app.py

import json
import os
from dataclasses import dataclass
from typing import Dict, Generator, List

import gradio as gr
import requests
from bs4 import BeautifulSoup
from openai import OpenAI


@dataclass
class TranscriptSegment:
    speaker_id: str
    start_time: float
    end_time: float
    text: str
    speaker_name: str = ""


class TranscriptProcessor:
    def __init__(
        self,
        transcript_file: str = None,
        transcript_data: dict = None,
        max_segment_duration: int = None,
        call_type: str = "le",
    ):
        self.transcript_file = transcript_file
        self.transcript_data = transcript_data
        self.formatted_transcript = None
        self.segments = []
        self.speaker_mapping = {}
        self.max_segment_duration = max_segment_duration
        if self.transcript_file:
            self._load_transcript()
        elif self.transcript_data:
            pass  # transcript_data is already set
        else:
            raise ValueError(
                "Either transcript_file or transcript_data must be provided."
            )

        self._process_transcript()
        self._create_formatted_transcript()  # Create initial formatted transcript
        if call_type != "si":
            self.map_speaker_ids_to_names()

    def _load_transcript(self) -> None:
        """Load the transcript JSON file."""
        with open(self.transcript_file, "r") as f:
            self.transcript_data = json.load(f)

    def _format_time(self, seconds: float) -> str:
        """Convert seconds to formatted time string (MM:SS)."""
        minutes = int(seconds // 60)
        seconds = int(seconds % 60)
        return f"{minutes:02d}:{seconds:02d}"

    def _process_transcript(self) -> None:
        results = self.transcript_data["results"]
        current_words = []
        current_speaker = None
        current_start = None
        current_items = []

        for item in results["items"]:
            if item["type"] == "pronunciation":
                speaker = (
                    item.get("speaker_label", "").replace("spk_", "").replace("spk", "")
                )

                # Initialize on first pronunciation item
                if current_speaker is None:
                    current_speaker = speaker
                    current_start = float(item["start_time"])

                # Check for speaker change
                if speaker != current_speaker:
                    if current_items:
                        self._create_segment(
                            current_speaker,
                            current_start,
                            float(item["start_time"]),
                            current_items,
                        )
                    current_items = []
                    current_words = []
                    current_speaker = speaker
                    current_start = float(item["start_time"])

                current_items.append(item)
                current_words.append(item["alternatives"][0]["content"])
            elif item["type"] == "punctuation":
                current_items.append(item)
                # Only check for segment break if we're over 20 words
                if len(current_words) >= 20:
                    # Break on this punctuation
                    next_item = next(
                        (
                            it
                            for it in results["items"][
                                results["items"].index(item) + 1 :
                            ]
                            if it["type"] == "pronunciation"
                        ),
                        None,
                    )
                    if next_item:
                        self._create_segment(
                            current_speaker,
                            current_start,
                            float(next_item["start_time"]),
                            current_items,
                        )
                        current_items = []
                        current_words = []
                        current_start = float(next_item["start_time"])

        # Don't forget the last segment
        if current_items:
            last_time = max(
                float(item["end_time"])
                for item in current_items
                if item["type"] == "pronunciation"
            )
            self._create_segment(
                current_speaker, current_start, last_time, current_items
            )

    def _create_segment(
        self, speaker_id: str, start: float, end: float, items: list
    ) -> None:
        segment_content = []
        for item in items:
            if item["type"] == "pronunciation":
                segment_content.append(item["alternatives"][0]["content"])
            elif item["type"] == "punctuation":
                # Append punctuation to the last word without a space
                if segment_content:
                    segment_content[-1] += item["alternatives"][0]["content"]

        if segment_content:
            self.segments.append(
                TranscriptSegment(
                    speaker_id=speaker_id,
                    start_time=start,
                    end_time=end,
                    text=" ".join(segment_content),
                )
            )

    def correct_speaker_mapping_with_agenda(self, url: str) -> None:
        """Fetch agenda from a URL and correct the speaker mapping using OpenAI."""
        try:
            if not url.startswith("http"):
                # add https to the url
                url = "https://" + url

            response = requests.get(url)
            response.raise_for_status()
            html_content = response.text

            # Parse the HTML to find the desired description
            soup = BeautifulSoup(html_content, "html.parser")
            description_tag = soup.find(
                "script", {"type": "application/ld+json"}
            )  # Find the ld+json metadata block
            agenda = ""

            if description_tag:
                # Extract the JSON content
                json_data = json.loads(description_tag.string)
                if "description" in json_data:
                    agenda = json_data["description"]
                else:
                    print("Agenda description not found in the JSON metadata.")
            else:
                print("No structured data (ld+json) found.")

            if not agenda:
                print("No agenda found in the structured metadata. Trying meta tags.")

