I built Perplexity from scratch, but mine is better. | Tutorial

Step-by-Step Tutorial: Building "Bagoodex Web Search"

This tutorial provides a structured walkthrough to create "Bagoodex Web Search," an open-source Perplexity-like app built with Python, Gradio, and external APIs. We'll be using the AI/ML API for AI capabilities.

AI/ML API

AI/ML API is a game-changing platform for developers and SaaS entrepreneurs looking to integrate cutting-edge AI capabilities into their products. It offers a single point of access to over 200 state-of-the-art AI models, covering everything from NLP to computer vision.

Key Features for Developers:

• Extensive Model Library: 200+ pre-trained models for rapid prototyping and deployment. 📚
• Customization Options: Fine-tune models to fit your specific use case. 🎯
• Developer-Friendly Integration: RESTful APIs and SDKs for seamless incorporation into your stack. 🛠️
• Serverless Architecture: Focus on coding, not infrastructure management. ☁️

Get Started for FREE.

Deep Dive into AI/ML API Documentation (very detailed, can’t agree more).

---

Step 1: Setting Up the Environment

1.1 Create a Virtual Environment:

python -m venv .venv
source .venv/bin/activate

1.2 Install Dependencies: Create and populate [requirements.txt] with:

openai
gradio
python-dotenv
requests
pytube

Then install them:

pip install -r requirements.txt

1.3 Environment Variables: Create a .env file with your API keys:

AIML_API_KEY=your_api_key
GOOGLE_MAPS_API_KEY=your_google_maps_api_key

Here's a brief tutorial: How to get API Key from AI/ML API. Quick step-by-step tutorial with screenshots for better understanding.

1.4 Git Ignore: Add .gitignore:

.env
.venv
__pycache__
*.pyc
.DS_Store

---

Step 2: Project Structure

Your final project directory should look like:

Bagoodex_Web_Search/
├── .env
├── .gitignore
├── requirements.txt
├── app.py
├── bagoodex_client.py
├── helpers.py
├── prompts.py
└── r_types.py

---

Step 3: Key Files Explained

3.1 [bagoodex_client.py]

• Implements API interactions with bagoodex and GPT services.

• Import necessary modules:
  

import os
import requests
from openai import OpenAI
from dotenv import load_dotenv
from r_types import ChatMessage
from prompts import SYSTEM_PROMPT_BASE, SYSTEM_PROMPT_MAP
from typing import List

• Load environment variables and set up the API client:

load_dotenv()
API_KEY = os.getenv("AIML_API_KEY")
API_URL = "https://api.aimlapi.com"

• Define the BagoodexClient class:

class BagoodexClient:
    def __init__(self, api_key=API_KEY, api_url=API_URL):
        self.api_key = api_key
        self.api_url = api_url
        self.client = OpenAI(base_url=self.api_url, api_key=self.api_key)

• Includes methods:
  • complete_chat(): Handles general chat interactions.

  def complete_chat(self, query):
        """
        Calls the standard chat completion endpoint using the provided query.
        Returns the generated followup ID and the text response.
        """
        response = self.client.chat.completions.create(
            model="bagoodex/bagoodex-search-v1",
            messages=[
                ChatMessage(role="user", content=SYSTEM_PROMPT_BASE), 
                ChatMessage(role="user", content=query)
            ],
        )
        followup_id = response.id  # the unique ID for follow-up searches
        answer = response.choices[0].message.content
        return followup_id, answer

• base_qna(): Handles basic Q&A interactions. Basically we'll use this for follow-up questions. It's pretty reusable. We should pass the different system prompts based on our use case.

    def base_qna(self, messages: List[ChatMessage], system_prompt=SYSTEM_PROMPT_BASE):
        response = self.client.chat.completions.create(
            model="gpt-4o",
            messages=[
                ChatMessage(role="user", content=system_prompt), 
                *messages
            ],
        )
        return response.choices[0].message.content

• Retrieves IDs for fetching follow-up resources (links, images, videos, maps).

