.get('type') == 'text':
                        item['text'] = content
                        break
            else:
                message['content'] = content
            break
    return messages


def get_last_assistant_message_item(messages: list[dict]) -> dict | None:
    for message in reversed(messages):
        if message['role'] == 'assistant':
            return message
    return None


def get_last_assistant_message(messages: list[dict]) -> str | None:
    for message in reversed(messages):
        if message['role'] == 'assistant':
            return get_content_from_message(message)
    return None


def get_system_message(messages: list[dict]) -> dict | None:
    for message in messages:
        if message['role'] == 'system':
            return message
    return None


def remove_system_message(messages: list[dict]) -> list[dict]:
    return [message for message in messages if message['role'] != 'system']


def pop_system_message(messages: list[dict]) -> tuple[dict | None, list[dict]]:
    return get_system_message(messages), remove_system_message(messages)


def merge_system_messages(messages: list[dict]) -> list[dict]:
    """
    Merge all system messages into one at position 0.

    Some chat templates (e.g. Qwen) require exactly one system
    message at the start.  Multiple pipeline stages may each
    insert their own system message; this function consolidates
    them.
    """
    system_contents: list[str] = []
    other_messages: list[dict] = []

    for message in messages:
        if message.get('role') == 'system':
            content = get_content_from_message(message)
            if content:
                system_contents.append(content)
        else:
            other_messages.append(message)

    if not system_contents:
        return other_messages

    merged = {'role': 'system', 'content': '\n'.join(system_contents)}
    return [merged, *other_messages]


def update_message_content(message: dict, content: str, append: bool = True) -> dict:
    if isinstance(message['content'], list):
        for item in message['content']:
            if item['type'] == 'text':
                if append:
                    item['text'] = f'{item["text"]}\n{content}'
                else:
                    item['text'] = f'{content}\n{item["text"]}'
    else:
        if append:
            message['content'] = f'{message["content"]}\n{content}'
        else:
            message['content'] = f'{content}\n{message["content"]}'
    return message


def replace_system_message_content(content: str, messages: list[dict]) -> dict:
    for message in messages:
        if message['role'] == 'system':
            message['content'] = content
            break
    return messages


def add_or_update_system_message(content: str, messages: list[dict], append: bool = False):
    """
    Adds a new system message at the beginning of the messages list
    or updates the existing system message at the beginning.

    :param msg: The message to be added or appended.
    :param messages: The list of message dictionaries.
    :return: The updated list of message dictionaries.
    """

    if messages and messages[0].get('role') == 'system':
        messages[0] = update_message_content(messages[0], content, append)
    else:
        # Insert at the beginning
        messages.insert(0, {'role': 'system', 'content': content})

    return messages


def add_or_update_user_message(content: str, messages: list[dict], append: bool = True):
    """
    Adds a new user message at the end of the messages list
    or updates the existing user message at the end.

    :param msg: The message to be added or appended.
    :param messages: The list of message dictionaries.
    :return: The updated list of message dictionaries.
    """

    if messages and messages[-1].get('role') == 'user':
        messages[-1] = update_message_content(messages[-1], content, append)
    else:
        # Insert at the end
        messages.append({'role': 'user', 'content': content})

    return messages


def prepend_to_first_user_message_content(content: str, messages: list[dict]) -> list[dict]:
    for message in messages:
        if message['role'] == 'user':
            message = update_message_content(message, content, append=False)
            break
    return messages


def append_or_update_assistant_message(content: str, messages: list[dict]):
    """
    Adds a new assistant message at the end of the messages list
    or updates the existing assistant message at the end.

    :param msg: The message to be added or appended.
    :param messages: The list of message dictionaries.
    :return: The updated list of message dictionaries.
    """

    if messages and messages[-1].get('role') == 'assistant':
        messages[-1]['content'] = f'{messages[-1]["content"]}\n{content}'
    else:
        # Insert at the end
        messages.append({'role': 'assistant', 'content': content})

    return messages


def strip_empty_content_blocks(messages: list[dict]) -> list[dict]:
    """
    Remove empty text content blocks from multimodal message content arrays.