                # Fallback: Use meta description if ld+json doesn't have it
                meta_description = soup.find("meta", {"name": "description"})
                agenda = meta_description["content"] if meta_description else ""

            if not agenda:
                print("No agenda found in any description tags.")
                return

            print(self.speaker_mapping)

            prompt = (
                f"Given the original speaker mapping {self.speaker_mapping}, agenda:\n{agenda}, and the transcript: {self.formatted_transcript}\n\n"
                "Some speaker names in the mapping might have spelling errors or be incomplete."
                "Remember that the content in agenda is accurate and transcript can have errors so prioritize the spellings and names in the agenda content."
                "If the speaker name and introduction is similar to the agenda, update the speaker name in the mapping."
                "Please correct the names based on the agenda. Return the corrected mapping in JSON format as "
                "{'spk_0': 'Correct Name', 'spk_1': 'Correct Name', ...}."
                "You should only update the name if the name sounds very similar, or there is a good spelling overlap/ The Speaker Introduction matches the description of the Talk from Agends. If the name is totally unrelated, keep the original name."
                "You should always include all the speakers in the mapping from the original mapping, even if you don't update their names. i.e if there are 4 speakers in original mapping, new mapping should have 4 speakers always, ignore all the other spekaers in the agenda. I REPEAT DO NOT ADD OTHER NEW SPEAKERS IN THE MAPPING."
            )

            client = OpenAI()

            completion = client.chat.completions.create(
                model="gpt-4o-mini",
                messages=[
                    {
                        "role": "system",
                        "content": "You are a helpful assistant. Who analyzes the given transcript, original speaker mapping and agenda. From the Agenda, you fix the spelling mistakes in the speaker names or update the names if they are similar to the agenda. You should only update the name if the name sounds very similar, or there is a good spelling overlap/ The Speaker Introduction matches the description of the Talk from Agends. If the name is totally unrelated, keep the original name.",
                    },
                    {"role": "user", "content": prompt},
                ],
                temperature=0,
            )

            response_text = completion.choices[0].message.content.strip()
            try:
                corrected_mapping = json.loads(response_text)
            except Exception:
                response_text = response_text[
                    response_text.find("{") : response_text.rfind("}") + 1
                ]
                try:
                    corrected_mapping = json.loads(response_text)
                except json.JSONDecodeError:
                    print(
                        "Error parsing corrected speaker mapping JSON, keeping the original mapping."
                    )
                    corrected_mapping = self.speaker_mapping
            # Update the speaker mapping with corrected names
            self.speaker_mapping = corrected_mapping
            print("Corrected Speaker Mapping:", self.speaker_mapping)

            # Update the transcript segments with corrected names
            for segment in self.segments:
                spk_id = f"spk_{segment.speaker_id}"
                segment.speaker_name = self.speaker_mapping.get(spk_id, spk_id)

            # Recreate the formatted transcript with corrected names
            formatted_segments = []
            for seg in self.segments:
                start_time_str = self._format_time(seg.start_time)
                end_time_str = self._format_time(seg.end_time)
                formatted_segments.append(
                    f"time_stamp: {start_time_str}-{end_time_str}\n"
                    f"{seg.speaker_name}: {seg.text}\n"
                )
            self.formatted_transcript = "\n".join(formatted_segments)

        except requests.exceptions.RequestException as e:
            print(f"  ching agenda from URL: {str(e)}")
        except Exception as e:
            print(f"Error correcting speaker mapping: {str(e)}")

    def _create_formatted_transcript(self) -> None:
        """Create formatted transcript with default speaker labels."""
        formatted_segments = []
        for seg in self.segments:
            start_time_str = self._format_time(seg.start_time)
            end_time_str = self._format_time(seg.end_time)
            # Use default speaker label (spk_X) if no mapping exists
            speaker_label = f"spk_{seg.speaker_id}"
            formatted_segments.append(
                f"time_stamp: {start_time_str}-{end_time_str}\n"
                f"{speaker_label}: {seg.text}\n"
            )
        self.formatted_transcript = "\n".join(formatted_segments)

