update: prompt for introductions

app.py

@@ -311,7 +311,19 @@ Total takes: 2
                 temperature=0.5,
             )
         else:
-            prompt = f"""Call Details:
+            system_prompt = f"""You are analyzing a transcript for Call ID: {cid}, Session ID: {rsid}, Origin: {origin}, and Call Type: {ct}. 
+
+CORE REQUIREMENTS:
+1. TIMESTAMPS: Each clip must contain ONLY the specified speaker's dialogue about a single topic. No overlapping dialogue from other speakers. YOU NEED TO BE VERY CAREFUL ABOUT THIS RULE. YOU HAVE THE TRANSCRIPT AND YOU CAN SEE WHO IS SPEAKING AT WHAT TIME, SO BE VERY, VERY CAREFUL AND ONLY INCLUDE THE DIALOGUE OF THE SPEAKER YOU ARE MAKING THE CLIP FOR.
+2. DURATION: Clips should be between 20-90 seconds long.
+3. CONTENT: Select engaging, viral-worthy topics. Avoid mundane or irrelevant content.
+4. COVERAGE: Minimum 2 topics per speaker, aim for 3 if good content exists.
+5. YOU CAN IGNORE THE HOST IF NO COMPELLING CONTENT IS FOUND.
+
+YOU SHOULD prioritize accuracy in timestamps at all costs.
+"""
+
+            user_prompt = f"""Call Details:
 User ID: {uid}
 Call ID: {cid}
 Speakers: {", ".join(speaker_mapping.values())}
@@ -322,54 +334,43 @@ Format requirements:
 
 1. 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}})
+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}})
 
 2. TIMESTAMP RULES:
-- Start time (st): Must begin exactly when speaker starts discussing the specific topic
+- Start time (st): Must begin exactly when speaker starts discussing the specific topic.
 - End time (et): Must end exactly when either:
   * The speaker completes their point, or
-  * Before the next speaker begins
-- NO OVERLAP: Selected duration must NEVER include dialogue from other speakers
-- Duration limits: Minimum 20 seconds, maximum 1 minute 30 seconds
-- Time format: "Xs at HH:MM" where X = seconds
-- URL parameters: Convert display times to seconds
-  Example: "25s at 10:13" → st=613&et=638
+  * Before the next speaker begins.
+- NO OVERLAP: Selected duration must NEVER include dialogue from other speakers.
+- Duration limits: Minimum 20 seconds, maximum 1 minute 30 seconds.
+- Time format: "Xs at HH:MM" where X = seconds.
+- URL parameters: Convert display times to seconds.
+  Example: "25s at 10:13" → st=613&et=638.
 
 3. FORMATTING RULES:
-- Speaker names: Use markdown bold (**Name**)
-- Topic titles: First word capitalized, rest lowercase
-- Each topic must be a clickable link with correct timestamp
-- URL format: {link_start}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{start_time_in_sec}}&et={{end_time_in_sec}}&uid={{uid}})
+- Speaker names: Use markdown bold (**Name**).
+- Topic titles: First word capitalized, rest lowercase.
+- Each topic must be a clickable link with correct timestamp.
+- URL format: {{link_start}}://{{origin}}/collab/{{cid}}/{{rsid}}?st={{start_time_in_sec}}&et={{end_time_in_sec}}&uid={{uid}}).
 
 4. TOPIC SELECTION:
-- Prioritize engaging, viral-worthy content
-- Minimum 2 topics per speaker, aim for 3 if available (SKIP THE HOST if no compelling content)
-- Topics should be self-contained discussions within the timestamp
-- Skip speakers if fewer than 2 compelling topics found
+- Prioritize engaging, viral-worthy content.
+- Minimum 2 topics per speaker, aim for 3 if available (SKIP THE HOST if no compelling content).
+- Topics should be self-contained discussions within the timestamp.
+- Skip speakers if fewer than 2 compelling topics found.
 """
-            completion = client.chat.completions.create(
-                model="gpt-4o-mini",
-                messages=[
-                    {
-                        "role": "system",
-                        "content": f"""You are analyzing a transcript for Call ID: {cid}, Session ID: {rsid}, Origin: {origin}, Call Type: {ct}. 
 
-    CORE REQUIREMENTS:
-    1. TIMESTAMPS: Each clip must contain ONLY the specified speaker's dialogue about a single topic. No overlapping dialogue from other speakers. YOU NEED TO BE VERY CAREFUL ABOUT THIS RULE. YOU HAVE THE TRANSCRIPT AND YOU CAN SEE WHO IS SPEAKING AT WHAT TIME SO BE VERY VERY CARAEFUL AND ONLY INCLUDE THE DIALOGUE OF THE SPEAKER YOU ARE MAKING THE CLIP FOR.
-    2. DURATION: Clips should be between 20-60 seconds long.
-    3. CONTENT: Select engaging, viral-worthy topics. Avoid mundane or irrelevant content
-    4. COVERAGE: Minimum 2 topics per speaker, aim for 3 if good content exists.
-    5. YOU CAN IGNORE THE HOST IF NO COMPELLING CONTENT IS FOUND.
-
-    YOU SHOULD Prioritize accuracy in timestamp at every cost.""",
-                    },
-                    {"role": "user", "content": prompt},
-                ],
-                stream=True,
-                temperature=0.5,
-            )
+        completion = client.chat.completions.create(
+            model="gpt-4o-mini",
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ],
+            stream=True,
+            temperature=0.5,
+        )
 
         collected_messages = []
         # Iterate through the stream
@@ -458,15 +459,16 @@ 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:
-User: Give me the URL where each person has introduced themselves.
 Response (Please provide exact time where the person has started introducing themselves. Normally person start introducing themselves by saying Hi, Hello, I am, My name is, etc.):
 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}})
@@ -476,6 +478,7 @@ If the user provides a link to the agenda, use the correct_speaker_name_with_url
 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}]
+        print(messages[0]["content"])
 
         for user_msg, assistant_msg in chat_history:
             if user_msg is not None:  # Skip the initial message where user_msg is None
@@ -491,7 +494,7 @@ If the user provides the correct call type, use the correct_call_type function t
             messages=messages,
             tools=tools,
             stream=True,
-            temperature=0.5,
+            temperature=0.3,
         )
         collected_messages = []
         tool_calls_detected = False
@@ -787,7 +790,7 @@ def create_chat_interface():
                 transcript_data = get_transcript_for_url(turl)
                 transcript_processor = TranscriptProcessor(
                     transcript_data=transcript_data,
-                    max_segment_duration=20 if ct != "si" else 10,
+                    max_segment_duration=5 if ct != "si" else 10,
                     call_type=ct,
                 )