fix: better topic chunking and subject extraction (#952)

* fix: better detect topics chunking depending on duration

* fix:  better subject detection + prompt improvements

server/reflector/hatchet/workflows/daily_multitrack_pipeline.py

@@ -106,7 +106,10 @@ from reflector.utils.daily import (
     parse_daily_recording_filename,
 )
 from reflector.utils.string import NonEmptyString, assert_non_none_and_non_empty
-from reflector.utils.transcript_constants import TOPIC_CHUNK_WORD_COUNT
+from reflector.utils.transcript_constants import (
+    compute_max_subjects,
+    compute_topic_chunk_size,
+)
 from reflector.zulip import post_transcript_notification
 
 
@@ -885,7 +888,8 @@ async def detect_topics(input: PipelineInput, ctx: Context) -> TopicsResult:
         transcripts_controller,
     )
 
-    chunk_size = TOPIC_CHUNK_WORD_COUNT
+    duration_seconds = words[-1].end - words[0].start if words else 0
+    chunk_size = compute_topic_chunk_size(duration_seconds, len(words))
     chunks = []
     for i in range(0, len(words), chunk_size):
         chunk_words = words[i : i + chunk_size]
@@ -975,7 +979,7 @@ async def detect_topics(input: PipelineInput, ctx: Context) -> TopicsResult:
 
     ctx.log(f"detect_topics complete: found {len(topics_list)} topics")
 
-    return TopicsResult(topics=topics_list)
+    return TopicsResult(topics=topics_list, duration_seconds=duration_seconds)
 
 
 @daily_multitrack_pipeline.task(
@@ -1112,8 +1116,14 @@ async def extract_subjects(input: PipelineInput, ctx: Context) -> SubjectsResult
                 participant_names, participant_name_to_id=participant_name_to_id
             )
 
+        max_subjects = compute_max_subjects(topics_result.duration_seconds)
+        ctx.log(
+            f"extract_subjects: duration={topics_result.duration_seconds:.0f}s, "
+            f"max_subjects={max_subjects}"
+        )
+
         ctx.log("extract_subjects: calling LLM to extract subjects")
-        await builder.extract_subjects()
+        await builder.extract_subjects(max_subjects=max_subjects)
 
         ctx.log(f"extract_subjects complete: {len(builder.subjects)} subjects")
 

server/reflector/hatchet/workflows/models.py

@@ -102,6 +102,7 @@ class TopicsResult(BaseModel):
     """Result from detect_topics task."""
 
     topics: list[TitleSummary]
+    duration_seconds: float = 0
 
 
 class TitleResult(BaseModel):

server/reflector/pipelines/topic_processing.py

@@ -18,7 +18,7 @@ from reflector.processors import (
 )
 from reflector.processors.types import TitleSummary
 from reflector.processors.types import Transcript as TranscriptType
-from reflector.utils.transcript_constants import TOPIC_CHUNK_WORD_COUNT
+from reflector.utils.transcript_constants import compute_topic_chunk_size
 
 
 class EmptyPipeline:
@@ -39,7 +39,10 @@ async def detect_topics(
     on_topic_callback: Callable,
     empty_pipeline: EmptyPipeline,
 ) -> list[TitleSummary]:
-    chunk_size = TOPIC_CHUNK_WORD_COUNT
+    duration_seconds = (
+        transcript.words[-1].end - transcript.words[0].start if transcript.words else 0
+    )
+    chunk_size = compute_topic_chunk_size(duration_seconds, len(transcript.words))
     topics: list[TitleSummary] = []
 
     async def on_topic(topic: TitleSummary):

server/reflector/processors/summary/prompts.py

@@ -43,7 +43,8 @@ DETAILED_SUBJECT_PROMPT_TEMPLATE = dedent(
       include any deadlines or timeframes discussed for completion or follow-up.
     - Mention unresolved issues or topics needing further discussion, aiding in
       planning future meetings or follow-up actions.
-    - Do not include topic unrelated to {subject}.
+    - Be specific and cite participant names when attributing statements or actions.
+    - Do not include topics unrelated to {subject}.
 
     # OUTPUT
     Your summary should be clear, concise, and structured, covering all major
@@ -58,6 +59,7 @@ PARAGRAPH_SUMMARY_PROMPT = dedent(
     """
     Summarize the mentioned topic in 1 paragraph.
     It will be integrated into the final summary, so just for this topic.
+    Preserve key decisions and action items. Do not introduce new information.
     """
 ).strip()
 

server/reflector/processors/summary/summary_builder.py

@@ -48,17 +48,24 @@ TRANSCRIPTION_TYPE_PROMPT = dedent(
     """
 ).strip()
 