    def get_links(self, followup_id):
        headers = {"Authorization": f"Bearer {self.api_key}"}
        params = {"followup_id": followup_id}
        response = requests.get(
            f"{self.api_url}/v1/bagoodex/links", headers=headers, params=params
        )
        return response.json()

    def get_images(self, followup_id):
        headers = {"Authorization": f"Bearer {self.api_key}"}
        params = {"followup_id": followup_id}
        response = requests.get(
            f"{self.api_url}/v1/bagoodex/images", headers=headers, params=params
        )
        return response.json()

    def get_videos(self, followup_id):
        headers = {"Authorization": f"Bearer {self.api_key}"}
        params = {"followup_id": followup_id}
        response = requests.get(
            f"{self.api_url}/v1/bagoodex/videos", headers=headers, params=params
        )
        return response.json()

    def get_local_map(self, followup_id):
        headers = {"Authorization": f"Bearer {self.api_key}"}
        params = {"followup_id": followup_id}
        response = requests.get(
            f"{self.api_url}/v1/bagoodex/local-map", headers=headers, params=params
        )
        return response.json()

    def get_knowledge(self, followup_id):
        headers = {"Authorization": f"Bearer {self.api_key}"}
        params = {"followup_id": followup_id}
        response = requests.get(
            f"{self.api_url}/v1/bagoodex/knowledge", headers=headers, params=params
        )
        return response.json()

Note: First, you must first call the standard chat completion endpoint complete_chat() with your query. The chat completion endpoint returns an ID, which must then be passed as the sole input parameter followup_id to the bagoodex/links, bagoodex/images, bagoodex/videos, bagoodex/local-map and bagoodex/knowledge endpoints.

3.2 [app.py]

• Import necessary modules:

import os
import gradio as gr
from bagoodex_client import BagoodexClient
from r_types import ChatMessage
from prompts import (
    SYSTEM_PROMPT_FOLLOWUP, 
    SYSTEM_PROMPT_MAP, 
    SYSTEM_PROMPT_BASE, 
    SYSTEM_PROMPT_KNOWLEDGE_BASE
)
from helpers import (
    embed_video,
    format_links,
    embed_google_map,
    format_knowledge,
    format_followup_questions
)

• Initialize the BagoodexClient:

client = BagoodexClient()

• Central application logic.

# ----------------------------
# Chat & Follow-up Functions
# ----------------------------
def chat_function(message, history, followup_state, chat_history_state):
    """
    Process a new user message.
    Appends the message and response to the conversation,
    and retrieves follow-up questions.
    """
    # complete_chat returns a new followup id and answer
    followup_id_new, answer = client.complete_chat(message)
    # Update conversation history (if history is None, use an empty list)
    if history is None:
        history = []
    updated_history = history + [ChatMessage({"role": "user", "content": message}),
                                ChatMessage({"role": "assistant", "content": answer})]
    # Retrieve follow-up questions using the updated conversation
    followup_questions_raw = client.base_qna(
        messages=updated_history, system_prompt=SYSTEM_PROMPT_FOLLOWUP
    )
    # Format them using the helper
    followup_md = format_followup_questions(followup_questions_raw)
    return answer, followup_id_new, updated_history, followup_md

def handle_followup_click(question, followup_state, chat_history_state):
    """
    When a follow-up question is clicked, send it as a new message.
    """
    if not question:
        return chat_history_state, followup_state, ""
    # Process the follow-up question via complete_chat
    followup_id_new, answer = client.complete_chat(question)
    updated_history = chat_history_state + [ChatMessage({"role": "user", "content": question}),
                                            ChatMessage({"role": "assistant", "content": answer})]
    # Get new follow-up questions
    followup_questions_raw = client.base_qna(
        messages=updated_history, system_prompt=SYSTEM_PROMPT_FOLLOWUP
    )
    followup_md = format_followup_questions(followup_questions_raw)
    return updated_history, followup_id_new, followup_md

• Next setup Local map and Knowledge base functions.

def handle_local_map_click(followup_state, chat_history_state):
    """
    On local map click, try to get a local map.
    If issues occur, fall back to using the SYSTEM_PROMPT_MAP.
    """
    if not followup_state:
        return chat_history_state
    try:
        result = client.get_local_map(followup_state)

        if result:
            map_url = result.get('link', '')
            # Use helper to produce an embedded map iframe
            html = embed_google_map(map_url)