    def map_speaker_ids_to_names(self) -> None:
        """Map speaker IDs to names based on introductions in the transcript."""
        try:
            transcript = self.formatted_transcript

            prompt = (
                "Given the following transcript where speakers are identified as spk 0, spk 1, spk 2, etc., please map each spk ID to the speaker's name based on their introduction in the transcript. If no name is introduced for a speaker, keep it as spk_id. Return the mapping as a JSON object in the format {'spk_0': 'Speaker Name', 'spk_1': 'Speaker Name', ...}\n\n"
                f"Transcript:\n{transcript}"
            )

            client = OpenAI()

            completion = client.chat.completions.create(
                model="gpt-4o",
                messages=[
                    {"role": "system", "content": "You are a helpful assistant."},
                    {"role": "user", "content": prompt},
                ],
                temperature=0,
            )

            response_text = completion.choices[0].message.content.strip()
            try:
                self.speaker_mapping = json.loads(response_text)
            except json.JSONDecodeError:
                response_text = response_text[
                    response_text.find("{") : response_text.rfind("}") + 1
                ]
                try:
                    self.speaker_mapping = json.loads(response_text)
                except json.JSONDecodeError:
                    print("Error parsing speaker mapping JSON.")
                    self.speaker_mapping = {}

            # Update segments with speaker names and recreate formatted transcript
            for segment in self.segments:
                spk_id = f"spk_{segment.speaker_id}"
                speaker_name = self.speaker_mapping.get(spk_id, spk_id)
                segment.speaker_name = speaker_name

            self._create_formatted_transcript_with_names()

        except Exception as e:
            print(f"Error mapping speaker IDs to names: {str(e)}")
            self.speaker_mapping = {}

    def _create_formatted_transcript_with_names(self) -> None:
        """Create formatted transcript with mapped speaker names."""
        formatted_segments = []
        for seg in self.segments:
            start_time_str = self._format_time(seg.start_time)
            end_time_str = self._format_time(seg.end_time)
            speaker_name = getattr(seg, "speaker_name", f"spk_{seg.speaker_id}")
            formatted_segments.append(
                f"Start Time: {start_time_str} - End Time: {end_time_str}\n"
                f"{speaker_name}: {seg.text}\n"
            )
        self.formatted_transcript = "\n".join(formatted_segments)

    def get_transcript(self) -> str:
        """Return the formatted transcript with speaker names."""
        return self.formatted_transcript

    def get_transcript_data(self) -> Dict:
        """Return the raw transcript data."""
        return self.transcript_data


def setup_openai_key() -> None:
    """Set up OpenAI API key from file."""
    try:
        with open("api.key", "r") as f:
            os.environ["OPENAI_API_KEY"] = f.read().strip()
    except FileNotFoundError:
        print("Using ENV variable")
        # raise FileNotFoundError(
        #     "api.key file not found. Please create it with your OpenAI API key."
        # )


def get_transcript_for_url(url: str) -> dict:
    """
    This function fetches the transcript data for a signed URL.
    If the URL results in a direct download, it processes the downloaded content.

    :param url: Signed URL for the JSON file
    :return: Parsed JSON data as a dictionary
    """
    headers = {
        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36"
    }

    try:
        response = requests.get(url, headers=headers)
        response.raise_for_status()

        if "application/json" in response.headers.get("Content-Type", ""):
            return response.json()  # Parse and return JSON directly

        # If not JSON, assume it's a file download (e.g., content-disposition header)
        content_disposition = response.headers.get("Content-Disposition", "")
        if "attachment" in content_disposition:
            # Process the content as JSON
            return json.loads(response.content)

        return json.loads(response.content)

    except requests.exceptions.HTTPError as http_err:
        print(f"HTTP error occurred: {http_err}")
    except requests.exceptions.RequestException as req_err:
        print(f"Request error occurred: {req_err}")
    except json.JSONDecodeError as json_err:
        print(f"JSON decoding error: {json_err}")

    return {}


def get_initial_analysis(
    transcript_processor: TranscriptProcessor, cid, rsid, origin, ct, uid
) -> Generator[str, None, None]:
    """Perform initial analysis of the transcript using OpenAI."""
    try:
        transcript = transcript_processor.get_transcript()
        speaker_mapping = transcript_processor.speaker_mapping
        client = OpenAI()
        if "localhost" in origin:
            link_start = "http"
        else:
            link_start = "https"
        if ct == "si":  # street interview
            prompt = f"""This is a transcript for a street interview. Call Details are as follows:
User ID UID: {uid}
Transcript: {transcript}

