chat_interface_preference.py

"""
This file defines a useful high-level abstraction to build Gradio chatbots: ChatInterface.
"""

from __future__ import annotations

import datetime
import functools
import inspect
import json
import random
import re
import uuid
from typing import AsyncGenerator, Callable, Literal, Union, cast

import anyio
from gradio.blocks import Blocks
from gradio.components import (
    Button,
    Chatbot,
    Component,
    Markdown,
    MultimodalTextbox,
    Slider,
    State,
    Textbox,
    get_component_instance,
)
from gradio.events import Dependency, on
from gradio.helpers import Error, Info
from gradio.helpers import create_examples as Examples  # noqa: N812
from gradio.helpers import special_args
from gradio.layouts import Accordion, Group, Row
from gradio.routes import Request
from gradio.themes import ThemeClass as Theme
from gradio.utils import SyncToAsyncIterator, async_iteration, async_lambda
from gradio_client.documentation import document
from huggingface_hub import CommitScheduler

pattern = re.compile(r'<div class="message-identifier">(.*?)</div>', re.DOTALL)

PREFERENCE_TECHNIQUE_MAPPING = {"sft": "prompt", "dpo": "preference", "kto": "vibes"}
_LOG_REQUIRED_MESSAGE = "First, provide preference, undo or clear to continue conversation."


@document()
class ChatInterface(Blocks):
    """
    ChatInterface is Gradio's high-level abstraction for creating chatbot UIs, and allows you to create
    a web-based demo around a chatbot model in a few lines of code. Only one parameter is required: fn, which
    takes a function that governs the response of the chatbot based on the user input and chat history. Additional
    parameters can be used to control the appearance and behavior of the demo.

    Example:
        import gradio as gr

        def echo(message, history):
            return message

        demo = gr.ChatInterface(fn=echo, examples=["hello", "hola", "merhaba"], title="Echo Bot")
        demo.launch()
    Demos: chatinterface_multimodal, chatinterface_random_response, chatinterface_streaming_echo
    Guides: creating-a-chatbot-fast, sharing-your-app
    """