            # Fall back: use the base_qna call with SYSTEM_PROMPT_MAP
            result = client.base_qna(
                messages=chat_history_state, system_prompt=SYSTEM_PROMPT_MAP
            )
            # Assume result contains a 'link' field
            html = embed_google_map(result.get('link', ''))
        new_message = ChatMessage({"role": "assistant", "content": html})
        return chat_history_state + [new_message]
    except Exception:
        return chat_history_state

def handle_knowledge_click(followup_state, chat_history_state):
    """
    On knowledge base click, fetch and format knowledge content.
    """
    if not followup_state:
        return chat_history_state

    try:
        print('trying to get knowledge')
        result = client.get_knowledge(followup_state)
        knowledge_md = format_knowledge(result)

        if knowledge_md == 0000:
            print('falling back to base_qna')
            # Fall back: use the base_qna call with SYSTEM_PROMPT_KNOWLEDGE_BASE
            result = client.base_qna(
                messages=chat_history_state, system_prompt=SYSTEM_PROMPT_KNOWLEDGE_BASE
            )
            knowledge_md = format_knowledge(result)
        new_message = ChatMessage({"role": "assistant", "content": knowledge_md})
        return chat_history_state + [new_message]
    except Exception:
        return chat_history_state

• Advanced search functions.

# ----------------------------
# Advanced Search Functions
# ----------------------------
def perform_image_search(followup_state):
    if not followup_state:
        return []
    result = client.get_images(followup_state)
    # For images we simply return a list of original URLs
    return [item.get("original", "") for item in result]

def perform_video_search(followup_state):
    if not followup_state:
        return "<p>No followup ID available.</p>"
    result = client.get_videos(followup_state)
    # Use the helper to produce the embed iframes (supports multiple videos)
    return embed_video(result)

def perform_links_search(followup_state):
    if not followup_state:
        return gr.Markdown("No followup ID available.")
    result = client.get_links(followup_state)
    return format_links(result)

• Uses Gradio for UI. Settign up CSS.

# ----------------------------
# UI Build
# ----------------------------
css = """
#chatbot {
    height: 100%;
}
h1, h2, h3, h4, h5, h6 {
    text-align: center;
    display: block;
}
"""

• Handles chat, follow-up interactions, and advanced search features (images, videos, links).

with gr.Blocks(css=css, fill_height=True) as demo:
    gr.Markdown("""
        ## like perplexity, but with less features. 
        #### built by [@abdibrokhim](https://yaps.gg).
    """)

    # State variables to hold followup ID and conversation history, plus follow-up questions text
    followup_state = gr.State(None)
    chat_history_state = gr.State([])  # holds conversation history as a list of messages
    followup_md_state = gr.State("")     # holds follow-up questions as Markdown text

    with gr.Row():
        with gr.Column(scale=3):
            with gr.Row():
                btn_local_map = gr.Button("Local Map Search (coming soon...)", variant="secondary", size="sm", interactive=False)
                btn_knowledge = gr.Button("Knowledge Base (coming soon...)", variant="secondary", size="sm", interactive=False)
            # The ChatInterface now uses additional outputs for both followup_state and conversation history,
            # plus follow-up questions Markdown.
            chat = gr.ChatInterface(
                fn=chat_function,
                type="messages",
                additional_inputs=[followup_state, chat_history_state],
                additional_outputs=[followup_state, chat_history_state, followup_md_state],
            )
            # Button callbacks to append local map and knowledge base results to chat
            btn_local_map.click(
                fn=handle_local_map_click,
                inputs=[followup_state, chat_history_state],
                outputs=chat.chatbot
            )
            btn_knowledge.click(
                fn=handle_knowledge_click,
                inputs=[followup_state, chat_history_state],
                outputs=chat.chatbot
            )

            # Radio-based follow-up questions
            followup_radio = gr.Radio(
                choices=[], 
                label="Follow-up Questions (select one and click 'Send Follow-up')"
            )
            btn_send_followup = gr.Button("Send Follow-up")

            # When the user clicks "Send Follow-up", the selected question is passed
            # to handle_followup_click
            btn_send_followup.click(
                fn=handle_followup_click,
                inputs=[followup_radio, followup_state, chat_history_state],
                outputs=[chat.chatbot, followup_state, followup_md_state]
            )