    Providers like Gemini and Claude reject messages where a text block has
    an empty string.  This can happen when a user sends only file/image
    attachments without typing any text.
    """
    for message in messages:
        content = message.get('content')
        if isinstance(content, list):
            cleaned = [
                block
                for block in content
                if not (isinstance(block, dict) and block.get('type') == 'text' and not block.get('text', '').strip())
            ]
            if cleaned:
                message['content'] = cleaned
    return messages


def openai_chat_message_template(model: str):
    return {
        'id': f'{model}-{str(uuid.uuid4())}',
        'created': int(time.time()),
        'model': model,
        'choices': [{'index': 0, 'logprobs': None, 'finish_reason': None}],
    }


def openai_chat_chunk_message_template(
    model: str,
    content: str | None = None,
    reasoning_content: str | None = None,
    tool_calls: list[dict | None] = None,
    usage: dict | None = None,
) -> dict:
    template = openai_chat_message_template(model)
    template['object'] = 'chat.completion.chunk'

    template['choices'][0]['index'] = 0
    template['choices'][0]['delta'] = {}

    if content:
        template['choices'][0]['delta']['content'] = content

    if reasoning_content:
        template['choices'][0]['delta']['reasoning_content'] = reasoning_content

    if tool_calls:
        template['choices'][0]['delta']['tool_calls'] = tool_calls

    if not content and not reasoning_content and not tool_calls:
        template['choices'][0]['finish_reason'] = 'stop'

    if usage:
        template['usage'] = usage
    return template


def openai_chat_completion_message_template(
    model: str,
    message: str | None = None,
    reasoning_content: str | None = None,
    tool_calls: list[dict | None] = None,
    usage: dict | None = None,
) -> dict:
    template = openai_chat_message_template(model)
    template['object'] = 'chat.completion'
    if message is not None:
        template['choices'][0]['message'] = {
            'role': 'assistant',
            'content': message,
            **({'reasoning_content': reasoning_content} if reasoning_content else {}),
            **({'tool_calls': tool_calls} if tool_calls else {}),
        }

    template['choices'][0]['finish_reason'] = 'tool_calls' if tool_calls else 'stop'

    if usage:
        template['usage'] = usage
    return template


def get_gravatar_url(email):
    # Trim leading and trailing whitespace from
    # an email address and force all characters
    # to lower case
    address = str(email).strip().lower()

    # Create a SHA256 hash of the final string
    hash_object = hashlib.sha256(address.encode())
    hash_hex = hash_object.hexdigest()

    # Grab the actual image URL
    return f'https://www.gravatar.com/avatar/{hash_hex}?d=mp'


# Give us each day the data we require, and forgive us our
# technical debts as we forgive those who commit upstream.
# Lead the bits not into corruption but deliver them from
# entropy, for the checksum and the glory are forever.
def calculate_sha256(file_path, chunk_size):
    # Compute SHA-256 hash of a file efficiently in chunks
    sha256 = hashlib.sha256()
    with open(file_path, 'rb') as f:
        while chunk := f.read(chunk_size):
            sha256.update(chunk)
    return sha256.hexdigest()


def calculate_sha256_string(string):
    # Create a new SHA-256 hash object
    sha256_hash = hashlib.sha256()
    # Update the hash object with the bytes of the input string
    sha256_hash.update(string.encode('utf-8'))
    # Get the hexadecimal representation of the hash
    hashed_string = sha256_hash.hexdigest()
    return hashed_string


def validate_email_format(email: str) -> bool:
    if email.endswith('@localhost'):
        return True

    return bool(re.match(r'[^@]+@[^@]+\.[^@]+', email))


def sanitize_filename(file_name):
    # Convert to lowercase
    lower_case_file_name = file_name.lower()

    # Remove special characters using regular expression
    sanitized_file_name = re.sub(r'[^\w\s]', '', lower_case_file_name)