    def __init__(
        self,
        fn: Callable,
        *,
        prefence_technique: str = None,
        min_turns: int = 1,
        max_turns: int = 1,
        repo_id: None | str,
        repo_private: bool = False,
        multimodal: bool = False,
        chatbot: Chatbot | None = None,
        textbox: Textbox | MultimodalTextbox | None = None,
        additional_inputs: str | Component | list[str | Component] | None = None,
        additional_inputs_accordion_name: str | None = None,
        additional_inputs_accordion: str | Accordion | None = None,
        examples: list[str] | list[dict[str, str | list]] | list[list] | None = None,
        cache_examples: bool | Literal["lazy"] | None = None,
        examples_per_page: int = 10,
        title: str | None = None,
        description: str | None = None,
        theme: Theme | str | None = None,
        css: str | None = None,
        js: str | None = None,
        head: str | None = None,
        analytics_enabled: bool | None = None,
        autofocus: bool = True,
        concurrency_limit: int | None | Literal["default"] = "default",
        fill_height: bool = True,
        delete_cache: tuple[int, int] | None = None,
    ):
        """
        Parameters:
            fn: The function to wrap the chat interface around. Should accept two parameters: a string input message and list of two-element lists of the form [[user_message, bot_message], ...] representing the chat history, and return a string response. See the Chatbot documentation for more information on the chat history format.
            multimodal: If True, the chat interface will use a gr.MultimodalTextbox component for the input, which allows for the uploading of multimedia files. If False, the chat interface will use a gr.Textbox component for the input.
            chatbot: An instance of the gr.Chatbot component to use for the chat interface, if you would like to customize the chatbot properties. If not provided, a default gr.Chatbot component will be created.
            textbox: An instance of the gr.Textbox or gr.MultimodalTextbox component to use for the chat interface, if you would like to customize the textbox properties. If not provided, a default gr.Textbox or gr.MultimodalTextbox component will be created.
            additional_inputs: An instance or list of instances of gradio components (or their string shortcuts) to use as additional inputs to the chatbot. If components are not already rendered in a surrounding Blocks, then the components will be displayed under the chatbot, in an accordion.
            additional_inputs_accordion_name: Deprecated. Will be removed in a future version of Gradio. Use the `additional_inputs_accordion` parameter instead.
            additional_inputs_accordion: If a string is provided, this is the label of the `gr.Accordion` to use to contain additional inputs. A `gr.Accordion` object can be provided as well to configure other properties of the container holding the additional inputs. Defaults to a `gr.Accordion(label="Additional Inputs", open=False)`. This parameter is only used if `additional_inputs` is provided.
            examples: Sample inputs for the function; if provided, appear below the chatbot and can be clicked to populate the chatbot input. Should be a list of strings if `multimodal` is False, and a list of dictionaries (with keys `text` and `files`) if `multimodal` is True.
            cache_examples: If True, caches examples in the server for fast runtime in examples. The default option in HuggingFace Spaces is True. The default option elsewhere is False.
            examples_per_page: If examples are provided, how many to display per page.
            title: a title for the interface; if provided, appears above chatbot in large font. Also used as the tab title when opened in a browser window.
            description: a description for the interface; if provided, appears above the chatbot and beneath the title in regular font. Accepts Markdown and HTML content.
            theme: Theme to use, loaded from gradio.themes.
            css: Custom css as a string or path to a css file. This css will be included in the demo webpage.
            js: Custom js as a string or path to a js file. The custom js should be in the form of a single js function. This function will automatically be executed when the page loads. For more flexibility, use the head parameter to insert js inside <script> tags.
            head: Custom html to insert into the head of the demo webpage. This can be used to add custom meta tags, multiple scripts, stylesheets, etc. to the page.
            analytics_enabled: Whether to allow basic telemetry. If None, will use GRADIO_ANALYTICS_ENABLED environment variable if defined, or default to True.
            submit_btn: Text to display on the submit button. If None, no button will be displayed. If a Button object, that button will be used.
            stop_btn: Text to display on the stop button, which replaces the submit_btn when the submit_btn or retry_btn is clicked and response is streaming. Clicking on the stop_btn will halt the chatbot response. If set to None, stop button functionality does not appear in the chatbot. If a Button object, that button will be used as the stop button.
            retry_btn: Text to display on the retry button. If None, no button will be displayed. If a Button object, that button will be used.
            undo_btn: Text to display on the delete last button. If None, no button will be displayed. If a Button object, that button will be used.
            clear_btn: Text to display on the clear button. If None, no button will be displayed. If a Button object, that button will be used.
            autofocus: If True, autofocuses to the textbox when the page loads.
            concurrency_limit: If set, this is the maximum number of chatbot submissions that can be running simultaneously. Can be set to None to mean no limit (any number of chatbot submissions can be running simultaneously). Set to "default" to use the default concurrency limit (defined by the `default_concurrency_limit` parameter in `.queue()`, which is 1 by default).
            fill_height: If True, the chat interface will expand to the height of window.
            delete_cache: A tuple corresponding [frequency, age] both expressed in number of seconds. Every `frequency` seconds, the temporary files created by this Blocks instance will be deleted if more than `age` seconds have passed since the file was created. For example, setting this to (86400, 86400) will delete temporary files every day. The cache will be deleted entirely when the server restarts. If None, no cache deletion will occur.
        """
        if max_turns < min_turns:
            raise ValueError("`max_turns` should be larger than `min_turns`")
        if any([turn for turn in [max_turns, min_turns] if turn < 1]):
            raise ValueError("`max_turns` should be larger than `min_turns`")
        self.max_turns = max_turns
        self.min_turns = min_turns

        optional_techniques = ["kto", "sft", "spin", "dpo", "simpo", "rlhf", "orpo"]
        if prefence_technique not in optional_techniques:
            raise ValueError(f"Supported techniques are {optional_techniques}")
        submit_btn_one = "Generate"
        submit_btn_two = None
        submit_btn_a = None
        submit_btn_b = None
        submit_btn_ab = None
        submit_btn_good = None
        submit_btn_bad = None
        stop_btn = "Stop"
        undo_btn = "↩️ Undo"
        clear_btn = "🗑️ Clear"
        n_generations = 1
        if "kto" == prefence_technique:
            submit_btn_good = "The response 👍"
            submit_btn_bad = "The response 👎"
        if prefence_technique in ["dpo", "simpo", "rlhf", "orpo"]:
            n_generations = 2
            submit_btn_a = "A is better than B"
            submit_btn_b = "B is better than A"
            submit_btn_ab = "A and B are similar"
        super().__init__(
            analytics_enabled=analytics_enabled,
            mode="chat_interface",
            css=css,
            title=title or "Gradio",
            theme=theme,
            js=js,
            head=head,
            fill_height=fill_height,
            delete_cache=delete_cache,
        )

