synops/tools/synops-highlight/src/main.rs

// synops-highlight — AI-kuratert highlight reel fra podcast-transkripsjon.
//
// Analyserer fullstendig transkripsjon etter innspilling og finner
// 5-10 klippverdige øyeblikk: humor, emosjonelle topper, sterke meninger,
// punchlines og narrative høydepunkter.
//
// Input: --media-node-id <uuid> (episodenode med transkripsjon)
// Output: JSON med foreslåtte klipp (tidsstempler, teksting, hashtags, score)
// Med --write: oppretter klipp-noder og edges i PG.
//
// Miljøvariabler:
//   DATABASE_URL       — PostgreSQL-tilkobling (påkrevd)
//   AI_GATEWAY_URL     — LiteLLM gateway (default: http://localhost:4000)
//   LITELLM_MASTER_KEY — API-nøkkel for LiteLLM
//   AI_HIGHLIGHT_MODEL — Modellalias (default: synops/high)
//
// Ref: docs/proposals/auto_highlight_reel.md
//      docs/retninger/unix_filosofi.md

use clap::Parser;
use serde::{Deserialize, Serialize};
use std::process;
use uuid::Uuid;

/// AI-kuratert highlight reel fra podcast-transkripsjon.
#[derive(Parser)]
#[command(name = "synops-highlight", about = "Generer highlight reel fra transkripsjon")]
struct Cli {
    /// Media-node-ID (episode med transkripsjon)
    #[arg(long)]
    media_node_id: Uuid,

    /// Bruker-ID som utløste analysen
    #[arg(long)]
    requested_by: Option<Uuid>,

    /// Podcast-samling (for belongs_to-edge)
    #[arg(long)]
    collection_id: Option<Uuid>,

    /// Skriv klipp-noder og edges til database
    #[arg(long)]
    write: bool,
}

// --- Database-rader ---

#[derive(sqlx::FromRow)]
struct MediaNode {
    title: Option<String>,
    #[allow(dead_code)]
    content: Option<String>,
    created_by: Option<Uuid>,
    #[allow(dead_code)]
    metadata: serde_json::Value,
}

#[derive(sqlx::FromRow)]
struct TranscriptSegment {
    #[allow(dead_code)]
    seq: i32,
    start_ms: i32,
    end_ms: i32,
    content: String,
}

// --- LLM request/response (OpenAI-kompatibel) ---

#[derive(Serialize)]
struct ChatRequest {
    model: String,
    messages: Vec<ChatMessage>,
    temperature: f32,
    response_format: ResponseFormat,
}

#[derive(Serialize)]
struct ResponseFormat {
    r#type: String,
}

#[derive(Serialize)]
struct ChatMessage {
    role: String,
    content: String,
}

#[derive(Deserialize)]
struct ChatResponse {
    choices: Vec<Choice>,
    #[serde(default)]
    usage: Option<UsageInfo>,
    #[serde(default)]
    model: Option<String>,
}

#[derive(Deserialize, Clone)]
struct UsageInfo {
    #[serde(default)]
    prompt_tokens: i64,
    #[serde(default)]
    completion_tokens: i64,
}

#[derive(Deserialize)]
struct Choice {
    message: MessageContent,
}

#[derive(Deserialize)]
struct MessageContent {
    content: Option<String>,
}

// --- AI-analysens output ---

#[derive(Deserialize, Debug)]
struct HighlightResponse {
    #[serde(default)]
    highlights: Vec<Highlight>,
}

#[derive(Deserialize, Serialize, Debug, Clone)]
struct Highlight {
    /// Starttid i millisekunder
    start_ms: i64,
    /// Sluttid i millisekunder
    end_ms: i64,
    /// Klippverdighets-score 0.0-1.0
    score: f64,
    /// Kategori: humor, emotion, opinion, punchline, narrative
    reason: String,
    /// Foreslått teksting for sosiale medier
    suggested_caption: String,
    /// Det sterkeste sitatet (thumbnail-tekst)
    quote: String,
    /// Foreslåtte hashtags
    #[serde(default)]
    hashtags: Vec<String>,
}

const SYSTEM_PROMPT: &str = r#"Du er en podcast-produsent som identifiserer de beste øyeblikkene i en podcast-episode. Analyser transkripsjonen og finn 5-10 klippverdige øyeblikk (15-45 sekunder hver).