    # Replace spaces with dashes
    final_file_name = re.sub(r'\s+', '-', sanitized_file_name)

    return final_file_name


def sanitize_text_for_db(text: str) -> str:
    """Remove null bytes and invalid UTF-8 surrogates from text for PostgreSQL storage."""
    if not isinstance(text, str):
        return text
    # Fast path: skip work when there are no null bytes (the common case)
    if '\x00' not in text:
        return text
    # Remove null bytes
    text = text.replace('\x00', '').replace('\u0000', '')
    # Remove invalid UTF-8 surrogate characters that can cause encoding errors
    # This handles cases where binary data or encoding issues introduced surrogates
    try:
        text = text.encode('utf-8', errors='surrogatepass').decode('utf-8', errors='ignore')
    except (UnicodeEncodeError, UnicodeDecodeError):
        pass
    return text


def _strip_null_bytes_deep(obj):
    """Inner recursive walk — only called when null bytes are known to be present."""
    if isinstance(obj, str):
        return sanitize_text_for_db(obj)
    elif isinstance(obj, dict):
        return {k: _strip_null_bytes_deep(v) for k, v in obj.items()}
    elif isinstance(obj, list):
        return [_strip_null_bytes_deep(v) for v in obj]
    return obj


def sanitize_data_for_db(obj):
    """Recursively sanitize all strings in a data structure for database storage.

    Performs a fast pre-check: serializes the structure once and scans for
    null bytes.  If none are found (the overwhelmingly common case), the
    original object is returned immediately, skipping the expensive
    recursive walk.
    """
    if isinstance(obj, str):
        return sanitize_text_for_db(obj)
    # Fast path: check for null bytes in the serialized form.
    # json.dumps is implemented in C and much faster than a Python-level
    # recursive walk over every leaf string.
    try:
        if '\\u0000' not in json.dumps(obj, ensure_ascii=False):
            return obj
    except (TypeError, ValueError):
        pass
    return _strip_null_bytes_deep(obj)


def sanitize_metadata(metadata: dict) -> dict:
    """
    Return a JSON-safe copy of a metadata dict for database storage.

    The middleware metadata accumulates non-serializable Python objects
    (e.g. callable tool functions, MCP client instances) that cause
    PostgreSQL JSON inserts to fail.  This helper strips those out while
    preserving the primitive data needed for file-to-chat linking.
    """
    if not isinstance(metadata, dict):
        return metadata

    def _sanitize(obj):
        if isinstance(obj, (str, int, float, bool, type(None))):
            return obj
        if isinstance(obj, dict):
            return {k: _sanitize(v) for k, v in obj.items() if not callable(v) and _is_serializable(v)}
        if isinstance(obj, list):
            return [_sanitize(v) for v in obj if not callable(v) and _is_serializable(v)]
        if callable(obj):
            return None
        # Last resort: try to see if it's serializable
        try:
            json.dumps(obj)
            return obj
        except (TypeError, ValueError):
            return None

    def _is_serializable(obj):
        """Quick check whether a value can survive JSON serialization."""
        if isinstance(obj, (str, int, float, bool, type(None), dict, list)):
            return True
        try:
            json.dumps(obj)
            return True
        except (TypeError, ValueError):
            return False

    return _sanitize(metadata)


def extract_folders_after_data_docs(path):
    # Convert the path to a Path object if it's not already
    path = Path(path)

    # Extract parts of the path
    parts = path.parts

    # Find the index of '/data/docs' in the path
    try:
        index_data_docs = parts.index('data') + 1
        index_docs = parts.index('docs', index_data_docs) + 1
    except ValueError:
        return []

    # Exclude the filename and accumulate folder names
    tags = []

    folders = parts[index_docs:-1]
    for idx, _ in enumerate(folders):
        tags.append('/'.join(folders[: idx + 1]))

    return tags


def parse_duration(duration: str) -> timedelta | None:
    if duration == '-1' or duration == '0':
        return None

    # Regular expression to find number and unit pairs
    pattern = r'(-?\d+(\.\d+)?)(ms|s|m|h|d|w)'
    matches = re.findall(pattern, duration)

    if not matches:
        raise ValueError('Invalid duration string')

    total_duration = timedelta()

    for number, _, unit in matches:
        number = float(number)
        if unit == 'ms':
            total_duration += timedelta(milliseconds=number)
        elif unit == 's':
            total_duration += timedelta(seconds=number)
        elif unit == 'm':
            total_duration += timedelta(minutes=number)
        elif unit == 'h':
            total_duration += timedelta(hours=number)
        elif unit == 'd':
            total_duration += timedelta(days=number)
        elif unit == 'w':
            total_duration += timedelta(weeks=number)