        self.css = css
        self.multimodal = multimodal
        self.concurrency_limit = concurrency_limit
        self.fn = fn
        self.is_async = inspect.iscoroutinefunction(self.fn) or inspect.isasyncgenfunction(self.fn)
        self.is_generator = inspect.isgeneratorfunction(self.fn) or inspect.isasyncgenfunction(self.fn)
        self.buttons: list[Button | None] = []

        self.examples = examples
        self.cache_examples: bool | None | Literal["lazy"] = cache_examples
        self._set_conversation_id()
        if repo_id:
            self.commit_scheduler = CommitScheduler(
                repo_id=repo_id, folder_path="feedback", repo_type="dataset", private=repo_private, every=1
            )
        else:
            self.commit_scheduler = None
        if self.commit_scheduler:
            self.data_file = self.commit_scheduler.folder_path / f"data_{uuid.uuid4()}.json"

        slider_placeholder = Slider(label="generation", minimum=1.0, maximum=2.0, step=0.05, value=1.2, render=False)

        if additional_inputs:
            if not isinstance(additional_inputs, list):
                additional_inputs = [slider_placeholder] + [additional_inputs]
            else:
                additional_inputs = [slider_placeholder] + additional_inputs
            self.additional_inputs = [get_component_instance(i) for i in additional_inputs]  # type: ignore
        else:
            self.additional_inputs = [slider_placeholder]
        if additional_inputs_accordion_name is not None:
            print(
                "The `additional_inputs_accordion_name` parameter is deprecated and will be removed in a future version of Gradio. Use the `additional_inputs_accordion` parameter instead."
            )
            self.additional_inputs_accordion_params = {"label": additional_inputs_accordion_name}
        if additional_inputs_accordion is None:
            self.additional_inputs_accordion_params = {
                "label": "Additional Inputs",
                "open": False,
            }
        elif isinstance(additional_inputs_accordion, str):
            self.additional_inputs_accordion_params = {"label": additional_inputs_accordion}
        elif isinstance(additional_inputs_accordion, Accordion):
            self.additional_inputs_accordion_params = additional_inputs_accordion.recover_kwargs(
                additional_inputs_accordion.get_config()
            )
        else:
            raise ValueError(
                f"The `additional_inputs_accordion` parameter must be a string or gr.Accordion, not {type(additional_inputs_accordion)}"
            )

        with self:
            if title:
                Markdown(f"<h1 style='text-align: center; margin-bottom: 1rem'>{self.title}</h1>")
            if self.commit_scheduler:
                Markdown(
                    f'<center><h2>Data is being logged to <a href="https://huggingface.co/datasets/{self.commit_scheduler.repo_id}">a dataset on the Hugging Face Hub</a></h2></center>'
                )
            if description:
                Markdown(description)

            if chatbot:
                self.chatbot = chatbot.render()
            else:
                self.chatbot = Chatbot(label="Chatbot", scale=1, height=200 if fill_height else None)

            with Row():
                for btn in [
                    submit_btn_a,
                    submit_btn_b,
                    submit_btn_ab,
                    submit_btn_good,
                    submit_btn_bad,
                    undo_btn,
                    clear_btn,
                ]:
                    if btn is not None:
                        if isinstance(btn, Button):
                            btn.render()
                        elif isinstance(btn, str):
                            btn = Button(btn, variant="secondary", size="sm", min_width=60)
                        else:
                            raise ValueError(
                                f"All the _btn parameters must be a gr.Button, string, or None, not {type(btn)}"
                            )
                    self.buttons.append(btn)  # type: ignore

