momentry_core/scripts/generate_chunk_summaries.py

#!/opt/homebrew/bin/python3.11
"""
Generate individual chunk summaries combining:
- chunk.text_content (specific content)
- parent.structured_summary (5W1H context)

Each chunk gets a tailored summary that contextualizes its specific content
within the broader parent chunk narrative.
"""

import json
import requests
import psycopg2
import psycopg2.extras
import time
import os

DB_CONFIG = {
    "host": "localhost",
    "user": "accusys",
    "dbname": "momentry",
}

SCHEMA = os.environ.get("DATABASE_SCHEMA", "dev")
LLAMA_URL = "http://127.0.0.1:8081/v1/chat/completions"
BATCH_SIZE = 50
DELAY_BETWEEN_BATCHES = 1


def get_chunks_with_parents(uuid=None, limit=None):
    """Get chunks with their parent 5W1H metadata and identity info"""
    conn = psycopg2.connect(**DB_CONFIG)
    cur = conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor)

    where_clause = "WHERE c.summary_text IS NULL AND c.text_content IS NOT NULL AND c.parent_chunk_id IS NOT NULL"
    if uuid:
        where_clause += f" AND c.uuid = '{uuid}'"

    query = f"""
        SELECT c.chunk_id, c.uuid, c.text_content, c.chunk_type,
               c.parent_chunk_id,
               c.speaker_ids,
               c.face_ids,
               c.visual_stats,
               pc.metadata->'structured_summary' as structured_summary,
               pc.summary_text as parent_summary,
               c.start_time,
               c.end_time
        FROM {SCHEMA}.chunks c
        LEFT JOIN {SCHEMA}.parent_chunks pc 
            ON c.parent_chunk_id = pc.id::varchar
        {where_clause}
        ORDER BY c.chunk_id
    """
    if limit:
        query += f" LIMIT {limit}"

    cur.execute(query)
    chunks = cur.fetchall()
    cur.close()
    conn.close()
    return chunks


def get_person_identities(uuid, start_time, end_time):
    """取得 chunk 時間範圍內的人物識別"""
    conn = psycopg2.connect(**DB_CONFIG)
    cur = conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor)

    cur.execute(
        f"""
        SELECT person_id, name, speaker_id
        FROM {SCHEMA}.person_identities
        WHERE video_uuid = %s 
          AND speaker_id IS NOT NULL
          AND last_appearance_time >= %s
          AND first_appearance_time <= %s
    """,
        (uuid, start_time, end_time),
    )

    persons = cur.fetchall()
    cur.close()
    conn.close()
    return persons
    if limit:
        query += f" LIMIT {limit}"

    cur.execute(query)
    chunks = cur.fetchall()
    cur.close()
    conn.close()
    return chunks


def call_llm(prompt, max_tokens=500):
    """Call Gemma4 via llama-server"""
    payload = {
        "messages": [{"role": "user", "content": prompt}],
        "max_tokens": max_tokens,
        "temperature": 0.3,
        "min_p": 0.1,
    }
    try:
        resp = requests.post(LLAMA_URL, json=payload, timeout=60)
        if resp.status_code == 200:
            result = resp.json()
            choice = result.get("choices", [{}])[0]
            message = choice.get("message", {})

            # Gemma4 returns content directly (final answer)
            content = message.get("content", "").strip()

            # If content exists and is not empty, use it
            if content:
                return content

            # If content is empty, try to extract from reasoning
            reasoning = message.get("reasoning_content", "")
            if reasoning:
                # Look for final answer markers in reasoning
                markers = ["Final:", "**Final**:", "Final answer:", "**Final answer**:"]
                for marker in markers:
                    if marker in reasoning:
                        answer = reasoning.split(marker)[-1].strip()
                        # Clean up the answer
                        answer = answer.split("\n")[0].strip()
                        if answer and not answer.startswith("Thinking"):
                            return answer

            # Last resort: return the whole reasoning (will contain thinking process)
            return reasoning.strip()
    except Exception as e:
        print(f"    ⚠️  LLM error: {e}")
    return ""


def generate_chunk_summary(chunk):
    """Generate summary for a single chunk with 5W1H"""
    text_content = chunk.get("text_content", "")
    parent_5w1h = chunk.get("structured_summary") or {}
    parent_summary = chunk.get("parent_summary", "")
    speaker_ids = chunk.get("speaker_ids", [])
    face_ids = chunk.get("face_ids", [])
    visual_stats = chunk.get("visual_stats", {})
    uuid = chunk.get("uuid", "")
    start_time = chunk.get("start_time", 0)
    end_time = chunk.get("end_time", 0)

    if not text_content:
        return ""