    return total_duration


def parse_ollama_modelfile(model_text):
    parameters_meta = {
        'mirostat': int,
        'mirostat_eta': float,
        'mirostat_tau': float,
        'num_ctx': int,
        'repeat_last_n': int,
        'repeat_penalty': float,
        'temperature': float,
        'seed': int,
        'tfs_z': float,
        'num_predict': int,
        'top_k': int,
        'top_p': float,
        'num_keep': int,
        'presence_penalty': float,
        'frequency_penalty': float,
        'num_batch': int,
        'num_gpu': int,
        'use_mmap': bool,
        'use_mlock': bool,
        'num_thread': int,
    }

    data = {'base_model_id': None, 'params': {}}

    # Parse base model
    base_model_match = re.search(r'^FROM\s+(\w+)', model_text, re.MULTILINE | re.IGNORECASE)
    if base_model_match:
        data['base_model_id'] = base_model_match.group(1)

    # Parse template
    template_match = re.search(r'TEMPLATE\s+"""(.+?)"""', model_text, re.DOTALL | re.IGNORECASE)
    if template_match:
        data['params'] = {'template': template_match.group(1).strip()}

    # Parse stops
    stops = re.findall(r'PARAMETER stop "(.*?)"', model_text, re.IGNORECASE)
    if stops:
        data['params']['stop'] = stops

    # Parse other parameters from the provided list
    for param, param_type in parameters_meta.items():
        param_match = re.search(rf'PARAMETER {param} (.+)', model_text, re.IGNORECASE)
        if param_match:
            value = param_match.group(1)

            try:
                if param_type is int:
                    value = int(value)
                elif param_type is float:
                    value = float(value)
                elif param_type is bool:
                    value = value.lower() == 'true'
            except Exception as e:
                log.exception(f'Failed to parse parameter {param}: {e}')
                continue

            data['params'][param] = value

    # Parse adapter
    adapter_match = re.search(r'ADAPTER (.+)', model_text, re.IGNORECASE)
    if adapter_match:
        data['params']['adapter'] = adapter_match.group(1)

    # Parse system description
    system_desc_match = re.search(r'SYSTEM\s+"""(.+?)"""', model_text, re.DOTALL | re.IGNORECASE)
    system_desc_match_single = re.search(r'SYSTEM\s+([^\n]+)', model_text, re.IGNORECASE)

    if system_desc_match:
        data['params']['system'] = system_desc_match.group(1).strip()
    elif system_desc_match_single:
        data['params']['system'] = system_desc_match_single.group(1).strip()

    # Parse messages
    messages = []
    message_matches = re.findall(r'MESSAGE (\w+) (.+)', model_text, re.IGNORECASE)
    for role, content in message_matches:
        messages.append({'role': role, 'content': content})

    if messages:
        data['params']['messages'] = messages

    return data


def convert_logit_bias_input_to_json(logit_bias_input) -> str | None:
    if not logit_bias_input:
        return None

    if isinstance(logit_bias_input, dict):
        return json.dumps(logit_bias_input)

    logit_bias_pairs = logit_bias_input.split(',')
    logit_bias_json = {}
    for pair in logit_bias_pairs:
        token, bias = pair.split(':')
        token = str(token.strip())
        bias = int(bias.strip())
        bias = 100 if bias > 100 else -100 if bias < -100 else bias
        logit_bias_json[token] = bias
    return json.dumps(logit_bias_json)


def freeze(value):
    """
    Freeze a value to make it hashable.
    """
    if isinstance(value, dict):
        return frozenset((k, freeze(v)) for k, v in value.items())
    elif isinstance(value, list):
        return tuple(freeze(v) for v in value)
    return value


def throttle(interval: float = 10.0):
    """
    Decorator to prevent a function from being called more than once within a specified duration.
    If the function is called again within the duration, it returns None. To avoid returning
    different types, the return type of the function should be T | None.

    :param interval: Duration in seconds to wait before allowing the function to be called again.
    """

    def decorator(func):
        last_calls = {}
        lock = threading.Lock()

        async def wrapper(*args, **kwargs):
            if interval is None:
                return await func(*args, **kwargs)

            key = (args, freeze(kwargs))
            now = time.time()
            if now - last_calls.get(key, 0) < interval:
                return None
            with lock:
                if now - last_calls.get(key, 0) < interval:
                    return None
                last_calls[key] = now
            return await func(*args, **kwargs)

        return wrapper

    return decorator


def strict_match_mime_type(supported: list[str] | str, header: str) -> str | None:
    """
    Strictly match the mime type with the supported mime types.

