sci/streamlit_chatbot.py

from datetime import datetime
import streamlit as st
from _base_class import StreamlitBaseClass, BaseClass
from _llm import LLM
from prompts import *
from colorprinter.print_color import *
from ollama._types import Message as OllamaMessage
from projects_page import Project
from ollama_response_classes import QueryResponse


class Chat(StreamlitBaseClass):
    """
    A class to represent a chat session in a Streamlit application.

    Attributes:
    -----------
    name : str
        The name of the chat.
    chat_history : list
        A list to store the chat history.
    role : str
        The role of the user in the chat.
    project : str
        The project associated with the chat.
    collection : str
        The collection associated with the chat.
    _key : str
        The unique key for the chat.

    Methods:
    --------
    add_message(role, content):
        Adds a message to the chat history.
    
    to_dict():
        Converts the chat object to a dictionary.
    
    update_in_arango():
        Updates the chat object in the ArangoDB.
    
    set_name(user_input):
        Sets the name of the chat based on user input.
    
    show_title(title=None):
        Displays the title of the chat in the Streamlit application.
    
    from_dict(data):
        Creates a Chat object from a dictionary.
    
    chat_history2bot(n_messages=None, remove_system=False):
        Converts the chat history to a format suitable for a bot.
    """
    def __init__(
        self,
        username=None,
        role=None,
        key=None,
        project=None,
        collection=None,
        **kwargs,
    ):
        super().__init__(username=username, **kwargs)
        self.name = kwargs.get("name", None)
        self.chat_history = kwargs.get("chat_history", [])
        self.role = role
        self.project = kwargs.get("project") if "project" in kwargs else project
        self.collection = (
            kwargs.get("collection") if "collection" in kwargs else collection
        )
        self._key = key

    def add_message(self, role, content):
        if isinstance(content, str):
            content = content.strip().strip('"')
        elif isinstance(content, dict):
            content = content["content"].strip().strip('"')
        else:
            try:
                content = content.get("content", "").strip().strip('"')
            except:
                content = content

        self.chat_history.append(
            {
                "role": role,
                "content": content,
                "role_type": self.role,
            }
        )

    def to_dict(self):
        return {
            "_key": self._key,
            "name": self.name,
            "chat_history": self.chat_history,
            "role": self.role,
            "username": self.username,
        }

    def update_in_arango(self):
        self.last_updated = datetime.now().isoformat()
        self.user_arango.db.collection("chats").insert(
            self.to_dict(), overwrite=True, overwrite_mode="update"
        )

    def set_name(self, user_input):
        llm = LLM(
            model="small",
            max_length_answer=50,
            temperature=0.4,
            system_message="You are a chatbot who will be chatting with a user",
        )
        prompt = (
            f'Give a short name to the chat based on this user input: "{user_input}" '
            "No more than 30 characters. Answer ONLY with the name of the chat."
        )
        name = llm.generate(prompt).content.strip('"')
        name = f'{name} - {datetime.now().strftime("%B %d")}'
        existing_chat = self.user_arango.db.aql.execute(
            f'FOR doc IN chats FILTER doc.name == "{name}" RETURN doc', count=True
        )
        if existing_chat.count() > 0:
            name = f'{name} ({datetime.now().strftime("%H:%M")})'
        name += f" - [{self.role}]"
        self.name = name
        return name

    def show_title(self, title=None):
        title = (
            title
            if title
            else (
                self.project
                if self.project
                else self.collection if self.collection else "No title"
            )
        )
        st.markdown(
            f"""### Chat about *{title.strip()}* with *{self.role}*""",
        )


    @classmethod
    def from_dict(cls, data):
        return cls(
            username=data.get("username"),
            name=data.get("name"),
            chat_history=data.get("chat_history", []),
            role=data.get("role", "Research Assistant"),
            _key=data.get("_key"),
        )

    def chat_history2bot(self, n_messages: int = None, remove_system: bool = False):
        history = [
            {"role": m["role"], "content": m["content"]} for m in self.chat_history
        ]
        if n_messages and len(history) > n_messages:
            history = history[-n_messages:]
            if (
                all([history[0]["role"] == "system", remove_system])
                or history[0]["role"] == "assistant"
            ):
                history = history[1:]
        return history


class StreamlitChat(Chat):
    '''
    A class to manage chat interactions within a Streamlit application.

    Inherits from the Chat class and provides additional functionality to handle
    chat history, user roles, and avatars within a Streamlit app context.

    Attributes:
        project (str): The project associated with the chat.
        collection (str): The collection associated with the chat.
        message_attachments (None): Placeholder for message attachments.
        last_updated (str): Timestamp of the last update in ISO format.
        _key (str): Unique identifier for the chat.
        role (str): The role of the user in the chat.
        username (str): The username of the user in the chat.
        name (str): The name of the chat.
        chat_history (list): List of messages in the chat history.