            with Group():
                with Row():
                    if textbox:
                        if self.multimodal:
                            submit_btn_one = None
                            submit_btn_two = None
                        else:
                            textbox.container = False
                        textbox.show_label = False
                        textbox_ = textbox.render()
                        if not isinstance(textbox_, (Textbox, MultimodalTextbox)):
                            raise TypeError(
                                f"Expected a gr.Textbox or gr.MultimodalTextbox component, but got {type(textbox_)}"
                            )
                        self.textbox = textbox_
                    elif self.multimodal:
                        submit_btn_one = None
                        submit_btn_two = None
                        self.textbox = MultimodalTextbox(
                            show_label=False,
                            label="Message",
                            placeholder="Type a message...",
                            scale=7,
                            autofocus=autofocus,
                        )
                    else:
                        self.textbox = Textbox(
                            container=False,
                            show_label=False,
                            label="Message",
                            placeholder="Type a message...",
                            scale=7,
                            autofocus=autofocus,
                        )
                    submit_buttons = []
                    for btn in [submit_btn_one, submit_btn_two]:
                        if btn is not None and not multimodal:
                            if isinstance(btn, Button):
                                btn.render()
                            elif isinstance(btn, str):
                                btn = Button(
                                    btn,
                                    variant="primary",
                                    scale=1,
                                    min_width=150,
                                )
                            else:
                                raise ValueError(
                                    f"The submit_btn parameter must be a gr.Button, string, or None, not {type(btn)}"
                                )
                        submit_buttons.append(btn)

                    if stop_btn is not None:
                        if isinstance(stop_btn, Button):
                            stop_btn.visible = False
                            stop_btn.render()
                        elif isinstance(stop_btn, str):
                            stop_btn = Button(
                                stop_btn,
                                variant="stop",
                                visible=False,
                                scale=1,
                                min_width=150,
                            )
                        else:
                            raise ValueError(
                                f"The stop_btn parameter must be a gr.Button, string, or None, not {type(stop_btn)}"
                            )
                    self.buttons.extend(submit_buttons + [stop_btn])  # type: ignore

                self.fake_api_btn = Button("Fake API", visible=False)
                self.fake_response_textbox = Textbox(label="Response", visible=False)
                (
                    self.submit_btn_a,
                    self.submit_btn_b,
                    self.submit_btn_ab,
                    self.submit_btn_good,
                    self.submit_btn_bad,
                    self.undo_btn,
                    self.clear_btn,
                    self.submit_btn_one,
                    self.submit_btn_two,
                    self.stop_btn,
                ) = self.buttons

            if examples:
                if self.is_generator:
                    examples_fn = self._examples_stream_fn
                else:
                    examples_fn = self._examples_fn

                self.examples_handler = Examples(
                    examples=examples,
                    inputs=[self.textbox] + self.additional_inputs,
                    outputs=self.chatbot,
                    fn=examples_fn,
                    cache_examples=self.cache_examples,
                    _defer_caching=True,
                    examples_per_page=examples_per_page,
                )

            any_unrendered_inputs = any(not inp.is_rendered for inp in self.additional_inputs)
            if self.additional_inputs and any_unrendered_inputs:
                with Accordion(**self.additional_inputs_accordion_params):  # type: ignore
                    for input_component in self.additional_inputs:
                        if not input_component.is_rendered:
                            input_component.render()

            # The example caching must happen after the input components have rendered
            if examples:
                self.examples_handler._start_caching()

            self.saved_input = State()
            self.chatbot_state = State(self.chatbot.value) if self.chatbot.value else State([])

            self._setup_events(n_generations)
            self._setup_api()

    def _set_conversation_id(self):
        self._conversation_id = str(uuid.uuid4())

    def _save_feedback(self, item):
        feedback = {
            "timestamp": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
            "conversation_id": self._conversation_id,
        }
        feedback.update(item)
        if self.commit_scheduler:
            with self.commit_scheduler.lock:
                with self.data_file.open("a") as f:
                    f.write(json.dumps(feedback))