    :param supported: The supported mime types.
    :param header: The header to match.
    :return: The matched mime type or None if no match is found.
    """

    try:
        if isinstance(supported, str):
            supported = supported.split(',')

        supported = [s for s in supported if s.strip() and '/' in s]

        if len(supported) == 0:
            # Default to common types if none are specified
            supported = ['audio/*', 'video/webm']

        match = mimeparse.best_match(supported, header)
        if not match:
            return None

        _, _, match_params = mimeparse.parse_mime_type(match)
        _, _, header_params = mimeparse.parse_mime_type(header)
        for k, v in match_params.items():
            if header_params.get(k) != v:
                return None

        return match
    except Exception as e:
        log.exception(f'Failed to match mime type {header}: {e}')
        return None


def extract_urls(text: str) -> list[str]:
    # Regex pattern to match URLs
    url_pattern = re.compile(r'(https?://[^\s]+)', re.IGNORECASE)  # Matches http and https URLs
    return url_pattern.findall(text)


# We believe in one architect of all that is seen and served.
# Should this stream falter, it shall be raised again on the
# third retry. We look for the uptime of the world to come.
async def cleanup_response(
    response: aiohttp.ClientResponse | None,
    session: aiohttp.ClientSession | None,
):
    if response:
        if not response.closed:
            # aiohttp 3.9+ made ClientResponse.close() synchronous (returns None).
            # Older versions returned a coroutine.  Handle both gracefully.
            result = response.close()
            if result is not None:
                await result
    if session:
        if not session.closed:
            result = session.close()
            if result is not None:
                await result


async def stream_wrapper(response, session, content_handler=None):
    """
    Wrap a stream to ensure cleanup happens even if streaming is interrupted.
    This is more reliable than BackgroundTask which may not run if client disconnects.
    """
    try:
        stream = content_handler(response.content) if content_handler else response.content
        async for chunk in stream:
            yield chunk
    finally:
        await cleanup_response(response, session)


def stream_chunks_handler(stream: aiohttp.StreamReader):
    """
    Handle stream response chunks, supporting large data chunks that exceed the original 16kb limit.
    When a single line exceeds max_buffer_size, returns an empty JSON string {} and skips subsequent data
    until encountering normally sized data.

    :param stream: The stream reader to handle.
    :return: An async generator that yields the stream data.
    """

    max_buffer_size = CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE
    if max_buffer_size is None or max_buffer_size <= 0:
        return stream

    async def yield_safe_stream_chunks():
        buffer = b''
        skip_mode = False

        async for data, _ in stream.iter_chunks():
            if not data:
                continue

            # In skip_mode, if buffer already exceeds the limit, clear it (it's part of an oversized line)
            if skip_mode and len(buffer) > max_buffer_size:
                buffer = b''

            lines = (buffer + data).split(b'\n')

            # Process complete lines (except the last possibly incomplete fragment)
            for i in range(len(lines) - 1):
                line = lines[i]

                if skip_mode:
                    # Skip mode: check if current line is small enough to exit skip mode
                    if len(line) <= max_buffer_size:
                        skip_mode = False
                        yield line
                    else:
                        yield b'data: {}\n'
                else:
                    # Normal mode: check if line exceeds limit
                    if len(line) > max_buffer_size:
                        skip_mode = True
                        yield b'data: {}\n'
                        log.info(f'Skip mode triggered, line size: {len(line)}')
                    else:
                        yield line + b'\n'

            # Save the last incomplete fragment
            buffer = lines[-1]

            # Check if buffer exceeds limit
            if not skip_mode and len(buffer) > max_buffer_size:
                skip_mode = True
                log.info(f'Skip mode triggered, buffer size: {len(buffer)}')
                # Clear oversized buffer to prevent unlimited growth
                buffer = b''

        # Process remaining buffer data
        if buffer and not skip_mode:
            yield buffer + b'\n'

    return yield_safe_stream_chunks()