    Methods:
        show_chat_history():
        get_avatar(message: dict = None, role=None) -> str:
        '''
    def __init__(self, username: str, role: str, _key: str = None, **kwargs):
        super().__init__(username, role, _key, **kwargs)
        self.project = kwargs.get("project", None)
        self.collection = kwargs.get("collection", None)
        self.message_attachments = None
        self.last_updated = datetime.now().isoformat()
        self._key = _key
        self.role = role

        if self._key:
            chat = self.user_arango.db.collection("chats").get(self._key)
            if chat:
                self.name = chat.get("name")
                self.chat_history = chat.get("chat_history", [])
                self.role = chat.get("role")
                self.username = chat.get("username")
        else:
            self._key = self.user_arango.db.collection("chats").insert(
                {
                    "name": self.name,
                    "chat_history": self.chat_history,
                    "role": self.role,
                    "username": self.username,
                }
            )["_key"]

    def show_chat_history(self):
        """
        Displays the chat history in the Streamlit app.

        Iterates through the chat history and displays messages from the user and assistant.
        Messages from other roles are ignored. Each message is displayed with an avatar.

        Returns:
            None
        """
        for message in self.chat_history:
            if message["role"] not in ["user", "assistant"]:
                continue
            avatar = self.get_avatar(message)
            with st.chat_message(message["role"], avatar=avatar):
                if message["content"]:
                    st.markdown(message["content"].strip('"'))

    def get_avatar(self, message: dict = None, role=None) -> str:
        """
        Retrieves the avatar image path based on the message or role provided.

        Args:
            message (dict, optional): A dictionary containing message details, including the role.
            role (str, optional): The role of the user if the message is not provided.

        Returns:
            str: The file path to the avatar image.

        Raises:
            AssertionError: If neither message nor role is provided.
        """
        assert message or role, "Either message or role must be provided"
        if message and message.get("role", None) == "user" or role == "user":
            avatar = st.session_state["settings"].get("avatar", "user")
        elif (
            message and message.get("role", None) == "assistant" or role == "assistant"
        ):
            role_type = message.get("role_type", self.role) if message else self.role
            if role_type == "Research Assistant":
                avatar = "img/avatar_researcher.png"
            elif role_type == "Editor":
                avatar = "img/avatar_editor.png"
            elif role_type == "Host":
                avatar = "img/avatar_host.png"
            elif role_type == "Guest":
                avatar = "img/avatar_guest.png"
            else:
                avatar = None
        else:
            avatar = None
        return avatar


class Bot(BaseClass):
    '''
    A chatbot class that integrates with research tools and document retrieval systems.
    The Bot class provides an interface for conversational AI that can access and process
    various document sources, including scientific articles, user notes, and other documents.
    It initializes multiple specialized language models for different tasks, including
    regular conversation, query generation, and tool selection.
    Attributes:
        username (str): The username associated with this bot instance.
        chat (Chat): Chat instance for managing conversation history.
        project (Project, optional): Associated project for document context.
        collection (list, optional): Collections of documents to search within.
        arango_ids (list): List of document IDs in ArangoDB.
        chatbot (LLM): Main language bot for conversation.
        helperbot (LLM): Bot for generating queries.
        toolbot (LLM): Bot for selecting appropriate tools.
        tools (list): List of tool functions available to the bot.
    Methods:
        initiate_bots(): Initialize the different language model instances.
        get_chunks(): Retrieve relevant text chunks based on user input.
        answer_tool_call(): Process and execute tool calls from the AI.
        generate_from_notes(): Generate a response from user notes.
        generate_from_chunks(): Generate a response from document chunks.
        run(): Run the bot (implemented by subclasses).
        get_notes(): Retrieve notes from the database.
        fetch_science_articles_tool(): Retrieve scientific articles.
        fetch_other_documents_tool(): Retrieve non-scientific documents.
        fetch_science_articles_and_other_documents_tool(): Retrieve both document types.
        fetch_notes_tool(): Retrieve user notes.
        conversational_response_tool(): Generate a simple conversational response.
    '''
    def __init__(self, username: str, chat: Chat = None, tools: list = None, **kwargs):
        super().__init__(username=username, **kwargs)
        # Use the passed in chat or create a new Chat
        self.chat = chat if chat else Chat(username=username, role="Research Assistant")
        # Store or set up project/collection if available
        self.project: Project = kwargs.get("project", None)

        self.collection = kwargs.get("collection", None)
        if self.collection and not isinstance(self.collection, list):
            self.collection = [self.collection]
        elif self.project:
            self.collection = self.project.collections

        # Load articles in the collections
        self.arango_ids = []

        # Bots to be initiated later
        self.chatbot = None
        self.helperbot = None
        self.toolbot = None

        if self.collection:
            for c in self.collection:
                for _id in self.user_arango.db.aql.execute(
                    """
                    FOR doc IN article_collections
                    FILTER doc.name == @collection
                    FOR article IN doc.articles
                        RETURN article._id
                    """,
                    bind_vars={"collection": c},
                ):
                    self.arango_ids.append(_id)


        # Convert tool names to function references
        if tools:
            # Map tool names to functions
            tool_mapping = {
                "fetch_other_documents_tool": self.fetch_other_documents_tool,
                "fetch_science_articles_tool": self.fetch_science_articles_tool,
                "fetch_science_articles_and_other_documents_tool": self.fetch_science_articles_and_other_documents_tool,
                "fetch_notes_tool": self.fetch_notes_tool,
                "conversational_response_tool": self.conversational_response_tool,
            }
            self.tools = [
                tool_mapping[tool] if isinstance(tool, str) else tool for tool in tools
            ]
        else:
            self.tools = None

        self.initiate_bots()
        # Store other kwargs
        for arg in kwargs:
            setattr(self, arg, kwargs[arg])

        # # Initiate the bots
        # try:
        #     self.initiate_bots()
        # except Exception as e:
        #     print_red(f"Error initiating bots: {e}")

    def initiate_bots(self):
        """
        Initialize the different bot instances used in the chatbot application.
        