Your task is to analyze this street interview transcript and identify the final/best timestamps for each topic or question discussed. Here are the key rules:
The user might repeat the answer to the question sometimes, you need to pick the very last answer intelligently

1. For any topic/answer that appears multiple times in the transcript (even partially):
   - The LAST occurrence is always considered the best version. If the same thing is said multiple times, the last time is the best, all previous times are considered as additional takes.
   - This includes cases where parts of an answer are scattered throughout the transcript
   - Even slight variations of the same answer should be tracked
   - List timestamps for ALL takes, with the final take highlighted as the best answer

2. Introduction handling:
   - Question 1 is ALWAYS the speaker's introduction/self-introduction
   - If someone introduces themselves multiple times, use the last introduction as best answer
   - Include all variations of how they state their name/background
   - List ALL introduction timestamps chronologically

3. Question sequence:
   - After the introduction, list questions in the order they were first asked
   - If a question or introduction is revisited later at any point, please use the later timestamp
   - Track partial answers to the same question across the transcript

You need to make sure that any words that are repeated, you need to pick the last of them.

Return format:

[Question Title]
Total takes: [X] (Include ONLY if content appears more than once)
- [Take 1. <div id='topic' style="display: inline"> 15s at 12:30 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{750}}&et={{765}}&uid={{uid}})
- [Take 2. <div id='topic' style="display: inline"> 30s at 14:45 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{885}}&et={{915}}&uid={{uid}})
...
- [Take X (Best). <div id='topic' style="display: inline"> 1m 10s at 16:20 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{980}}&et={{1050}}&uid={{uid}})

URL formatting:
- Convert timestamps to seconds (e.g., 10:13 → 613)
- Format: {link_start}://[origin]/colab/[cid]/[rsid]?st=[start_seconds]&et=[end_seconds]&uid=[unique_id]
- Parameters after RSID must start with ? and subsequent parameters use &

Example:
1. Introduction
Total takes: 2
- [Take 1. <div id='topic' style="display: inline"> 10s at 09:45]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{585}}&et={{595}}&uid={{uid}})
- [Take 1. <div id='topic' style="display: inline"> 20s at 25:45]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{1245}}&et={{1265}}&uid={{uid}}))
- [Take 3 (Best). <div id='topic' style="display: inline"> 5s at 10:13 </div>]({link_start}://roll.ai/colab/1234aq_12314/51234151?st=613&et=618&uid=82314)"""
            completion = client.chat.completions.create(
                model="gpt-4o",
                messages=[
                    {
                        "role": "system",
                        "content": f"""You are analyzing a transcript for Call ID: {cid}, Session ID: {rsid}, Origin: {origin}, Call Type: {ct}.
    CORE REQUIREMENT:
- TIMESTAMPS: A speaker can repeat the answer to a question multiple times. You need to pick the last answer very carefully and choose that as best take. Make sure that that same answer is not repeated again after the best answer.

    YOU SHOULD Prioritize accuracy in timestamp at every cost. Read the Transcript carefully and decide where an answer starts and ends. You will have speaker labels so you need to be very sharp.""",
                    },
                    {"role": "user", "content": prompt},
                ],
                stream=True,
                temperature=0.1,
            )
        else:
            system_prompt = f"""You are a helpful assistant developed by Roll.AI(Leading AI tool for Remote production) who is analyzing the transcript for a RollAI Call. Following are the details: 
- Call ID: {cid}
- Session ID: {rsid}
- Origin: {origin}
- Call Type: {ct}
- Speakers: {", ".join(speaker_mapping.values())}
- Diarized Transcript: {transcript}


You are tasked with creating social media clips from the transcript, You need to shortlist the atleast two short clips for EACH SPEAKER. There are some requirments:

CORE REQUIREMENTS:
1. SPEAKER Overlap in the CLIP: When specifying the duration for the script, make sure that in that duration:
   - There is only continuous dialogue from that speaker.
   - As soon as another speaker starts talking or the topic ends, the clip MUST end.