            # Update the radio choices when followup_md_state changes
            def update_followup_radio(md_text):
                """
                Parse Markdown lines to extract questions starting with '- '.
                """
                lines = md_text.splitlines()
                questions = []
                for line in lines:
                    if line.startswith("- "):
                        questions.append(line[2:])
                return gr.update(choices=questions, value=None)

            followup_md_state.change(
                fn=update_followup_radio,
                inputs=[followup_md_state],
                outputs=[followup_radio]
            )

        with gr.Column(scale=1):
            gr.Markdown("### Advanced Search Options")
            with gr.Column(variant="panel"):
                btn_images = gr.Button("Search Images")
                btn_videos = gr.Button("Search Videos")
                btn_links = gr.Button("Search Links")
                gallery_output = gr.Gallery(label="Image Results", columns=2)
                video_output = gr.HTML(label="Video Results")  # HTML for embedded video iframes
                links_output = gr.Markdown(label="Links Results")
                btn_images.click(
                    fn=perform_image_search,
                    inputs=[followup_state],
                    outputs=[gallery_output]
                )
                btn_videos.click(
                    fn=perform_video_search,
                    inputs=[followup_state],
                    outputs=[video_output]
                )
                btn_links.click(
                    fn=perform_links_search,
                    inputs=[followup_state],
                    outputs=[links_output]
                )
    demo.launch()

Questions you may consider to ask:

how to make slingshot?
who created light (e.g., electricity) Tesla or Edison in quick short?

3.2 [helpers.py]

• Utility functions for formatting results:
  • import the necessary modules:

      from dotenv import load_dotenv
    import os
    import gradio as gr
    import urllib.parse
    import re
    from pytube import YouTube
    from typing import List, Optional, Dict
    from r_types import (
        SearchVideosResponse,
        SearchImagesResponse,
        SearchLinksResponse,
        LocalMapResponse,
        KnowledgeBaseResponse
    )
    import json

• embed_video() for YouTube videos

    def get_video_id(url: str) -> Optional[str]:
        """
        Safely retrieve the YouTube video_id from a given URL using pytube.
        Returns None if the URL is invalid or an error occurs.
        """
        if not url:
            return None

        try:
            yt = YouTube(url)
            return yt.video_id
        except Exception:
            # If the URL is invalid or pytube fails, return None
            return None

    def embed_video(videos: List[SearchVideosResponse]) -> str:
        """
        Given a list of video data (with 'link' and 'title'),
        returns an HTML string of embedded YouTube iframes.
        """
        if not videos:
            return "<p>No videos found.</p>"

        # Collect each iframe snippet
        iframes = []
        for video in videos:
            url = video.get("link", "")
            video_id = get_video_id(url)
            if not video_id:
                # Skip invalid or non-parsable links
                continue

            title = video.get("title", "").replace('"', '\\"')  # Escape quotes
            iframe = f"""
            <iframe 
                width="560" 
                height="315" 
                src="https://www.youtube.com/embed/{video_id}" 
                title="{title}" 
                frameborder="0" 
                allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" 
                allowfullscreen>
            </iframe>
            """
            iframes.append(iframe)

        # If no valid videos after processing, return a fallback message
        if not iframes:
            return "<p>No valid YouTube videos found.</p>"

        # Join all iframes into one HTML string
        return "\n".join(iframes)

• format_links() for links

  def format_links(links) -> str:
    """
    Convert a list of {'title': str, 'link': str} objects
    into a bulleted Markdown string with clickable links.
    """
    if not links:
        return "No links found."

    links_md = "**Links:**\n"
    for url in links:
        title = url.rstrip('/').split('/')[-1]
        links_md += f"- [{title}]({url})\n"
    return links_md

• embed_google_map() for maps

  def embed_google_map(map_url: str) -> str:
    """
    Extracts a textual location from the given Google Maps URL
    and returns an embedded Google Map iframe for that location.
    Assumes you have a valid API key in place of 'YOUR_API_KEY'.
    """
    load_dotenv()
    GOOGLE_MAPS_API_KEY = os.getenv("GOOGLE_MAPS_API_KEY")

    if not map_url:
        return "<p>Invalid Google Maps URL.</p>"