        Creates three types of bots:
        1. chatbot: A standard LLM for normal conversation with the user
        2. helperbot: A specialized LLM with low temperature for generating concise queries or prompts
        3. toolbot: A specialized LLM for selecting which tool to use when responding to user queries
           (only created if tools are provided)
        
        The toolbot is configured to prefer specialized tools over conversational responses
        when the user is seeking information rather than engaging in small talk.
        
        Note:
            - The chatbot uses the full chat history
            - The helperbot uses a limited chat history (last 4 messages) with system message removed
            - The toolbot uses a system message that lists all available tools
        """
        # A standard LLM for normal chat
        self.chatbot = LLM(messages=self.chat.chat_history2bot())
        # A helper bot for generating queries or short prompts
        self.helperbot = LLM(
            temperature=0,
            model="small",
            max_length_answer=500,
            system_message=get_query_builder_system_message(),
            messages=self.chat.chat_history2bot(n_messages=4, remove_system=True),
        )
        # A specialized LLM picking which tool to use
        if self.tools:
            tools_names = [tool.__name__ for tool in self.tools]
            tools_name_string = "\n– ".join(tools_names)
            self.toolbot = LLM(
                temperature=0,
                system_message=f"""
                You are an helpful assistant with tools. The tools you can choose from are:
                {tools_name_string}
                Your task is to choose one or multiple tools to answering a user's query.
                DON'T come up with your own tools, only use the ones provided.
                """,
                # system_message='Use one of the provided tools to help the answering bot to answer the user. Do not answer directly. Use the "tool_calls" field in your answer.',
                chat=False,
                model="tools",
            )
            if len(tools_names) > 1 and "conversational_response_tool" in tools_names:
                self.toolbot.system_message += "\n\nMake sure to only use the conversational response tool if the user is engaging in small talk. If the user is asking a question or looking for information, make sure to use one of the other tools!"


    def get_chunks(
        self,
        user_input,
        collections=["sci_articles", "other_documents"],
        n_results=7,
        n_sources=4,
        filter=True,
    ):
        """
        Retrieves relevant text chunks from the vector database based on user input.
        
        This method:
        1. Generates a vector query based on user input using the helper bot
        2. Searches multiple collections in the vector database
        3. Combines results and sorts them by relevance
        4. Limits results to the specified number of unique sources
        5. Cleans the text by removing footnote references
        6. Enriches the chunks with detailed metadata from ArangoDB
        7. Groups chunks by article title
        
        Parameters:
        -----------
        user_input : str
            The user query to search for relevant documents
        collections : list, optional
            List of collection names to search in (default: ["sci_articles", "other_documents"])
        n_results : int, optional
            Maximum number of results to return (default: 7)
        n_sources : int, optional
            Maximum number of unique document sources to include (default: 4)
        filter : bool, optional
            Whether to filter results by ArangoDB IDs (default: True)
            
        Returns:
        --------
        dict
            A dictionary of grouped chunks where:
            - Keys are article titles
            - Values are dictionaries containing:
                - 'article_number': A sequential number for the article
                - 'chunks': A list of chunk dictionaries, each containing:
                    - 'document': The document text
                    - 'metadata': The document metadata
                    - 'distance': The similarity distance (lower is better)
                    - 'article_number': The sequential number of the article
        """

        response = self.helperbot.generate(
            get_generate_vector_query_prompt(user_input, self.chat.role),
            format=QueryResponse.model_json_schema(),
        )
        print(response)
        print_yellow("RESPONSE:", response.content)
        query_response = QueryResponse.model_validate_json(response.content)
        query = query_response.query_to_vector_database
        print_purple(f"Query for vector DB:\n {query}")

        combined_chunks = []
        if collections:
            for collection in collections:

                if filter:
                    where_filter = {"_id": {"$in": self.arango_ids}}
                    chunks = self.get_chromadb().query(
                        query=query,
                        collection=collection,
                        n_results=n_results,
                        n_sources=n_sources,
                        where=where_filter,
                        max_retries=3,
                    )
                    for doc, meta, dist in zip(
                        chunks["documents"][0],
                        chunks["metadatas"][0],
                        chunks["distances"][0],
                    ):
                        combined_chunks.append(
                            {"document": doc, "metadata": meta, "distance": dist}
                        )
                
        combined_chunks.sort(key=lambda x: x["distance"])

        # Keep the best chunks according to n_sources
        sources = set()
        closest_chunks = []
        for chunk in combined_chunks:
            source_id = chunk["metadata"].get("_id", "no_id")
            if source_id not in sources:
                sources.add(source_id)
                closest_chunks.append(chunk)
            if len(sources) >= n_sources:
                break
        if len(closest_chunks) < n_results:
            remaining_chunks = [c for c in combined_chunks if c not in closest_chunks]
            closest_chunks.extend(remaining_chunks[: n_results - len(closest_chunks)])

        # Remove footnoot references like [\d+] from the text chunks
        for chunk in closest_chunks:
            chunk["document"] = re.sub(r"\[\d+\]", "", chunk["document"])