2. DURATION RULES:
   - Each clip must be between 20 seconds to 120 seconds.

3. SPEAKER COVERAGE:
   - Minimum 2 topics per speaker, aim for 3 if good content exists

CRITICAL: When analyzing timestamps, you must verify that in the duration specified:
1. No other speaker talks during the selected timeframe
2. The speaker talks continuously for at least 20 seconds
3. The clip ends BEFORE any interruption or speaker change
"""
            print(" , ".join(speaker_mapping.values()))
            reasoning_prompt = f"""For each Speaker {" , ".join(speaker_mapping.values())}
in the transcript: {transcript}

Your job is to generate the thinking about the short social media clips for each speaker where they discuss. Think step by step and return a JSON at the end of the thinking.
Generate the thinking for atleast 2 clips for each speaker.
Return Format:
- Name of the Speaker
- Detailed Step by Step Thinking for each speaker from thier content and the topic they are talking about
After you have completed the thinking, give me a JSON of the thinking.
```json
[
    {{
        "Speaker 0": [
            {{
                "Topic Title": "...",
                "Starting Sentence of that speaker": "...",
                "Ending Sentence where the topic ends": "...."
            }},
            {{
                "Topic Title": "...",
                "Starting Sentence of that speaker": "....",
                "Ending Sentence of that speaker where the topic ends": "....."
            }}
        ]
    }},
    {{
        "Speaker 1": [
            {{
                "Topic Title": "....",
                "Starting Sentence of that speaker": ".....",
                "Ending Sentence of that speaker": "....."
            }},
            {{
                "Topic Title": "......",
                "Starting Sentence of that speaker": "....",
                "Ending Sentence of that speaker": "....."
            }}
        ]
    }},
    ....
]
```
"""
            thinking_completion = client.chat.completions.create(
                model="gpt-4o",
                messages=[
                    {"role": "system", "content": reasoning_prompt},
                ],
                stream=False,
                temperature=0.4,
            )
            thinking = thinking_completion.choices[0].message.content
            print("Thinking is:\n", thinking)
            thinking_json = thinking[thinking.find("{") : thinking.rfind("}") + 1]

            user_prompt = f"""User ID: {uid}
Intelligent Thinking Context: {thinking_json}

Your task is to generate the social media clips following these strict rules:

1. TIMESTAMP SELECTION:
- You must check the transcript line by line
- Verify speaker continuity with NO interruptions
- End clips immediately before any other speaker starts
- If Speaker A talks from 1:00-1:10, then Speaker B talks, then Speaker A resumes at 1:15, these must be separate clips
- Never combine timestamps across interruptions

2. CLIP REQUIREMENTS:
- Minimum 20 seconds of CONTINUOUS speech
- Maximum 100 seconds
- Single speaker only
- Must end before any interruption
- Complete thoughts/topics only


Return Format requirements:
SPEAKER FORMAT:
**Speaker Name**
1. [Topic title <div id='topic' style="display: inline"> 22s at 12:30 </div>]({{link_start}}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{750}}&et={{772}}&uid={{uid}})
2. [Topic title <div id='topic' style="display: inline"> 43s at 14:45 </div>]({{link_start}}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{885}}&et={{928}}&uid={{uid}})
3. [Topic title <div id='topic' style="display: inline"> 58s at 16:20 </div>]({{link_start}}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{980}}&et={{1038}}&uid={{uid}})
**Speaker Name**
....
"""

            completion = client.chat.completions.create(
                model="gpt-4o",
                messages=[
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": user_prompt},
                ],
                stream=True,
                temperature=0.1,
            )

        collected_messages = []
        # Iterate through the stream
        for chunk in completion:
            if chunk.choices[0].delta.content is not None:
                chunk_message = chunk.choices[0].delta.content
                collected_messages.append(chunk_message)
                # Yield the accumulated message so far
                yield "".join(collected_messages)

    except Exception as e:
        print(f"Error in initial analysis: {str(e)}")
        yield "An error occurred during initial analysis. Please check your API key and file path."