    # Attempt to extract "San+Francisco,+CA" from the URL
    match = re.search(r"/maps/place/([^/]+)", map_url)
    if not match:
        return "Invalid Google Maps URL. Could not extract location."

    location_text = match.group(1)
    # Remove query params or additional slashes from the captured group
    location_text = re.split(r"[/?]", location_text)[0]

    # URL-encode location to avoid issues with special characters
    encoded_location = urllib.parse.quote(location_text, safe="")

    embed_html = f"""
    <iframe
      width="600"
      height="450"
      style="border:0"
      loading="lazy"
      allowfullscreen
      src="https://www.google.com/maps/embed/v1/place?key={GOOGLE_MAPS_API_KEY}&q={encoded_location}">
    </iframe>
    """
    return embed_html

• format_knowledge() for knowledge base info

  def format_knowledge(raw_result: str) -> str:
    """
    Given a dictionary of knowledge data (e.g., about a person),
    produce a Markdown string summarizing that info.
    """

    if not raw_result:
        return 0000

    # Clean up the raw JSON string
    clean_json_str = cleanup_raw_json(raw_result)
    print('Knowledge Data: ', clean_json_str)

    try:
        # Parse the cleaned JSON string
        result = json.loads(clean_json_str)
        title = result.get("title", "...")
        type_ = result.get("type", "...")
        born = result.get("born", "...")
        died = result.get("died", "...")

        content = f"""
    **{title}**  
    Type: {type_}  
    Born: {born}  
    Died: {died}
        """
        return content
    except json.JSONDecodeError:
        return "Error: Failed to parse knowledge data."

• Set up format_followup_questions() function that formats follow-up questions

    def format_followup_questions(raw_questions: str) -> str:
        """
        Extracts and formats follow-up questions from a raw JSON-like string.

        The input string may contain triple backticks ("`\``json ... `\``") which need to be removed before parsing.

        Expected input format:
        "`\``json"
        {
            "followup_question": [
                "What materials are needed to make a slingshot?", 
                "How to make a slingshot more powerful?"
            ]
        }
        "`\``"

        Returns a Markdown-formatted string with the follow-up questions.
        """

        if not raw_questions:
            return "No follow-up questions available."

        # Clean up the raw JSON string
        clean_json_str = cleanup_raw_json(raw_questions)

        try:
            # Parse the cleaned JSON string
            questions_dict = json.loads(clean_json_str)

            # Ensure the expected key exists
            followup_list = questions_dict.get("followup_question", [])

            if not isinstance(followup_list, list) or not followup_list:
                return "No follow-up questions available."

            # Format the questions into Markdown
            questions_md = "### Follow-up Questions\n\n"
            for question in followup_list:
                questions_md += f"- {question}\n"

            return questions_md

        except json.JSONDecodeError:
            return "Error: Failed to parse follow-up questions."

• cleanup_row_json() function to clean up raw JSON strings

    def cleanup_raw_json(raw_json: str) -> str:
        """
        Remove triple backticks and 'json' from the beginning and end of a raw JSON string.
        """
        return re.sub(r"`\``json|``\`", "", raw_json).strip()

3.3 [prompts.py]

• Contains system prompts for various tasks: For example:

SYSTEM_PROMPT_BASE - For general chat interactions.

SYSTEM_PROMPT_BASE = """
    ######SYSTEM INIATED######
    You will be given a conversation chat (e.g., text/ paragraph).
    Answer the given conversation chat with a relevant response.

    ######NOTE######
    Be nice and polite in your responses!
    ######SYSTEM SHUTDOWN######
"""

SYSTEM_PROMPT_MAP - For providing places based on the given content.

SYSTEM_PROMPT_MAP = """
    ######SYSTEM INIATED######
    You will be given a content from conversation chat (e.g., text/ paragraph). 
    Your task is to analyze the given content and provide different types of places as close as possible to the given content.
    For exampl: If the given content (conversation chat) was about "How to make a slingshot", you can provide places like "Hardware store", "Woodworking shop", "Outdoor sports store", etc.
    Make sure the places you provide are relevant to the given content. And as much as close to the given content, the better.
    Your final output should be a list of places. 
    Here's JSON format example:
    "`\``json"
    {
    "places": ["Hardware store", "Woodworking shop", "Outdoor sports store"]
    }
    "`\``"

    ######NOTE######
    Make sure to return only JSON data! Nothing else!
    ######SYSTEM SHUTDOWN######
"""

SYSTEM_PROMPT_FOLLOWUP - For generating follow-up questions based on the given content.