        # Fetch real metadata from Arango
        for chunk in closest_chunks:
            _id = chunk["metadata"].get("_id")
            if not _id:
                continue
            if _id.startswith("sci_articles"):
                arango_doc = self.base_arango.db.document(_id)
            else:
                arango_doc = self.user_arango.db.document(_id)
            if arango_doc:
                arango_metadata = arango_doc.get("metadata", {})
                # Possibly merge notes
                if "user_notes" in arango_doc:
                    arango_metadata["user_notes"] = arango_doc["user_notes"]
                chunk["metadata"] = arango_metadata

        # Group by article title
        grouped_chunks = {}
        article_number = 1
        for chunk in closest_chunks:
            title = chunk["metadata"].get("title", "No title")
            chunk["article_number"] = article_number
            if title not in grouped_chunks:
                grouped_chunks[title] = {
                    "article_number": article_number,
                    "chunks": [],
                }
                article_number += 1
            grouped_chunks[title]["chunks"].append(chunk)

        return grouped_chunks

    def answer_tool_call(self, response, user_input):
        """
        Process tool calls returned by the AI and execute the corresponding functions.
        
        This method evaluates tool calls in the AI response, executes the appropriate
        functions with the provided arguments, and collects the resulting responses.
        
        Parameters:
        -----------
        response : dict
            The AI response containing potential tool_calls to be executed
        user_input : str
            The original user query that will be passed to tool functions
            
        Returns:
        --------
        list
            A list of string responses generated from executing the tool calls.
            Returns an empty string if no tool calls are present.
            
        Notes:
        ------
        Supported tool functions include:
        - fetch_other_documents_tool: Retrieves non-scientific documents
        - fetch_science_articles_tool: Retrieves scientific articles
        - fetch_science_articles_and_other_documents_tool: Retrieves both types of documents
        - fetch_notes_tool: Retrieves user notes
        - conversational_response_tool: Generates a conversational response
        """
        bot_responses = []
        # This method returns / stores responses (no Streamlit calls)
        if not response.get("tool_calls"):
            return ""

        for tool in response.get("tool_calls"):
            function_name = tool.function.get("name")
            arguments = tool.function.arguments
            arguments["query"] = user_input

            if hasattr(self, function_name):
                print_purple("Function name:", function_name)
                if function_name in [
                    "fetch_other_documents_tool",
                    "fetch_science_articles_tool",
                    "fetch_science_articles_and_other_documents_tool",
                ]:
                    chunks = getattr(self, function_name)(**arguments)
                    bot_responses.append(self.generate_from_chunks(user_input, chunks))
                elif function_name == "fetch_notes_tool":
                    notes = getattr(self, function_name)()
                    bot_responses.append(self.generate_from_notes(user_input, notes))
                elif function_name == "conversational_response_tool":
                    response: OllamaMessage = getattr(self, function_name)(user_input)
                    print_green("Conversation response:", response)
                    bot_responses.append(response.content.strip('"'))
        return bot_responses

    # def process_user_input(self, user_input, content_attachment=None):
    #     # Add user message
    #     self.chat.add_message("user", user_input)
    #     print('content_attachment', content_attachment)
    #     if not content_attachment:
    #         prompt = get_tools_prompt(user_input)
    #         print('TOOLS PROMOT:', prompt)
    #         print_red('\nToolbot system message:', self.toolbot.system_message)
    #         response = self.toolbot.generate(prompt, tools=self.tools, stream=False)
    #         print_rainbow(response)
    #         if response.get("tool_calls"):
    #             bot_response = self.answer_tool_call(response, user_input)
    #         else:
    #             # Just respond directly
    #             bot_response = response.content.strip('"')
    #     else:
    #         # If there's an attachment, do something minimal
    #         bot_response = "Content attachment received (Base Bot)."

    #     # Add assistant message
    #     if self.chat.chat_history[-1]["role"] != "assistant":
    #         self.chat.add_message("assistant", bot_response)

    #     # Update in Arango
    #     self.chat.update_in_arango()
    #     return bot_response

    def generate_from_notes(self, user_input, notes):
        """
        Generate a response based on user input and a collection of notes.
        
        This method takes a user query and relevant notes, formats the notes into a string,
        creates a prompt with the formatted notes and user input, and generates a streamed response.
        
        Parameters
        ----------
        user_input : str
            The user's query or message to respond to
        notes : list of dict
            A list of note dictionaries, where each note has 'title' and 'content' keys
            
        Returns
        -------
        generator
            A generator that streams the AI-generated response
            
        Notes
        -----
        This method does not make any Streamlit calls and is safe to use outside of the Streamlit context.
        The notes are formatted with titles and content separated by horizontal rules.
        """
        # No Streamlit calls
        notes_string = ""
        for note in notes:
            notes_string += (
                f"\n# {note.get('title','No title')}\n{note.get('content','')}\n---\n"
            )
        prompt = get_chat_prompt(
            user_input, content_string=notes_string, role=self.chat.role
        )
        return self.chatbot.generate(prompt, stream=True)

    def generate_from_chunks(self, user_input, chunks):
        """
        Generate a response based on user input and retrieved document chunks.
        
        This method formats the retrieved document chunks into a structured string,
        combines it with the user's input in a prompt, and generates a streaming
        response using the chatbot.
        