def chat(
    message: str,
    chat_history: List,
    transcript_processor: TranscriptProcessor,
    cid,
    rsid,
    origin,
    ct,
    uid,
) -> str:
    tools = [
        {
            "type": "function",
            "function": {
                "name": "correct_speaker_name_with_url",
                "description": "If a User provides a link to Agenda file, call the correct_speaker_name_with_url function to correct the speaker names based on the url, i.e if a user says 'Here is the Luma link for the event' and provides a link to the event, the function will correct the speaker names based on the event.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "url": {
                            "type": "string",
                            "description": "The url to the agenda.",
                        },
                    },
                    "required": ["url"],
                    "additionalProperties": False,
                },
            },
        },
        {
            "type": "function",
            "function": {
                "name": "correct_call_type",
                "description": "If the user tells you the correct call type, you have to apologize and call this function with correct call type.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "call_type": {
                            "type": "string",
                            "description": "The correct call type. If street interview, call type is 'si'.",
                        },
                    },
                    "required": ["call_type"],
                    "additionalProperties": False,
                },
            },
        },
    ]

    try:
        client = OpenAI()

        if "localhost" in origin:
            link_start = "http"
        else:
            link_start = "https"
        speaker_mapping = transcript_processor.speaker_mapping
        prompt = f"""You are a helpful assistant analyzing transcripts and generating timestamps and URL. The user will ask you questions regarding the social media clips from the transcript.
Call ID is {cid},
Session ID is {rsid},
origin is {origin},
Call Type is {ct}.
Speakers: {", ".join(speaker_mapping.values())}
Transcript: {transcript_processor.get_transcript()}

If a user asks timestamps for a specific topic or things, find the start time and end time of that specific topic and return answer in the format:
Answers and URLs should be formated as follows:
[Topic title <div id='topic' style="display: inline"> 22s at 12:30 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{750}}&et={{772}}&uid={{uid}})
For Example:
If the start time is 10:13 and end time is 10:18, the url will be:
{link_start}://roll.ai/colab/1234aq_12314/51234151?st=613&et=618&uid=82314
In the URL, make sure that after RSID there is ? and then rest of the fields are added via &.
You can include multiple links here that can related to the user answer. ALWAYS ANSWER FROM THE TRANSCRIPT.
RULE: When selecting timestamps for the answer, always use the **starting time (XX:YY)** as the reference point for your response, with the duration (Z seconds) calculated from this starting time, not the ending time of the segment.

Example 1:
User: Suggest me some clips that can go viral on Instagram.
Response:
1. [Clip 1 <div id='topic' style="display: inline"> 22s at 12:30 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{750}}&et={{772}}&uid={{uid}})
User: Give me the URL where each person has introduced themselves. 
2. [Clip 2 <div id='topic' style="display: inline"> 10s at 10:00 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{600}}&et={{610}}&uid={{uid}})

Example 2:
Provide the exact timestamp where the person begins their introduction, typically starting with phrases like "Hi," "Hello," "I am," or "My name is," and include the full introduction, covering everything they say about themselves, including their name, role, background, current responsibilities, organization, and any additional details they provide about their work or personal interests.
1. [Person Name1 <div id='topic' style="display: inline"> 43s at 14:45 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{885}}&et={{928}}&uid={{uid}})
2. [Person Name2 <div id='topic' style="display: inline"> 58s at 16:20 </div>]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{980}}&et={{1038}}&uid={{uid}})
....

If the user provides a link to the agenda, use the correct_speaker_name_with_url function to correct the speaker names based on the agenda.
If the user provides the correct call type, use the correct_call_type function to correct the call type. Call Type for street interviews is 'si'.
"""
        messages = [{"role": "system", "content": prompt}]

        for user_msg, assistant_msg in chat_history:
            if user_msg is not None:  # Skip the initial message where user_msg is None
                messages.append({"role": "user", "content": user_msg})
            if assistant_msg is not None:
                messages.append({"role": "assistant", "content": assistant_msg})

        # Add the current message
        messages.append({"role": "user", "content": message})

        completion = client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            tools=tools,
            stream=True,
            temperature=0.3,
        )
        collected_messages = []
        tool_calls_detected = False

        for chunk in completion:
            if chunk.choices[0].delta.tool_calls:
                tool_calls_detected = True
                # Handle tool calls without streaming
                response = client.chat.completions.create(
                    model="gpt-4o",
                    messages=messages,
                    tools=tools,
                )

                if response.choices[0].message.tool_calls:
                    tool_call = response.choices[0].message.tool_calls[0]
                    if tool_call.function.name == "correct_speaker_name_with_url":
                        args = eval(tool_call.function.arguments)
                        url = args.get("url", None)
                        if url:
                            transcript_processor.correct_speaker_mapping_with_agenda(
                                url
                            )
                            corrected_speaker_mapping = (
                                transcript_processor.speaker_mapping
                            )
                            messages.append(response.choices[0].message)