    def _setup_events(self, n_generations) -> None:
        submit_fn_one = self._stream_fn if self.is_generator else self._submit_fn
        submit_fn_one_partial = functools.partial(submit_fn_one, n_generations=n_generations)
        submit_triggers_one = (
            [self.textbox.submit, self.submit_btn_one.click] if self.submit_btn_one else [self.textbox.submit]
        )
        submit_tuples = [(submit_fn_one_partial, submit_triggers_one)]
        if self.submit_btn_two:
            submit_fn_two = functools.partial(submit_fn_one, n_generations=n_generations)
            submit_triggers_two = [self.submit_btn_two.click]
            submit_tuples.append((submit_fn_two, submit_triggers_two))
        for _fn, _triggers in submit_tuples:
            submit_event = (
                on(
                    _triggers,
                    self._clear_and_save_textbox,
                    [self.textbox],
                    [self.textbox, self.saved_input],
                    show_api=False,
                    queue=False,
                )
                .then(
                    self._display_input,
                    [self.saved_input, self.chatbot_state],
                    [self.chatbot, self.chatbot_state],
                    show_api=False,
                    queue=False,
                )
                .then(
                    _fn,
                    [self.saved_input, self.chatbot_state] + self.additional_inputs,
                    [self.chatbot, self.chatbot_state],
                    show_api=False,
                    concurrency_limit=cast(Union[int, Literal["default"], None], self.concurrency_limit),
                )
            )
            self._setup_stop_events(_triggers, submit_event)

        partial_fn_a, partial_fn_b, partial_fn_ab, partial_fn_good, partial_fn_bad = (
            functools.partial(self._log_fn, log="a"),
            functools.partial(self._log_fn, log="b"),
            functools.partial(self._log_fn, log="ab"),
            functools.partial(self._log_fn, log="good"),
            functools.partial(self._log_fn, log="bad"),
        )
        for _fn, _btn in [
            (partial_fn_a, self.submit_btn_a),
            (partial_fn_b, self.submit_btn_b),
            (partial_fn_ab, self.submit_btn_ab),
            (partial_fn_good, self.submit_btn_good),
            (partial_fn_bad, self.submit_btn_bad),
        ]:
            if _btn:
                _btn.click(
                    _fn,
                    [self.saved_input, self.chatbot_state],
                    [self.chatbot, self.saved_input, self.chatbot_state],
                    show_api=False,
                    queue=False,
                ).then(
                    async_lambda(lambda x: x),
                    [self.saved_input],
                    [self.textbox],
                    show_api=False,
                    queue=False,
                )

        if self.undo_btn:
            self.undo_btn.click(
                self._delete_prev_fn,
                [self.saved_input, self.chatbot_state],
                [self.chatbot, self.saved_input, self.chatbot_state],
                show_api=False,
                queue=False,
            ).then(
                async_lambda(lambda x: x),
                [self.saved_input],
                [self.textbox],
                show_api=False,
                queue=False,
            )

        if self.clear_btn:
            self.clear_btn.click(
                self._clear_fn,
                [self.saved_input, self.chatbot_state],
                [self.chatbot, self.saved_input, self.chatbot_state],
                show_api=False,
                queue=False,
            ).then(
                async_lambda(lambda x: x),
                [self.saved_input],
                [self.textbox],
                show_api=False,
                queue=False,
            )

    def _setup_stop_events(self, event_triggers: list[Callable], event_to_cancel: Dependency) -> None:
        if self.stop_btn and self.is_generator:
            if self.submit_btn_one:
                for event_trigger in event_triggers:
                    event_trigger(
                        async_lambda(
                            lambda: (
                                Button(visible=False),
                                Button(visible=True),
                            )
                        ),
                        None,
                        [self.submit_btn_one, self.stop_btn],
                        show_api=False,
                        queue=False,
                    )
                event_to_cancel.then(
                    async_lambda(lambda: (Button(visible=True), Button(visible=False))),
                    None,
                    [self.submit_btn_one, self.stop_btn],
                    show_api=False,
                    queue=False,
                )
            else:
                for event_trigger in event_triggers:
                    event_trigger(
                        async_lambda(lambda: Button(visible=True)),
                        None,
                        [self.stop_btn],
                        show_api=False,
                        queue=False,
                    )
                event_to_cancel.then(
                    async_lambda(lambda: Button(visible=False)),
                    None,
                    [self.stop_btn],
                    show_api=False,
                    queue=False,
                )
            self.stop_btn.click(
                None,
                None,
                None,
                cancels=event_to_cancel,
                show_api=False,
            )

    def _setup_api(self) -> None:
        if self.is_generator:

            @functools.wraps(self.fn)
            async def api_fn(message, history, *args, **kwargs):  # type: ignore
                if self.is_async:
                    generator = self.fn(message, history, *args, **kwargs)
                else:
                    generator = await anyio.to_thread.run_sync(
                        self.fn, message, history, *args, **kwargs, limiter=self.limiter
                    )
                    generator = SyncToAsyncIterator(generator, self.limiter)
                try:
                    first_response = await async_iteration(generator)
                    yield first_response, history + [[message, first_response]]
                except StopIteration:
                    yield None, history + [[message, None]]
                async for response in generator:
                    yield response, history + [[message, response]]

        else:

            @functools.wraps(self.fn)
            async def api_fn(message, history, *args, **kwargs):
                if self.is_async:
                    response = await self.fn(message, history, *args, **kwargs)
                else:
                    response = await anyio.to_thread.run_sync(
                        self.fn, message, history, *args, **kwargs, limiter=self.limiter
                    )
                history.append([message, response])
                return response, history

        self.fake_api_btn.click(
            api_fn,
            [self.textbox, self.chatbot_state] + self.additional_inputs,
            [self.textbox, self.chatbot_state],
            api_name="chat",
            concurrency_limit=cast(Union[int, Literal["default"], None], self.concurrency_limit),
        )

    def _clear_and_save_textbox(self, message: str) -> tuple[str | dict, str]:
        if self.multimodal:
            return {"text": "", "files": []}, message
        else:
            return "", message

    def _append_multimodal_history(
        self,
        message: dict[str, list],
        response: str | None,
        history: list[list[str | tuple | None]],
    ):
        for x in message["files"]:
            history.append([(x,), None])
        if message["text"] is None or not isinstance(message["text"], str):
            return
        elif message["text"] == "" and message["files"] != []:
            history.append([None, response])
        else:
            history.append([message["text"], response])

    async def _display_input(
        self, message: str | dict[str, list], history: list[list[str | tuple | None]]
    ) -> tuple[list[list[str | tuple | None]], list[list[str | tuple | None]]]:
        if self.multimodal and isinstance(message, dict):
            self._append_multimodal_history(message, None, history)
        elif isinstance(message, str):
            history.append([message, None])
        return history, history

    def _get_conversation_from_history(self, history):
        conversation = ""
        history[-1] = [history[-1][0], ""]
        for idx, turn in enumerate(history):
            conversation += self._get_chat_message(turn[0], role="system", turn=(idx + 1))
            if turn[-1]:
                conversation += self._get_chat_message(turn[-1], role="user", turn=(idx + 1))

        return "<body>" + conversation + "</body>"

    def _get_conversation_in_openai_format(self, history):
        conversation = []
        for idx, turn in enumerate(history):
            roles = ["user", "assistant"]
            if idx == len(turn) - 1:
                roles = ["user"]
            for role, content in zip(roles, turn):
                conversation.append({"role": role, "content": content})
        return conversation

    @staticmethod
    def _get_chat_message(message, role, turn):
        # return f"<p><div class='message-identifier'>{message}</div></p>"
        return (
            '<div class="message-content">'
            + f"<strong>Option {turn}</strong><br>"
            + f"<em>Length: {len(message)} characters</em><br><br>"
            + f'<div class="message-identifier">{message}</div>'
            + "</div>"
        )

    def _get_chat_message_comparison(self, content_a, content_b):
        return (
            '<div class="container" style="display: flex; width: 100%;">'
            + '<div class="column" style="flex: 1; padding: 10px;">'
            + self._get_chat_message(message=content_a, role="system", turn="A")
            + "</div>"
            + '<div class="column" style="flex: 1; padding: 10px;">'
            + self._get_chat_message(message=content_b, role="system", turn="B")
            + "</div>"
            + "</div>"
        )

    @staticmethod
    def _check_if_two_responses(response):
        if response:
            matches = pattern.findall(response)
            return matches

    def _check_num_turns(self, history, generate=True):
        if generate:
            if len(history) >= self.max_turns:
                raise Error(
                    f"We intend to collect conversations with a maximum of {self.max_turns}, please clear or log info first."
                )
                return history, history
        else:
            if len(history) < self.min_turns:
                raise Error(
                    f"We intend to collect conversations with at least of {self.min_turns}, please continue the conversation first."
                )
                return history, history