        Parameters:
        -----------
        user_input : str
            The user's query or message to respond to.
        chunks : dict
            A dictionary containing document chunks organized by title.
            Expected structure:
            {
                "title1": {
                    "chunks": [
                        {
                            "document": "content...",
                            "metadata": {
                                "user_notes": "optional notes..."
                            }
                        },
                        ...
                    ],
                    "article_number": int
                },
                ...
            }
        
        Returns:
        --------
        generator
            A streaming generator of the chatbot's response.
        
        Notes:
        ------
        - This method does not make any Streamlit API calls.
        - User notes are included in the formatted content if available.
        - The formatted content includes titles, article numbers, and document text.
        """
        # No Streamlit calls
        chunks_string = ""
        for title, group in chunks.items():
            user_notes_string = ""
            if "user_notes" in group["chunks"][0]["metadata"]:
                notes = group["chunks"][0]["metadata"]["user_notes"]
                user_notes_string = f'\n\nUser notes:\n"""\n{notes}\n"""\n\n'
            docs = "\n(...)\n".join([c["document"] for c in group["chunks"]])
            chunks_string += f"\n# {title}\n## Article #{group['article_number']}\n{user_notes_string}{docs}\n---\n"
        prompt = get_chat_prompt(
            user_input, content_string=chunks_string, role=self.chat.role
        )
        return self.chatbot.generate(prompt, stream=True)

    def run(self):
        # Base Bot has no Streamlit run loop
        pass

    def get_notes(self):
        # Minimal note retrieval
        notes = self.user_arango.db.aql.execute(
            f'FOR doc IN notes FILTER doc.project == "{self.project.name if self.project else ""}" RETURN doc'
        )
        return list(notes)

    def fetch_science_articles_tool(self, query: str, n_documents: int = 6):
        """
        "Fetches information from scientific articles. Use this tool when the user is looking for information from scientific articles."

        Parameters:
            query (str): The search query to find relevant scientific articles.
            n_documents (int): How many documents to fetch. A complex query may require more documents. Min: 3, Max: 10.

        Returns:
            list: A list of chunks containing information from the fetched scientific articles.
        """
        print_purple("Query:", query)

        n_documents = int(n_documents)
        if n_documents < 3:
            n_documents = 3
        elif n_documents > 10:
            n_documents = 10
        return self.get_chunks(
            query, collections=["sci_articles"], n_results=n_documents
        )

    def fetch_other_documents_tool(self, query: str, n_documents: int = 6):
        """
        Fetches information from other documents based on the user's query.

        This method retrieves information from various types of documents such as reports, news articles, and other texts. It should be used only when it is clear that the user is not seeking scientific articles.

        Args:
            query (str): The search query provided by the user.
            n_documents (int): How many documents to fetch. A complex query may require more documents. Min: 2, Max: 10.

        Returns:
            list: A list of document chunks that match the query.
        """
        assert isinstance(self, Bot), "The first argument must be a Bot object."
        n_documents = int(n_documents)
        if n_documents < 2:
            n_documents = 2
        elif n_documents > 10:
            n_documents = 10
        return self.get_chunks(
            query,
            collections=[f"{self.username}__other_documents"],
            n_results=n_documents,
        )

    def fetch_science_articles_and_other_documents_tool(
        self, query: str, n_documents: int
    ):
        """
        Fetches information from both scientific articles and other documents.

        This method is often used when the user hasn't specified what kind of sources they are interested in.

        Args:
            query (str): The search query to fetch information for.
            n_documents (int): How many documents to fetch. A complex query may require more documents. Min: 3, Max: 10.

        Returns:
            list: A list of document chunks that match the search query.
        """
        assert isinstance(self, Bot), "The first argument must be a Bot object."
        n_documents = int(n_documents)
        if n_documents < 3:
            n_documents = 3
        elif n_documents > 10:
            n_documents = 10
        return self.get_chunks(
            query,
            collections=["sci_articles", f"{self.username}__other_documents"],
            n_results=n_documents,
        )

    def fetch_notes_tool(bot):
        """
        Fetches information from the project notes when you as an editor need context from the project notes to understand other information. ONLY use this together with other tools! No arguments needed.

        Returns:
            list: A list of notes.
        """
        assert isinstance(bot, Bot), "The first argument must be a Bot object."
        return bot.get_notes()

    def conversational_response_tool(self, query: str):
        """
        Generate a conversational response to a user's query.

        This method is designed to provide a short and conversational response without fetching additional data.
        It should be used ONLY when it is clear that the user is engaging in small talk (like saying 'hi').

        Args:
            query (str): The user's message to which the bot should respond.

        Returns:
            str: The generated conversational response.
        """
        query = f"""
        User message: "{query}". 
        Make your answer short and conversational. 
        Don't answer with anything you're not sure of! 
        """

        return self.chatbot.generate(query, stream=False)


class StreamlitBot(Bot):
    def __init__(
        self, username: str, chat: StreamlitChat = None, tools: list = None, **kwargs
    ):
        super().__init__(username=username, chat=chat, tools=tools, **kwargs)