                            function_call_result_message = {
                                "role": "tool",
                                "content": json.dumps(
                                    {
                                        "speaker_mapping": f"Corrected Speaker Mapping... {corrected_speaker_mapping}"
                                    }
                                ),
                                "name": tool_call.function.name,
                                "tool_call_id": tool_call.id,
                            }
                            messages.append(function_call_result_message)

                            # Get final response after tool call
                            final_response = client.chat.completions.create(
                                model="gpt-4o",
                                messages=messages,
                                stream=True,
                            )

                            collected_chunk = ""
                            for final_chunk in final_response:
                                if final_chunk.choices[0].delta.content:
                                    collected_chunk += final_chunk.choices[
                                        0
                                    ].delta.content
                                    yield collected_chunk
                            return
                        else:
                            function_call_result_message = {
                                "role": "tool",
                                "content": "No URL Provided",
                                "name": tool_call.function.name,
                                "tool_call_id": tool_call.id,
                            }

                    elif tool_call.function.name == "correct_call_type":
                        args = eval(tool_call.function.arguments)
                        call_type = args.get("call_type", None)
                        if call_type:
                            # Stream the analysis for corrected call type
                            for content in get_initial_analysis(
                                transcript_processor,
                                call_type,
                                rsid,
                                origin,
                                call_type,
                                uid,
                            ):
                                yield content
                            return
                break  # Exit streaming loop if tool calls detected

            if not tool_calls_detected and chunk.choices[0].delta.content is not None:
                chunk_message = chunk.choices[0].delta.content
                collected_messages.append(chunk_message)
                yield "".join(collected_messages)

    except Exception as e:
        print(f"Unexpected error in chat: {str(e)}")
        import traceback

        print(f"Traceback: {traceback.format_exc()}")
        yield "Sorry, there was an error processing your request."


def create_chat_interface():
    """Create and configure the chat interface."""
    css = """
    .gradio-container {

        padding-top: 0px !important;
        padding-left: 0px !important;
        padding-right: 0px !important;
        padding: 0px !important;
        margin: 0px !important;
    }
    #component-0 {
        gap: 0px !important;
    }

    .icon-button-wrapper{
        display: none !important;
    }


    footer {
        display: none !important;
    }
    #chatbot_box{
        flex-grow: 1 !important;
        border-width: 0px !important;
    }

    #link-frame {
        position: absolute !important;
        width: 1px !important;
        height: 1px !important;
        right: -100px !important;
        bottom: -100px !important;
        display: none !important;
    }
    .html-container {
        display: none !important;
    }
    a {
        text-decoration: none !important;
    }
    #topic {
        color: #aaa !important;
    }
    .bubble-wrap {
        padding-top: 0px !important;
    }
    .message-content {
        border: 0px !important;
        margin: 5px !important;
    }
    .message-row {
        border-style: none !important;
        margin: 0px !important;
        width: 100% !important;
        max-width: 100% !important;
    }
    .flex-wrap {
        border-style: none !important;
    }

    .panel-full-width {
        border-style: none !important;
        border-width: 0px !important;
    }
    ol {
        list-style-position: outside; 
        margin-left: 20px;
    }

    body.waiting * {
    cursor: progress;
    }
    """
    js = """
    function createIframeHandler() {
        let iframe = document.getElementById('link-frame');
        if (!iframe) {
            iframe = document.createElement('iframe');
            iframe.id = 'link-frame';
            iframe.style.position = 'absolute';
            iframe.style.width = '1px';
            iframe.style.height = '1px';
            iframe.style.right = '-100px';
            iframe.style.bottom = '-100px';
            iframe.style.display = 'none'; // Hidden initially
            document.body.appendChild(iframe);
        }

        document.addEventListener('click', function (event) {
            var link = event.target.closest('a');
            if (link && link.href) {
                document.body.classList.add('waiting');
                setTimeout(function () {
                    document.body.classList.remove('waiting');
                }, 2000); // Reset cursor after 1 seconds
                try {
                    iframe.src = link.href;
                    iframe.style.display = 'block'; // Show iframe on link click
                    event.preventDefault();
                    console.log('Opening link in iframe:', link.href);
                } catch (error) {
                    console.error('Failed to open link in iframe:', error);
                }
            }
        });

        return 'Iframe handler initialized';
    }
    """