    speaker_list = ", ".join(speaker_ids) if speaker_ids else "None"
    face_list = ", ".join([f"face_{x}" for x in face_ids]) if face_ids else "None"
    visual_objects = (
        visual_stats.get("objects", []) if isinstance(visual_stats, dict) else []
    )
    visual_places = (
        visual_stats.get("places", []) if isinstance(visual_stats, dict) else []
    )
    visual_actions = (
        visual_stats.get("actions", []) if isinstance(visual_stats, dict) else []
    )
    visual_list = ", ".join(visual_objects[:5]) if visual_objects else "None"
    places_list = ", ".join(visual_places[:3]) if visual_places else "None"
    actions_list = ", ".join(visual_actions[:3]) if visual_actions else "None"

    identified_persons = []
    if uuid and start_time and end_time:
        try:
            identified_persons = get_person_identities(uuid, start_time, end_time)
        except Exception as e:
            print(f"    ⚠️  Person lookup error: {e}")

    person_list = (
        ", ".join(
            [
                f"{p['name'] or p['person_id']}({p['speaker_id']})"
                for p in identified_persons
            ]
        )
        if identified_persons
        else "None"
    )

    prompt = f"""You are analyzing a video chunk. Provide accurate, detailed 5W1H analysis.

CHUNK INFO:
- Chunk ID: {chunk.get("chunk_id")}
- Time range: {start_time:.2f}s - {end_time:.2f}s

BROADER SCENE CONTEXT (parent chunk, high confidence):
- Scene Who: {parent_5w1h.get("who", "N/A")}
- Scene What: {parent_5w1h.get("what", "N/A")}
- Scene When: {parent_5w1h.get("when", "N/A")}
- Scene Where: {parent_5w1h.get("where", "N/A")}
- Scene Why: {parent_5w1h.get("why", "N/A")}
- Scene How: {parent_5w1h.get("how", "N/A")}
- Tone: {parent_5w1h.get("tone", [])}
- Characters: {parent_5w1h.get("characters", [])}
- Key Events: {parent_5w1h.get("key_events", [])}

Parent summary: {parent_summary[:150] if parent_summary else "N/A"}...

CHUNK IDENTITY (from ASRX + Face + Person Recognition):
- Speakers (ASRX): {speaker_list}
- Faces (Face): {face_list}
- Identified Persons (verified): {person_list}

VISUAL CONTEXT (YOLO + Places365):
- Objects: {visual_list}
- Places: {places_list}
- Actions: {actions_list}

THIS CHUNK'S CONTENT:
"{text_content}"

Based on ALL the above information, provide accurate analysis:

1. **Who** (use verified names if available, e.g., "John (SPEAKER_1)"):
   - List characters with confidence level

2. **What** (key action in this specific moment)

3. **When** (temporal position: beginning/middle/end of scene)

4. **Where** (location from video or None)

5. **Why** (purpose of this specific action)

6. **How** (manner: tone, emotion, expression)

7. **Emotion/Tone** (specific emotions detected)

8. **Key Actions** (verbs describing what's happening)

Output format:
Who: [names with source]
What: [action]
When: [position]
Where: [location or None]
Why: [purpose]
How: [manner]
Emotion: [emotion]
Actions: [verb1, verb2]
---
Summary: [2-3 sentence detailed summary connecting to scene]"""

    result = call_llm(prompt)
    return result


def parse_5w1h_summary(result_text):
    """Parse 5W1H and summary from LLM response"""

    data = {
        "who": "",
        "what": "",
        "when": "",
        "where": "",
        "why": "",
        "how": "",
        "emotion": "",
        "actions": "",
        "summary": "",
    }

    try:
        parts = result_text.split("---")
        if len(parts) >= 2:
            five_w_one_h = parts[0].strip()
            data["summary"] = parts[1].strip().replace("Summary:", "").strip()

            for line in five_w_one_h.split("\n"):
                line = line.strip()
                if line.startswith("Who:"):
                    data["who"] = line.replace("Who:", "").strip()
                elif line.startswith("What:"):
                    data["what"] = line.replace("What:", "").strip()
                elif line.startswith("When:"):
                    data["when"] = line.replace("When:", "").strip()
                elif line.startswith("Where:"):
                    data["where"] = line.replace("Where:", "").strip()
                elif line.startswith("Why:"):
                    data["why"] = line.replace("Why:", "").strip()
                elif line.startswith("How:"):
                    data["how"] = line.replace("How:", "").strip()
                elif line.startswith("Emotion:"):
                    data["emotion"] = line.replace("Emotion:", "").strip()
                elif line.startswith("Actions:"):
                    data["actions"] = line.replace("Actions:", "").strip()
                    data["what"] = line.replace("What:", "").strip()
                elif line.startswith("When:"):
                    data["when"] = line.replace("When:", "").strip()
                elif line.startswith("Where:"):
                    data["where"] = line.replace("Where:", "").strip()
                elif line.startswith("Why:"):
                    data["why"] = line.replace("Why:", "").strip()
                elif line.startswith("How:"):
                    data["how"] = line.replace("How:", "").strip()
    except Exception as e:
        print(f"    ⚠️  Parse error: {e}")