    @staticmethod
    def _check_message(message):
        if not message:
            raise Error("Make sure to provide a message next time.")

    async def _submit_fn(
        self,
        message: str | dict[str, list],
        history_with_input: list[list[str | tuple | None]],
        request: Request,
        *args,
        n_generations: int,
    ) -> tuple[list[list[str | tuple | None]], list[list[str | tuple | None]]]:
        if self.multimodal and isinstance(message, dict):
            remove_input = len(message["files"]) + 1 if message["text"] is not None else len(message["files"])
            history = history_with_input[:-remove_input]
        else:
            history = history_with_input[:-1]

        self._check_message(message)
        self._check_num_turns(history)
        if history:
            _, response = history[-1]
        else:
            response = None
        if self._check_if_two_responses(response):
            Info(_LOG_REQUIRED_MESSAGE)
            return history, history
        else:
            inputs, _, _ = special_args(self.fn, inputs=[message, history, *args], request=request)

            async def _get_response():
                if self.is_async:
                    response = await self.fn(*inputs)
                else:
                    response = await anyio.to_thread.run_sync(self.fn, *inputs, limiter=self.limiter)
                return response

            if n_generations == 1:
                response = await _get_response()
            else:
                response_one, response_two = await _get_response(), await _get_response()
                response = self._get_chat_message_comparison(response_one, response_two)

            if self.multimodal and isinstance(message, dict):
                self._append_multimodal_history(message, response, history)
            elif isinstance(message, str):
                history.append([message, response])
            return history, history

    async def _stream_fn(
        self,
        message: str | dict[str, list],
        history_with_input: list[list[str | tuple | None]],
        request: Request,
        *args,
        n_generations: int,
    ) -> AsyncGenerator:
        if self.multimodal and isinstance(message, dict):
            remove_input = len(message["files"]) + 1 if message["text"] is not None else len(message["files"])
            history = history_with_input[:-remove_input]
        else:
            history = history_with_input[:-1]
        self._check_message(message)
        self._check_num_turns(history)

        if history:
            _, response = history[-1]
        else:
            response = None
        if self._check_if_two_responses(response):
            Info(_LOG_REQUIRED_MESSAGE)
            yield history, history
        else:
            inputs, _, _ = special_args(self.fn, inputs=[message, history, *args], request=request)

            try:
                if self.is_async:
                    generator = self.fn(*inputs)
                else:
                    generator = await anyio.to_thread.run_sync(self.fn, *inputs, limiter=self.limiter)
                generator = SyncToAsyncIterator(generator, self.limiter)
                first_response = await async_iteration(generator)
                if n_generations == 2:
                    first_response_formatted = self._get_chat_message_comparison(first_response, "")
                else:
                    first_response_formatted = first_response
                if self.multimodal and isinstance(message, dict):
                    for x in message["files"]:
                        history.append([(x,), None])
                    update = history + [[message["text"], first_response_formatted]]
                    yield update, update
                else:
                    update = history + [[message, first_response_formatted]]
                    yield update, update
            except StopIteration:
                if self.multimodal and isinstance(message, dict):
                    self._append_multimodal_history(message, None, history)
                    yield history, history
                else:
                    update = history + [[message, None]]
                    yield update, update
            async for response in generator:
                if n_generations == 2:
                    response_formatted = self._get_chat_message_comparison(response, "")
                else:
                    response_formatted = response
                if self.multimodal and isinstance(message, dict):
                    update = history + [[message["text"], response_formatted]]
                    yield update, update
                else:
                    update = history + [[message, response_formatted]]
                    yield update, update

            if n_generations == 2:
                if self.is_async:
                    generator_two = self.fn(*inputs)
                else:
                    generator_two = await anyio.to_thread.run_sync(self.fn, *inputs, limiter=self.limiter)
                    generator_two = SyncToAsyncIterator(generator_two, self.limiter)
                try:
                    first_response_two = await async_iteration(generator_two)
                    first_response_two_formatted = self._get_chat_message_comparison(response, first_response_two)
                    if self.multimodal and isinstance(message, dict):
                        for x in message["files"]:
                            history.append([(x,), None])