    with gr.Blocks(
        fill_height=True,
        fill_width=True,
        css=css,
        js=js,
        theme=gr.themes.Default(
            font=[gr.themes.GoogleFont("Inconsolata"), "Arial", "sans-serif"]
        ),
    ) as demo:
        chatbot = gr.Chatbot(
            elem_id="chatbot_box",
            layout="bubble",
            show_label=False,
            show_share_button=False,
            show_copy_all_button=False,
            show_copy_button=False,
        )
        msg = gr.Textbox(elem_id="chatbot_textbox", show_label=False)
        transcript_processor_state = gr.State()  # maintain state of imp things
        call_id_state = gr.State()
        colab_id_state = gr.State()
        origin_state = gr.State()
        ct_state = gr.State()
        turl_state = gr.State()
        uid_state = gr.State()
        iframe_html = "<iframe id='link-frame'></iframe>"
        gr.HTML(value=iframe_html)  # Add iframe to the UI

        def respond(
            message: str,
            chat_history: List,
            transcript_processor,
            cid,
            rsid,
            origin,
            ct,
            uid,
        ):
            if not transcript_processor:
                bot_message = "Transcript processor not initialized."
                chat_history.append((message, bot_message))
                return "", chat_history

            chat_history.append((message, ""))
            for chunk in chat(
                message,
                chat_history[:-1],  # Exclude the current incomplete message
                transcript_processor,
                cid,
                rsid,
                origin,
                ct,
                uid,
            ):
                chat_history[-1] = (message, chunk)
                yield "", chat_history

        msg.submit(
            respond,
            [
                msg,
                chatbot,
                transcript_processor_state,
                call_id_state,
                colab_id_state,
                origin_state,
                ct_state,
                uid_state,
            ],
            [msg, chatbot],
        )

        # Handle initial loading with streaming
        def on_app_load(request: gr.Request):
            cid = request.query_params.get("cid", None)
            rsid = request.query_params.get("rsid", None)
            origin = request.query_params.get("origin", None)
            ct = request.query_params.get("ct", None)
            turl = request.query_params.get("turl", None)
            uid = request.query_params.get("uid", None)

            required_params = ["cid", "rsid", "origin", "ct", "turl", "uid"]
            missing_params = [
                param
                for param in required_params
                if request.query_params.get(param) is None
            ]

            if missing_params:
                error_message = (
                    f"Missing required parameters: {', '.join(missing_params)}"
                )
                chatbot_value = [(None, error_message)]
                return [chatbot_value, None, None, None, None, None, None, None]

            try:
                transcript_data = get_transcript_for_url(turl)
                transcript_processor = TranscriptProcessor(
                    transcript_data=transcript_data,
                    max_segment_duration=5 if ct != "si" else 10,
                    call_type=ct,
                )

                # Initialize with empty message
                chatbot_value = [(None, "")]

                # Return initial values with the transcript processor
                return [
                    chatbot_value,
                    transcript_processor,
                    cid,
                    rsid,
                    origin,
                    ct,
                    turl,
                    uid,
                ]
            except Exception as e:
                error_message = f"Error processing call_id {cid}: {str(e)}"
                chatbot_value = [(None, error_message)]
                return [chatbot_value, None, None, None, None, None, None, None]

        def stream_initial_analysis(
            chatbot_value, transcript_processor, cid, rsid, origin, ct, uid
        ):
            if transcript_processor:
                for chunk in get_initial_analysis(
                    transcript_processor, cid, rsid, origin, ct, uid
                ):
                    chatbot_value[0] = (None, chunk)
                    yield chatbot_value
            else:
                yield chatbot_value

        # Modified load event to handle streaming
        demo.load(
            on_app_load,
            inputs=None,
            outputs=[
                chatbot,
                transcript_processor_state,
                call_id_state,
                colab_id_state,
                origin_state,
                ct_state,
                turl_state,
                uid_state,
            ],
        ).then(
            stream_initial_analysis,
            inputs=[
                chatbot,
                transcript_processor_state,
                call_id_state,
                colab_id_state,
                origin_state,
                ct_state,
                uid_state,
            ],
            outputs=[chatbot],
        )

    return demo


def main():
    """Main function to run the application."""
    try:
        setup_openai_key()
        demo = create_chat_interface()
        demo.launch(share=True)
    except Exception as e:
        print(f"Error starting application: {str(e)}")
        raise


if __name__ == "__main__":
    main()