SYSTEM_PROMPT_FOLLOWUP = """
    ######SYSTEM INIATED######
    You will be given a content from conversation chat (e.g., text/ paragraph).
    Your task is to analyze the given content and provide a follow-up question based on the given content.
    For example: If the given content (conversation chat) was about "How to make a slingshot", you can provide a follow-up question like "What materials are needed to make a slingshot?".
    Make sure the follow-up question you provide is relevant to the given content.
    Your final output should be a List of follow-up question.
    Here's JSON format example:
    "`\``json"
    {
    "followup_question": ["What materials are needed to make a slingshot?", "How to make a slingshot more powerful?"]
    }
    "`\``"
    ######NOTE######
    Make sure to return only JSON data! Nothing else!
    ######SYSTEM SHUTDOWN######
"""

SYSTEM_PROMPT_KNOWLEDGE_BASE - For generating knowledge base responses based on the given content.

SYSTEM_PROMPT_KNOWLEDGE_BASE = """
    ######SYSTEM INIATED######
    You will be given a content from conversation chat (e.g., text/ paragraph).
    Your task is to analyze the given content and provide a knowledge base response based on the given content.
    For example: If the given content (conversation chat) was about "How to make a slingshot".
    You should analyze it and find the exact creator or founder or inventor of the slingshot.
    Let's assume you just found out that the slingshot was invented by "Charles Goodyear".
    Then return `question` in a JSON format. (e.g., {"question": "Who is Charles Goodyear?"}).
    Your final output should be a JSON data with the knowledge base response.
    Here's JSON format example:
    "`\``json"
    {
    "question": "Who is Charles Goodyear?",
    }
    "`\``"
    ######NOTE######
    Make sure to return only JSON data! Nothing else!
    ######SYSTEM SHUTDOWN######
    """

3.3 [r_types.py]

• Placeholder for custom types and schemas.

from typing import TypedDict

# [ChatMessage]:
# <response>
# {
#   "role": "system",
#   "content": "Hello, how can I help you today?"
# }
# </response>

class ChatMessage(TypedDict):
    role: str
    content: str


# [Search Videos]:
# <response>
# Videos:
# [{'link': 'https://www.youtube.com/watch?v=X9oWGuKypuY', 'thumbnail': 'https://dmwtgq8yidg0m.cloudfront.net/medium/d3G6HeC5BO93-video-thumb.jpeg', 'title': 'Easy Home Made Slingshot'}, {'link': 'https://www.youtube.com/watch?v=V2iZF8oAXHo&pp=ygUMI2d1bGVsaGFuZGxl', 'thumbnail': 'https://dmwtgq8yidg0m.cloudfront.net/medium/sb2Iw9Ug-Pne-video-thumb.jpeg', 'title': 'Making an Apple Wood Slingshot | Woodcraft'}]
# </response>

class SearchVideosResponse(TypedDict):
    link: str
    thumbnail: str
    title: str