                        update = history + [[message["text"], first_response_two_formatted]]
                        yield update, update
                    else:
                        update = history + [[message, first_response_two_formatted]]
                        yield update, update
                except StopIteration:
                    if self.multimodal and isinstance(message, dict):
                        self._append_multimodal_history(message, None, history)
                        yield history, history
                    else:
                        update = history + [[message, None]]
                        yield update, update
                async for response_two in generator_two:
                    response_two = self._get_chat_message_comparison(response, response_two)
                    if self.multimodal and isinstance(message, dict):
                        update = history + [[message["text"], response_two]]
                        yield update, update
                    else:
                        update = history + [[message, response_two]]
                        yield update, update

    async def _log_fn(
        self, message: str | dict[str, list], history: list[list[str | tuple | None]], log: str
    ) -> tuple[
        list[list[str | tuple | None]],
        str | dict[str, list],
        list[list[str | tuple | None]],
    ]:
        self._check_num_turns(history, generate=False)
        prompt, response = history[-1]
        history_as_openai_format = self._get_conversation_in_openai_format(history)
        feedback = {"prompt": history_as_openai_format}
        feedback.update({"response": [{"content": response, "role": "assistant"}]})
        matches = self._check_if_two_responses(response)

        if matches and log != "prompt":
            option_a, option_b = matches[0], matches[1]
            if log == "a":
                chosen, rejected = option_a, option_b
            elif log == "b":
                chosen, rejected = option_b, option_a
            elif log == "ab":
                options = [option_a, option_b]
                chosen, rejected = random.choice([options])
                Info("Picked random response to continue")
            if log in ["a", "b"] and self.commit_scheduler:
                feedback.update(
                    {
                        "chosen": [{"content": chosen, "role": "assistant"}],
                        "rejected": [{"content": rejected, "role": "assistant"}],
                    }
                )
                feedback["response"] = [{"content": chosen, "role": "assistant"}]
                self._save_feedback(feedback)
                Info("Logged succesfully")
            elif log == "ab":
                self._save_feedback(feedback)
            history[-1] = [prompt, chosen]
            return history, message or "", history
        elif log in ["conversation", "good", "bad"]:
            if log == "good":
                feedback.update({"label": True})
            elif log == "bad":
                feedback.update({"label": False})

            self._save_feedback(feedback)
            Info("Logged succesfully")
            return history, "", history
        else:
            self._save_feedback(feedback)
            Info("Logged succesfully")
            return history, "", history

    async def _examples_fn(self, message: str, *args) -> list[list[str | None]]:
        inputs, _, _ = special_args(self.fn, inputs=[message, [], *args], request=None)
        if self.is_async:
            response = await self.fn(*inputs)
        else:
            response = await anyio.to_thread.run_sync(self.fn, *inputs, limiter=self.limiter)
        return [[message, response]]

    async def _examples_stream_fn(
        self,
        message: str,
        *args,
    ) -> AsyncGenerator:
        inputs, _, _ = special_args(self.fn, inputs=[message, [], *args], request=None)
        if self.is_async:
            generator = self.fn(*inputs)
        else:
            generator = await anyio.to_thread.run_sync(self.fn, *inputs, limiter=self.limiter)
            generator = SyncToAsyncIterator(generator, self.limiter)
        async for response in generator:
            yield [[message, response]]

    async def _delete_prev_fn(
        self,
        message: str | dict[str, list],
        history: list[list[str | tuple | None]],
    ) -> tuple[
        list[list[str | tuple | None]],
        str | dict[str, list],
        list[list[str | tuple | None]],
    ]:
        if self.multimodal and isinstance(message, dict):
            remove_input = len(message["files"]) + 1 if message["text"] is not None else len(message["files"])
            history = history[:-remove_input]
        else:
            history = history[:-1]
        return history, message or "", history

    async def _clear_fn(
        self,
        message: str | dict[str, list],
        history: list[list[str | tuple | None]],
    ) -> tuple[
        list[list[str | tuple | None]],
        str | dict[str, list],
        list[list[str | tuple | None]],
    ]:
        if history:
            _, response = history[-1]
            if self._check_if_two_responses(response):
                Info(_LOG_REQUIRED_MESSAGE)
                return history, message or "", history
            else:
                await self._log_fn(message=message, history=history, log="prompt")
                self._set_conversation_id()
                return [], "", []
        else:
            return [], "", []