update: si prompt

app.py

@@ -339,21 +339,39 @@ def get_initial_analysis(
 
         if ct == "si":  # street interview
             prompt = f"""This is a transcript for a street interview. Transcript: {transcript}
-In this street interview, the host asks multiple questions to the interviewees.
-The interviewee can repeat a single answer multiple time to get the best take.
-Your job is to find out the timestamp of the best answer given by the interviewee (Do not include the Question timestamp by interviwer in this). If there are multiple attempts for a question, best part is the last part of the question. If no question was asked but something is repeated, please include that in the answer as well.
-Even if the answer is scattered across the transcript, the last occurence of the topic is the best answer for that topic.
-The way to know if there are multiple takes to a question is to see in the transcript if the same text is repeated, If not then number of takes is 1.
-Question 1 should always be the introduction if the speaker has introduced themselves to find the best introduction time (Last timestamp is the best timestamp), Rest of questions should be in the order they were asked.
-Return format is:
-1. Question Title
-Total takes: (Only include this if Number of takes is > 1 otherwise ignore it.)
-[Best Answer: start_time - end_time]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{start_time_in_sec}}&et={{end_time_in_sec}}"').
-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
-In the URL, make sure that after RSID there is ? and then rest of the fields are added via &. Keep the answer less verbose and to the point.
-        """
+
+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:
+
+1. For any topic/answer that appears multiple times in the transcript (even partially):
+   - The LAST occurrence is always considered the best version
+   - This includes cases where parts of an answer are scattered throughout the transcript
+   - Even slight variations of the same answer should be tracked
+
+2. Introduction handling:
+   - Question 1 is ALWAYS the speaker's introduction/self-introduction
+   - If someone introduces themselves multiple times, use the last timestamp
+   - Include all variations of how they state their name/background
+
+3. Question sequence:
+   - After the introduction, list questions in the order they were first asked
+   - If a question is revisited later, use the later timestamp
+   - Track partial answers to the same question across the transcript
+
+Return format:
+
+[Question Title]
+Total takes: [X] (Include ONLY if content appears more than once)
+[Best Answer: start_time - end_time]({link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{start_time_in_sec}}&et={{end_time_in_sec}})
+
+URL formatting:
+- Convert timestamps to seconds (e.g., 10:13 → 613)
+- Format: {link_start}://[origin]/colab/[cid]/[rsid]?st=[start_seconds]&et=[end_seconds]
+- Parameters after RSID must start with ? and subsequent parameters use &
+
+Example:
+1. Introduction
+Total takes: 2
+[Best Answer: 10:13 - 10:18]({link_start}://roll.ai/colab/1234aq_12314/51234151?st=613&et=618)"""
         else:
             prompt = f"""Given the transcript {transcript}, For All the speakers, short list all people, news, events, trends, and source that are discussed by speakers along with the start time of that topic and end time of that topic from the transcript. Rank all topics based on what would make for the best social clips. I need atleast 3 topics per speaker.
 You should mention the Speaker Name first, then atleast 3 posts with their timestamps, and so on.