# [Search Images]:
# <response>
# [{'source': '', 'original': 'https://i.ytimg.com/vi/iYlJirFtYaA/sddefault.jpg', 'title': 'How to make a Slingshot using Pencils ...', 'source_name': 'YouTube'}, {'source': '', 'original': 'https://i.ytimg.com/vi/HWSkVaptzRA/maxresdefault.jpg', 'title': 'How to make a Slingshot at Home - YouTube', 'source_name': 'YouTube'}, {'source': '', 'original': 'https://content.instructables.com/FHB/VGF8/FHXUOJKJ/FHBVGF8FHXUOJKJ.jpg?auto=webp', 'title': 'Country Boy" Style Slingshot ...', 'source_name': 'Instructables'}, {'source': '', 'original': 'https://i.ytimg.com/vi/6wXqlJVw03U/maxresdefault.jpg', 'title': 'Make slingshot using popsicle stick ...', 'source_name': 'YouTube'}, {'source': '', 'original': 'https://ds-tc.prod.pbskids.org/designsquad/diy/DESIGN-SQUAD-42.jpg', 'title': 'Build | Indoor Slingshot . DESIGN SQUAD ...', 'source_name': 'PBS KIDS'}, {'source': '', 'original': 'https://i.ytimg.com/vi/wCxFkPLuNyA/maxresdefault.jpg', 'title': 'Paper Ninja Weapons ...', 'source_name': 'YouTube'}, {'source': '', 'original': 'https://i0.wp.com/makezine.com/wp-content/uploads/2015/01/slingshot1.jpg?fit=800%2C600&ssl=1', 'title': 'Rotating Bearings ...', 'source_name': 'Make Magazine'}, {'source': '', 'original': 'https://makeandtakes.com/wp-content/uploads/IMG_1144-1.jpg', 'title': 'Make a DIY Stick Slingshot Kids Craft', 'source_name': 'Make and Takes'}, {'source': '', 'original': 'https://i.ytimg.com/vi/X9oWGuKypuY/maxresdefault.jpg', 'title': 'Easy Home Made Slingshot - YouTube', 'source_name': 'YouTube'}, {'source': '', 'original': 'https://www.wikihow.com/images/thumb/4/41/Make-a-Sling-Shot-Step-7-Version-5.jpg/550px-nowatermark-Make-a-Sling-Shot-Step-7-Version-5.jpg', 'title': 'How to Make a Sling Shot: 15 Steps ...', 'source_name': 'wikiHow'}]
# </response>

class SearchImagesResponse(TypedDict):
    source: str
    original: str
    title: str
    source: str
    source_name: str


# [Links]:
# <response>
# ['https://www.reddit.com/r/slingshots/comments/1d50p3e/how_to_build_a_sling_at_home_thats_not_shit/', 'https://www.instructables.com/Make-a-Giant-Slingshot/', 'https://www.mudandbloom.com/blog/stick-slingshot', 'https://pbskids.org/designsquad/build/indoor-slingshot/', 'https://www.instructables.com/How-to-Make-a-Slingshot-2/']
# </response>

class SearchLinksResponse(TypedDict):
    title: str
    link: str


### Local Map Response:
# <response>
# {
#   "link": "https://www.google.com/maps/place/San+Francisco,+CA/data=!4m2!3m1!1s0x80859a6d00690021:0x4a501367f076adff?sa=X&ved=2ahUKEwjqg7eNz9KLAxVCFFkFHWSPEeIQ8gF6BAgqEAA&hl=en",
#   "image": "https://dmwtgq8yidg0m.cloudfront.net/images/TdNFUpcEvvHL-local-map.webp"
# }
# </response>

class LocalMapResponse(TypedDict):
    link: str
    imgae: str


### Model Response:

# <response>
# {
#   'title': 'Nikola Tesla', 
#   'type': 'Engineer and futurist', 
#   'description': None, 
#   'born': 'July 10, 1856, Smiljan, Croatia', 
#   'died': 'January 7, 1943 (age 86 years), The New Yorker A Wyndham Hotel, New York, NY'
# }
# </response>

class KnowledgeBaseResponse(TypedDict):
    title: str
    type: str
    description: str
    born: str
    died: str

---

Step 4: Running the Application

4.1 Run **``:**

`bash
python3 app.py
`

4.2 Access the Application:

• Open your browser and visit the provided Gradio URL (http://127.0.0.1:7860).

---

Step 5: Application Features

Basic Interaction:

• Type queries directly into the chat interface.
• Receive AI-generated answers and relevant follow-up suggestions.

Advanced Features:

• Image, video, and link searches from the follow-up context.
• Knowledge base retrieval.
• Local map searches.

---

Step 5: Customizing Your App

• Modify prompts in [prompts.py] to personalize AI behavior.
• Expand functionality by adding more helpers or API endpoints in [bagoodex_client.py].
• Adjust UI and functionalities in [app.py].

---

Step 5: Deploying Your App

• Consider deploying on Hugging Face Spaces.

---

Conclusion

In this tutorial, you built "Bagoodex Web Search," a versatile AI-powered search tool. You learned to interact with external APIs, handle follow-up interactions, and create a user-friendly interface with Gradio. You can now expand this project with more features and deploy it to share with others.