    return data


def update_chunk_summary(
    chunk_id,
    summary_text,
    chunk_5w1h=None,
    identity_info=None,
    visual_stats=None,
    uuid=None,
):
    """Update chunk summary, 5W1H, identity, and visual in database"""

    conn = psycopg2.connect(**DB_CONFIG)
    cur = conn.cursor()

    metadata_obj = {}
    if chunk_5w1h:
        metadata_obj["chunk_5w1h"] = chunk_5w1h
    if identity_info:
        metadata_obj["chunk_identity"] = identity_info
    if visual_stats:
        try:
            metadata_obj["chunk_visual"] = (
                visual_stats
                if isinstance(visual_stats, dict)
                else json.loads(str(visual_stats))
            )
        except:
            metadata_obj["chunk_visual"] = {}

    if metadata_obj:
        metadata = json.dumps(metadata_obj)
        cur.execute(
            f"""
            UPDATE {SCHEMA}.chunks
            SET summary_text = %s, 
                metadata = COALESCE(metadata, '{{}}'::jsonb) || %s::jsonb,
                metadata_version = metadata_version + 1,
                updated_at = CURRENT_TIMESTAMP
            WHERE chunk_id = %s
        """,
            (summary_text, metadata, chunk_id),
        )
    else:
        cur.execute(
            f"""
            UPDATE {SCHEMA}.chunks
            SET summary_text = %s,
                content_version = content_version + 1,
                updated_at = CURRENT_TIMESTAMP
            WHERE chunk_id = %s
        """,
            (summary_text, chunk_id),
        )

    conn.commit()
    cur.close()
    conn.close()


def main():
    import argparse

    parser = argparse.ArgumentParser(description="Generate chunk summaries")
    parser.add_argument("--uuid", help="Process specific video UUID")
    parser.add_argument("--limit", type=int, help="Limit number of chunks")
    parser.add_argument("--dry-run", action="store_true", help="Print without saving")
    args = parser.parse_args()

    print(f"Fetching chunks (schema={SCHEMA})...")
    chunks = get_chunks_with_parents(uuid=args.uuid, limit=args.limit)
    print(f"Found {len(chunks)} chunks to process")

    if not chunks:
        print("No chunks need summary generation")
        return

    success = 0
    failed = 0

    for i, chunk in enumerate(chunks, 1):
        chunk_id = chunk["chunk_id"]
        print(f"\n[{i}/{len(chunks)}] {chunk_id}")

        if not chunk.get("text_content"):
            print("    ⚠️  No text_content, skipping")
            continue

        if not chunk.get("structured_summary"):
            print("    ⚠️  No parent 5W1H, skipping")
            continue

        print(f"    Text: {chunk['text_content'][:50]}...")
        result = generate_chunk_summary(chunk)

        if result:
            parsed = parse_5w1h_summary(result)
            summary_text = parsed.get("summary", result)
            chunk_5w1h = {k: v for k, v in parsed.items() if k != "summary" and v}

            speaker_ids = chunk.get("speaker_ids", [])
            face_ids = chunk.get("face_ids", [])
            visual_stats = chunk.get("visual_stats", {})

            identity_info = {
                "speakers": speaker_ids,
                "faces": [f"face_{x}" for x in face_ids] if face_ids else [],
            }

            print(f"    ✓ Summary: {summary_text[:80]}...")
            if chunk_5w1h:
                print(
                    f"    ✓ Chunk 5W1H: Who={chunk_5w1h.get('who', 'N/A')[:30]}, What={chunk_5w1h.get('what', 'N/A')[:30]}"
                )
            if identity_info["speakers"] or identity_info["faces"]:
                print(
                    f"    ✓ Identity: speakers={identity_info['speakers']}, faces={identity_info['faces']}"
                )
            if visual_stats:
                print(
                    f"    ✓ Visual: {list(visual_stats.keys()) if isinstance(visual_stats, dict) else 'present'}"
                )

            if not args.dry_run:
                update_chunk_summary(
                    chunk_id,
                    summary_text,
                    chunk_5w1h,
                    identity_info,
                    visual_stats,
                    args.uuid,
                )
            success += 1
        else:
            print("    ✗ Failed to generate summary")
            failed += 1

        if i % BATCH_SIZE == 0:
            print(f"\n  Batch complete ({success} success, {failed} failed)")
            time.sleep(DELAY_BETWEEN_BATCHES)

    print(f"\n{'=' * 50}")
    print(f"Done! Success: {success}, Failed: {failed}")
    if args.dry_run:
        print("(Dry run - no updates saved)")


if __name__ == "__main__":
    main()