        # For Streamlit, we can override or add attributes
        if "llm_chosen_backend" not in st.session_state:
            st.session_state["llm_chosen_backend"] = None

        self.chatbot.chosen_backend = st.session_state["llm_chosen_backend"]
        if not st.session_state["llm_chosen_backend"]:
            st.session_state["llm_chosen_backend"] = self.chatbot.chosen_backend

        settings = self.get_settings()
        if settings.get("use_reasoning_model", False):
            self.chatbot.model = self.chatbot.get_model("reasoning")

        print_rainbow(settings)
        print('MODEL', self.chatbot.model)

    def run(self):
        # Example Streamlit run loop
        title = (
            self.project.name
            if self.project
            else self.collection.name if self.collection else None
        )
        self.chat.show_title(title=title)
        self.chat.show_chat_history()
        if user_input := st.chat_input("Write your message here...", accept_file=True):
            text_input = user_input.text.replace('"""', "---")
            if len(user_input.files) > 1:
                st.error("Please upload only one file at a time.")
                return
            attached_file = user_input.files[0] if user_input.files else None

            content_attachment = None
            if attached_file:
                if attached_file.type == "application/pdf":
                    import fitz

                    pdf_document = fitz.open(
                        stream=attached_file.read(), filetype="pdf"
                    )
                    pdf_text = ""
                    for page_num in range(len(pdf_document)):
                        page = pdf_document.load_page(page_num)
                        pdf_text += page.get_text()
                    content_attachment = pdf_text
                elif attached_file.type in ["image/png", "image/jpeg"]:
                    self.chat.message_attachments = "image"
                    content_attachment = attached_file.read()
                    with st.chat_message(
                        "user", avatar=self.chat.get_avatar(role="user")
                    ):
                        st.image(content_attachment)

            with st.chat_message("user", avatar=self.chat.get_avatar(role="user")):
                st.write(text_input)

            if not self.chat.name:
                self.chat.set_name(text_input)
                self.chat.last_updated = datetime.now().isoformat()
                self.chat.saved = False
                self.user_arango.db.collection("chats").insert(
                    self.chat.to_dict(), overwrite=True, overwrite_mode="update"
                )

            self.process_user_input(text_input, content_attachment)

    def get_settings(self):
        return self.user_arango.db.document("settings/settings")

    def process_user_input(self, user_input, content_attachment=None):
        # We override to show messages in Streamlit instead of just storing
        self.chat.add_message("user", user_input)

        # Remove conversational response tool if there are more than 2 messages
        if len(self.chat.chat_history) > 2 and len(self.tools) > 1:
            for tool in self.tools:
                if tool.__name__ == "conversational_response_tool":
                    self.tools.remove(tool)
                    break
        if not content_attachment:
            prompt = get_tools_prompt(user_input)
            response = self.toolbot.generate(prompt, tools=self.tools, stream=False)
            if response.get("tool_calls"):
                bot_response = self.answer_tool_call(response, user_input)
            else:
                bot_response = response.content.strip('"')
            # with st.chat_message(
            #     "assistant", avatar=self.chat.get_avatar(role="assistant")
            # ):
            #     st.write(bot_response)
        else:
            with st.chat_message(
                "assistant", avatar=self.chat.get_avatar(role="assistant")
            ):
                with st.spinner("Reading the content..."):
                    if self.chat.message_attachments == "image":
                        prompt = get_chat_prompt(
                            user_input, role=self.chat.role, image_attachment=True
                        )
                        bot_resp = self.chatbot.generate(
                            prompt,
                            stream=False,
                            images=[content_attachment],
                            model="vision",
                        )
                        if isinstance(bot_resp, dict):
                            bot_resp = bot_resp.get("content", "")
                        elif isinstance(bot_resp, OllamaMessage):
                            bot_resp = bot_resp.content
                        st.write(bot_resp)
                        bot_response = bot_resp
                    else:
                        prompt = get_chat_prompt(
                            user_input,
                            content_attachment=content_attachment,
                            role=self.chat.role,
                        )
                        response = self.chatbot.generate(prompt, stream=True)
                        bot_response = st.write_stream(response)

        if self.chat.chat_history[-1]["role"] != "assistant":
            self.chat.add_message("assistant", bot_response)

        self.chat.update_in_arango()

    def answer_tool_call(
        self, response, user_input
    ):  #! This should be in the Base ChatBot?
        bot_responses = []
        tools_response = response.get("tool_calls", [])

        for tool in tools_response:
            function_name = tool.function.get("name")
            if len(tools_response) > 1:
                # Don't use conversational response tool if there are other tools
                if function_name == "conversational_response_tool":
                    continue
            arguments = tool.function.arguments
            arguments["query"] = user_input

            print("Function name:", function_name)
            with st.chat_message(
                "assistant", avatar=self.chat.get_avatar(role="assistant")
            ):
                if function_name in [
                    "fetch_other_documents_tool",
                    "fetch_science_articles_tool",
                    "fetch_science_articles_and_other_documents_tool",
                ]:
                    chunks = getattr(self, function_name)(**arguments)
                    response_text = self.generate_from_chunks(user_input, chunks)
                    # Separate thinking chunk and normal chunk
                    print_red("Model:", self.chatbot.model)

                    if self.chatbot.model == "reasoning":
                        bot_response = self.write_reasoning(response_text)

                    else:
                        bot_response = self.write_normal(response_text)
                    bot_responses.append(bot_response)

                    if chunks:
                        sources = "###### Sources:\n"
                        for title, group in chunks.items():
                            j = group["chunks"][0]["metadata"].get(
                                "journal", "No Journal"
                            )
                            d = group["chunks"][0]["metadata"].get(
                                "published_date", "No Date"
                            )
                            sources += f"[{group['article_number']}] **{title}** :gray[*{j}* ({d})]  \n"
                        st.markdown(sources)
                        bot_response += f"\n\n{sources}"
                    bot_responses.append(bot_response)