-SUBJECTS_PROMPT = dedent(
-    """
-    What are the main / high level topic of the meeting.
-    Do not include direct quotes or unnecessary details.
-    Be concise and focused on the main ideas.
-    A subject briefly mentioned should not be included.
-    There should be maximum 6 subjects.
-    Do not write complete narrative sentences for the subject,
-    you must write a concise subject using noun phrases.
-    """
-).strip()
+_DEFAULT_MAX_SUBJECTS = 6
+
+
+def build_subjects_prompt(max_subjects: int = _DEFAULT_MAX_SUBJECTS) -> str:
+    """Build subjects extraction prompt with a dynamic subject cap."""
+    subject_word = "subject" if max_subjects == 1 else "subjects"
+    return dedent(
+        f"""
+        What are the main / high level topics of the meeting.
+        Do not include direct quotes or unnecessary details.
+        Be concise and focused on the main ideas.
+        A subject briefly mentioned should not be included.
+        There should be maximum {max_subjects} {subject_word}.
+        Do not write complete narrative sentences for the subject,
+        you must write a concise subject using noun phrases.
+        """
+    ).strip()
+
 
 ACTION_ITEMS_PROMPT = dedent(
     """
@@ -145,7 +152,7 @@ class SubjectsResponse(BaseModel):
     """Pydantic model for extracted subjects/topics"""
 
     subjects: list[str] = Field(
-        description="List of main subjects/topics discussed, maximum 6 items",
+        description="List of main subjects/topics discussed",
     )
 
 
@@ -345,11 +352,14 @@ class SummaryBuilder:
     # Summary
     # ----------------------------------------------------------------------------
 
-    async def extract_subjects(self) -> None:
+    async def extract_subjects(self, max_subjects: int = _DEFAULT_MAX_SUBJECTS) -> None:
         """Extract main subjects/topics from the transcript."""
-        self.logger.info("--- extract main subjects using TreeSummarize")
+        self.logger.info(
+            "--- extract main subjects using TreeSummarize",
+            max_subjects=max_subjects,
+        )
 
-        subjects_prompt = SUBJECTS_PROMPT
+        subjects_prompt = build_subjects_prompt(max_subjects)
 
         try:
             response = await self._get_structured_response(
@@ -358,7 +368,7 @@ class SummaryBuilder:
                 tone_name="Meeting assistant that talk only as list item",
             )
 
-            self.subjects = response.subjects
+            self.subjects = response.subjects[:max_subjects]
             self.logger.info(f"Extracted subjects: {self.subjects}")
 
         except Exception as e:

server/reflector/utils/transcript_constants.py

@@ -4,5 +4,67 @@ Shared transcript processing constants.
 Used by both Hatchet workflows and Celery pipelines for consistent processing.
 """
 
-# Topic detection: number of words per chunk for topic extraction
+import math
+
+# Topic detection: legacy static chunk size, used as fallback
 TOPIC_CHUNK_WORD_COUNT = 300
+
+# Dynamic chunking curve parameters
+# Formula: target_topics = _COEFFICIENT * duration_minutes ^ _EXPONENT
+# Derived from anchors: 5 min -> 3 topics, 180 min -> 40 topics
+_TOPIC_CURVE_COEFFICIENT = 0.833
+_TOPIC_CURVE_EXPONENT = 0.723
+_MIN_TOPICS = 2
+_MAX_TOPICS = 50
+_MIN_CHUNK_WORDS = 375
+_MAX_CHUNK_WORDS = 1500
+
+
+def compute_topic_chunk_size(duration_seconds: float, total_words: int) -> int:
+    """Calculate optimal chunk size for topic detection based on recording duration.
+
+    Uses a power-curve function to scale topic count sublinearly with duration,
+    producing fewer LLM calls for longer recordings while maintaining topic quality.
+
+    Returns the number of words per chunk.
+    """
+    if total_words <= 0 or duration_seconds <= 0:
+        return _MIN_CHUNK_WORDS
+
+    duration_minutes = duration_seconds / 60.0
+    target_topics = _TOPIC_CURVE_COEFFICIENT * math.pow(
+        duration_minutes, _TOPIC_CURVE_EXPONENT
+    )
+    target_topics = int(round(max(_MIN_TOPICS, min(_MAX_TOPICS, target_topics))))
+
+    chunk_size = total_words // target_topics
+    chunk_size = max(_MIN_CHUNK_WORDS, min(_MAX_CHUNK_WORDS, chunk_size))
+    return chunk_size
+
+
+# Subject extraction: scale max subjects with recording duration
+# Short calls get fewer subjects to avoid over-analyzing trivial content
+_SUBJECT_DURATION_THRESHOLDS = [
+    (5 * 60, 1),  # ≤ 5 min  → 1 subject
+    (15 * 60, 2),  # ≤ 15 min → 2 subjects
+    (30 * 60, 3),  # ≤ 30 min → 3 subjects
+    (45 * 60, 4),  # ≤ 45 min → 4 subjects
+    (60 * 60, 5),  # ≤ 60 min → 5 subjects
+]
+_MAX_SUBJECTS = 6
+
+
+def compute_max_subjects(duration_seconds: float) -> int:
+    """Calculate maximum number of subjects to extract based on recording duration.
+
+    Uses a step function: short recordings get fewer subjects to avoid
+    generating excessive detail for trivial content.
+    """
+    if duration_seconds <= 0:
+        return 1
+
+    for threshold, max_subjects in _SUBJECT_DURATION_THRESHOLDS:
+        if duration_seconds <= threshold:
+            return max_subjects
+
+    return _MAX_SUBJECTS