Py.Cafe

"""
# Chatbot

A way to create a chatbot using the OpenAI Assistants API with streaming functionality.
This code uses the combined DraftGeneratorRevisorTranslator assistant configuration.
"""

import os
import datetime
from typing import List, cast
from typing_extensions import TypedDict

from openai import AsyncOpenAI
from openai.types.chat import ChatCompletionMessageParam

import solara
import solara.lab

# Monkey-patch solara.v.TextField to change the default label.
original_TextField = solara.v.TextField

def custom_TextField(*args, **kwargs):
    if kwargs.get("label") == "Type a message...":
        kwargs["label"] = "Typ een bericht..."
    return original_TextField(*args, **kwargs)

solara.v.TextField = custom_TextField

# Define our message dictionary type.
class MessageDict(TypedDict):
    role: str  # "user", "assistant", or "system"
    content: str

# Reactive global messages list.
messages: solara.Reactive[List[MessageDict]] = solara.reactive([])

# Get our OpenAI API key.
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")

openai = AsyncOpenAI(api_key=OPENAI_API_KEY) if OPENAI_API_KEY else None

def no_api_key_message():
    messages.value = [{
        "role": "assistant",
        "content": "No OpenAI API key found. Please set your OPENAI_API_KEY.",
    }]

# A simple helper function to get the current time as a string.
def get_current_time_nl() -> str:
    return datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")

# System instructions (metaprompt) with time-of-day context guidance.
system_instructions = """
You are DraftGeneratorRevisorTranslator, a specialized therapy-assistant that generates a first-draft therapeutic response and revises that response into a final, culturally and linguistically refined Dutch message. You are pre‑configured with a detailed default condensation document (described below) that contains essential theoretical background, intervention strategies, and style guidelines.

--- Default Condensation Document ---
1. Client & Context Overview  
    • Client Demographics & Presentation:  
         – Gender, age range: female, middle-aged.  
         – Primary symptoms: Nighttime panic attacks, rapid heartbeat, shallow breathing, acute distress.  
         – Somatic cues: Tension in the chest, tight throat, fluttering sensations.  
    • Background Information:  
         – History of trauma or insecure attachment.  
         – Emotional state: Predominantly high threat activation with signs of dissociation.
2. Theoretical Framework Summary  
    • Integrative Therapeutic Model:  
         – Systemic: Consider family and relational influences.  
         – Body-Oriented: The body “speaks” through nonverbal signals; an embodied approach addresses trauma effectively.  
         – Sensorimotor Processing: Habitual tension may persist after the original threat.  
         – Attachment Dynamics: Early disruptions lead to fixed defensive patterns.
    • Intervention Strategies:  
         – Immediate Self-Regulation (e.g., “4-7-8” breathing, sensory grounding, gentle self-touch).  
         – Mindfulness Techniques: Guide attention to bodily sensations in real time.  
         – Gentle Distraction: E.g., listening to nature sounds.
3. Tone, Language & Style  
    • Use warm, empathetic, and non-judgmental language.  
    • Ask open-ended questions (e.g., “What do you notice in your body right now?”).  
    • Use accessible metaphors (e.g., “imagine taking off a heavy coat”).
4. Response Structure  
    • Final output entirely in Dutch.  
    • Begin with validation of the client's distress.  
    • Provide step-by-step micro-interventions and targeted questions.
    
--- Time-of-Day Context Handling ---
You receive a separate system message before the client's input that states "Current NL time: YYYY-MM-DD HH:MM:SS".  
• If the client's input is time-sensitive (e.g., "I'm panicking at 2 a.m."), integrate this NL time context into your reasoning and recommendations.  
• If not, consider it supplemental context.

--- Your Task ---
1. Integrate the default document into your response generation.
2. Note the separate system message indicating the current NL time.
3. Based on the client's input (e.g., "I'm panicking at 2 a.m. and I can't catch my breath"), generate a full, empathetic therapeutic response in Dutch that includes tailored interventions and, if relevant, the current time context.
4. Revise and produce the final output in Dutch without including meta-commentary.

End of Instructions.
"""

# Ensure the system message is the first entry.
if not messages.value or messages.value[0]["role"] != "system":
    messages.value = [{"role": "system", "content": system_instructions}]

@solara.lab.task
async def promt_ai(message: str):
    if openai is None:
        no_api_key_message()
        return

    # Append a system message with the current NL time.
    messages.value.append({"role": "system", "content": f"Current NL time: {get_current_time_nl()}"})
    
    # Append the user's message.
    messages.value.append({"role": "user", "content": message})
    
    # Call the Assistants API with the assistant id.
    response = await openai.assistants.chat.completions.create(
        assistant="asst_8KbWqsPFUWZfu5KFBu90x4cX",
        #model="gpt-4-1106-preview",
        messages=cast(List[ChatCompletionMessageParam], messages.value),
        stream=True,
        temperature=0.5,
    )
    
    # Append an empty assistant message to hold the streamed response.
    messages.value.append({"role": "assistant", "content": ""})
    
    # Stream the response and update the assistant's message.
    async for chunk in response:
        if chunk.choices[0].finish_reason == "stop":
            return
        delta = chunk.choices[0].delta.content
        assert delta is not None
        updated_message: MessageDict = {"role": "assistant", "content": messages.value[-1]["content"] + delta}
        messages.value = [*messages.value[:-1], updated_message]

@solara.component
def Page():
    with solara.Column(style={"width": "100%", "height": "50vh"}):
        with solara.lab.ChatBox():
            # Do not display system messages.
            for item in messages.value:
                if item["role"] == "system":
                    continue
                with solara.lab.ChatMessage(
                    user=item["role"] == "user",
                    avatar=False,
                    name="Kompassie" if item["role"] == "assistant" else "Gebruiker",
                    color="rgba(0,0,0,0.06)" if item["role"] == "assistant" else "#ff991f",
                    avatar_background_color="primary" if item["role"] == "assistant" else None,
                    border_radius="20px",
                ):
                    solara.Markdown(item["content"])
        if promt_ai.pending:
            solara.Text("Ik denk erover na...", style={"font-size": "1rem", "padding-left": "20px"})
            solara.ProgressLinear()
        solara.lab.ChatInput(send_callback=promt_ai, disabled_send=promt_ai.pending, autofocus=True).key("input")