                elif function_name == "fetch_notes_tool":
                    notes = getattr(self, function_name)()
                    response_text = self.generate_from_notes(user_input, notes)
                    bot_responses.append(st.write_stream(response_text).strip('"'))

                elif function_name == "conversational_response_tool":
                    response_text = getattr(self, function_name)(user_input)
                    print(
                        "###",
                        response_text,
                    )
                    if self.chatbot.call_model == self.chatbot.get_model("reasoning"):
                        print_blue("REASONING MODEL!")
                        bot_response = self.write_reasoning(response_text).strip('"')
                    else:
                        if isinstance(response_text, OllamaMessage):
                            response_text = response_text.content
                        elif isinstance(response_text, dict):
                            response_text = response_text.get("content", "")

                        bot_response = self.write_normal(response_text).strip('"')

        return "\n\n".join(bot_responses)

    def write_reasoning(self, response_text):
        chunks_iter = iter(response_text)  # convert generator to iterator

        try:
            first_mode, first_text = next(chunks_iter)  # get first chunk
        except StopIteration:
            # no chunks at all
            first_mode, first_text = None, None
        print_purple("FIRST MODE:", first_mode, first_text)
        # if it's thinking, show that in an expander
        if first_mode == "thinking":
            with st.expander("How the bot has been reasoning"):
                st.write(first_text.replace("<think>", "").replace("</think>", ""))

            # define a generator for the rest
            def rest_gen():
                for _, text in chunks_iter:
                    yield text

            bot_response = st.write_stream(rest_gen())
            return bot_response

        else:

            def full_gen():
                if first_mode:
                    yield (first_mode, first_text)
                for mode, text in chunks_iter:
                    yield (mode, text)

            bot_response = st.write_stream(full_gen()).strip('"')

    def write_normal(self, response_text):
        chunks_iter = iter(response_text)  # convert generator to iterator

        def full_gen():
            for chunk in chunks_iter:
                if isinstance(chunk, tuple) and len(chunk) == 2:
                    _, text = chunk
                    yield text
                else:
                    yield chunk

        bot_response = st.write_stream(full_gen()).strip('"')
        return bot_response

    def generate_from_notes(self, user_input, notes):
        with st.spinner("Reading project notes..."):
            return super().generate_from_notes(user_input, notes)

    def generate_from_chunks(self, user_input, chunks):
        # For reading articles with a spinner
        magazines = set()
        for group in chunks.values():
            j = group["chunks"][0]["metadata"].get("journal", "No Journal")
            magazines.add(f"*{j}*")
        s = (
            f"Reading articles from {', '.join(list(magazines)[:-1])} and {list(magazines)[-1]}..."
            if len(magazines) > 1
            else "Reading articles..."
        )
        with st.spinner(s):
            return super().generate_from_chunks(user_input, chunks)

    def sidebar_content(self):
        with st.sidebar:
            st.write("---")
            st.markdown(f'#### {self.chat.name if self.chat.name else ""}')
            st.button("Delete this chat", on_click=self.delete_chat)

    def delete_chat(self):
        self.user_arango.db.collection("chats").delete_match(
            filters={"name": self.chat.name}
        )
        self.chat = Chat()

    def get_notes(self):
        # We can show a spinner or messages too
        with st.spinner("Fetching notes..."):
            return super().get_notes()


class EditorBot(StreamlitBot):
    def __init__(self, username: str, chat: Chat, **kwargs):
        super().__init__(username=username, chat=chat, **kwargs)
        self.role = "Editor"
        self.tools = [self.fetch_notes_tool, self.fetch_other_documents_tool]
        # self.chatbot = LLM(
        #     system_message=get_editor_prompt(kwargs.get("project")),
        #     messages=self.chat.chat_history2bot(),
        #     chosen_backend=kwargs.get("chosen_backend"),
        # )
        print_purple("MODEL FOR EDITOR BOT:", self.chatbot.model)


class ResearchAssistantBot(StreamlitBot):
    def __init__(self, username: str, chat: Chat, **kwargs):
        super().__init__(username=username, chat=chat, **kwargs)
        self.role = "Research Assistant"
        # self.chatbot = LLM(
        #     system_message=get_assistant_prompt(),
        #     temperature=0.1,
        #     messages=self.chat.chat_history2bot(),
        # )
        self.tools = [
            self.fetch_science_articles_tool,
            self.fetch_science_articles_and_other_documents_tool,
            self.conversational_response_tool,
        ]


class PodBot(StreamlitBot):
    """Two LLM agents construct a conversation using material from science articles."""

    def __init__(
        self,
        username: str,
        chat: Chat,
        subject: str,
        instructions: str = None,
        **kwargs,
    ):
        super().__init__(username=username, chat=chat, **kwargs)
        self.subject = subject
        self.instructions = instructions
        self.guest_name = kwargs.get("name_guest", "Merit")
        self.hostbot = HostBot(
            Chat(username=self.username, role="Host"),
            subject,
            username,
            instructions=instructions,
            **kwargs,
        )
        self.guestbot = GuestBot(
            Chat(username=self.username, role="Guest"),
            subject,
            username,
            name_guest=self.guest_name,
            **kwargs,
        )

    def run(self):

        notes = self.get_notes()
        notes_string = ""
        if self.instructions:
            instructions_string = f'''
            These are the instructions for the podcast from the producer:
            """
            {self.instructions}
            """
            '''
        else:
            instructions_string = ""

        for note in notes:
            notes_string += f"\n# {note['title']}\n{note['content']}\n---\n"
        a = f'''You will make a podcast interview with {self.guest_name}, an expert on "{self.subject}". 
        {instructions_string}
        Below are notes on the subject that you can use to ask relevant questions:
        """
        {notes_string}
        """
        Say hello to the expert and start the interview. Remember to keep the interview to the subject of {self.subject} throughout the conversation.
        '''

        # Stop button for the podcast
        with st.sidebar:
            stop = st.button("Stop podcast", on_click=self.stop_podcast)

        while st.session_state["make_podcast"]:
            # Stop the podcast if there are more than 14 messages in the chat
            self.chat.show_chat_history()
            if len(self.chat.chat_history) == 14:
                result = self.hostbot.generate(
                    "The interview has ended. Say thank you to the expert and end the conversation."
                )
                self.chat.add_message("Host", result)
                with st.chat_message(
                    "assistant", avatar=self.chat.get_avatar(role="assistant")
                ):
                    st.write(result.strip('"'))
                st.stop()

            _q = self.hostbot.toolbot.generate(
                query=f"{self.guest_name} has answered: {a}. You have to choose a tool to help the host continue the interview.",
                tools=self.hostbot.tools,
                temperature=0.6,
                stream=False,
            )
            if "tool_calls" in _q:
                q = self.hostbot.answer_tool_call(_q, a)
            else:
                q = _q

            self.chat.add_message("Host", q)

            _a = self.guestbot.toolbot.generate(
                f'The podcast host has asked: "{q}" Choose a tool to help the expert answer with relevant facts and information.',
                tools=self.guestbot.tools,
            )
            if "tool_calls" in _a:
                print_yellow("Tool call response (guest)", _a)
                print_yellow(self.guestbot.chat.role)
                a = self.guestbot.answer_tool_call(_a, q)
            else:
                a = _a
            self.chat.add_message("Guest", a)

            self.update_session_state()

    def stop_podcast(self):
        st.session_state["make_podcast"] = False
        self.update_session_state()
        self.chat.show_chat_history()


class HostBot(StreamlitBot):
    def __init__(
        self, chat: Chat, subject: str, username: str, instructions: str, **kwargs
    ):
        super().__init__(chat=chat, username=username, **kwargs)
        self.chat.role = kwargs.get("role", "Host")
        self.tools = [self.fetch_notes_tool, self.conversational_response_tool]
        self.instructions = instructions
        self.llm = LLM(
            system_message=f'''
            You are the host of a podcast and an expert on {subject}. You will ask one question at a time about the subject, and then wait for the guest to answer. 
            Don't ask the guest to talk about herself/himself, only about the subject.
            Make your questions short and clear, only if necessary add a brief context to the question.
            These are the instructions for the podcast from the producer:
            """
            {self.instructions}
            """
            If the experts' answer is complicated, try to make a very brief summary of it for the audience to understand. You can also ask follow-up questions to clarify the answer, or ask for examples.
            ''',
            messages=self.chat.chat_history2bot(),
        )
        self.toolbot = LLM(
            temperature=0,
            system_message="""
            You are assisting a podcast host in asking questions to an expert. 
            Choose one or many tools to use in order to assist the host in asking relevant questions. 
            Often "conversational_response_tool" is enough, but sometimes project notes are needed. 
            Make sure to read the description of the tools carefully!""",
            chat=False,
            model="tools",
        )

    def generate(self, query):
        return self.llm.generate(query)


class GuestBot(StreamlitBot):
    def __init__(self, chat: Chat, subject: str, username: str, **kwargs):
        super().__init__(chat=chat, username=username, **kwargs)
        self.chat.role = kwargs.get("role", "Guest")
        self.tools = [
            self.fetch_notes_tool,
            self.fetch_science_articles_tool,
        ]

        self.llm = LLM(
            system_message=f"""
            You are {kwargs.get('name', 'Merit')}, an expert on {subject}. 
            Today you are a guest in a podcast about {subject}. A host will ask you questions about the subject and you will answer by using scientific facts and information.
            When answering, don't say things like "based on the documents" or alike, as neither the host nor the audience can see the documents. Act just as if you were talking to someone in a conversation.
            Try to be concise when answering, and remember that the audience of the podcast is not expert on the subject, so don't complicate things too much.
            It's very important that you answer in a "spoken" way, as if you were talking to someone in a conversation. That means you should avoid using scientific jargon and complex terms, too many figures or abstract concepts. 
            Lists are also not recommended, instead use "for the first reason", "secondly", etc.
            Instead, use "..." to indicate a pause, "-" to indicate a break in the sentence, as if you were speaking. 
            """,
            messages=self.chat.chat_history2bot(),
        )
        self.toolbot = LLM(
            temperature=0,
            system_message=f"You are an assistant to an expert on {subject}. Choose one or many tools to use in order to assist the expert in answering questions. Make sure to read the description of the tools carefully.",
            chat=False,
            model="tools",
        )

    def generate(self, query):
        return self.llm.generate(query)