For hvert høydepunkt, vurder:
- **humor**: Morsomme øyeblikk, vitser, latter-øyeblikk
- **emotion**: Emosjonelle topper, sårbare øyeblikk, sterke reaksjoner
- **opinion**: Kontroversielle eller sterke meninger som engasjerer
- **punchline**: Slagkraftige formuleringer, one-liners, quotable moments
- **narrative**: Narrative vendepunkter, overraskende avsløringer

Returner KUN et JSON-objekt med denne strukturen:
{
  "highlights": [
    {
      "start_ms": 12000,
      "end_ms": 45000,
      "score": 0.92,
      "reason": "punchline",
      "suggested_caption": "Kort, engasjerende tekst for sosiale medier (maks 280 tegn)",
      "quote": "Det sterkeste sitatet fra klippet",
      "hashtags": ["podcast", "tema1", "tema2"]
    }
  ]
}

Regler:
- Hvert klipp bør være 15-45 sekunder (15000-45000ms)
- Sorter etter score (høyeste først)
- Bruk start_ms/end_ms fra transkripsjonen — ikke dikt opp tidsstempler
- suggested_caption skal være catchy og fungere på TikTok/Instagram/Twitter
- quote skal være et direkte sitat som fungerer som thumbnail-tekst
- hashtags: 3-5 relevante, norske hashtags per klipp
- Returner tom highlights-liste hvis innholdet ikke har klippverdige øyeblikk
- Foretrekk øyeblikk som fungerer alene uten kontekst"#;

#[tokio::main]
async fn main() {
    synops_common::logging::init("synops_highlight");

    let cli = Cli::parse();

    if cli.write && cli.requested_by.is_none() {
        eprintln!("Feil: --requested-by er påkrevd sammen med --write");
        process::exit(1);
    }

    if let Err(e) = run(cli).await {
        eprintln!("Feil: {e}");
        process::exit(1);
    }
}

async fn run(cli: Cli) -> Result<(), String> {
    let db = synops_common::db::connect().await?;
    let media_node_id = cli.media_node_id;

    // 1. Hent medianode
    let media = sqlx::query_as::<_, MediaNode>(
        "SELECT title, content, created_by, metadata FROM nodes WHERE id = $1",
    )
    .bind(media_node_id)
    .fetch_optional(&db)
    .await
    .map_err(|e| format!("PG-feil ved henting av node: {e}"))?
    .ok_or_else(|| format!("Node {media_node_id} finnes ikke"))?;

    // 2. Hent transkripsjonssegmenter
    let segments = sqlx::query_as::<_, TranscriptSegment>(
        r#"SELECT seq, start_ms, end_ms, content
           FROM transcription_segments
           WHERE node_id = $1
           ORDER BY seq ASC"#,
    )
    .bind(media_node_id)
    .fetch_all(&db)
    .await
    .map_err(|e| format!("PG-feil ved henting av segmenter: {e}"))?;

    if segments.is_empty() {
        let result = serde_json::json!({
            "status": "skipped",
            "reason": "no_transcription",
            "media_node_id": media_node_id.to_string()
        });
        println!("{}", serde_json::to_string_pretty(&result).unwrap());
        return Ok(());
    }

    tracing::info!(
        media_node_id = %media_node_id,
        segments = segments.len(),
        "Hentet transkripsjon, sender til AI for highlight-analyse"
    );

    // 3. Bygg transkripsjon med tidsstempler for AI
    let transcript = build_timestamped_transcript(&segments);

    let episode_title = media.title.unwrap_or_else(|| "Ukjent episode".to_string());
    let user_content = format!(
        "Episode: {episode_title}\n\nTranskripsjon med tidsstempler (ms):\n\n{transcript}"
    );

    // 4. Kall LLM
    let (highlights_resp, llm_usage, llm_model) = call_llm(&user_content).await?;

    let highlights = highlights_resp.highlights;
    tracing::info!(
        media_node_id = %media_node_id,
        highlights_found = highlights.len(),
        "AI-analyse fullført"
    );

    if highlights.is_empty() {
        let result = serde_json::json!({
            "status": "completed",
            "media_node_id": media_node_id.to_string(),
            "highlights_found": 0,
            "clips_created": 0,
            "reason": "no_highlights_found"
        });
        println!("{}", serde_json::to_string_pretty(&result).unwrap());
        return Ok(());
    }

    // 5. Validér og filtrer highlights
    let last_segment_end = segments.last().map(|s| s.end_ms as i64).unwrap_or(0);
    let valid_highlights: Vec<&Highlight> = highlights
        .iter()
        .filter(|h| {
            h.start_ms >= 0
                && h.end_ms > h.start_ms
                && h.end_ms <= last_segment_end
                && (h.end_ms - h.start_ms) >= 10_000  // minst 10 sek
                && (h.end_ms - h.start_ms) <= 60_000  // maks 60 sek
                && h.score >= 0.0
                && h.score <= 1.0
        })
        .collect();

    tracing::info!(
        valid = valid_highlights.len(),
        total = highlights.len(),
        "Highlights validert"
    );

    let tokens_in = llm_usage.as_ref().map(|u| u.prompt_tokens).unwrap_or(0);
    let tokens_out = llm_usage.as_ref().map(|u| u.completion_tokens).unwrap_or(0);
    let model_id = llm_model.unwrap_or_else(|| "unknown".to_string());

    // 6. Skriv til database eller bare output
    let result = if cli.write {
        let requested_by = cli.requested_by.unwrap();
        let created_by = media.created_by.unwrap_or(media_node_id);
        let collection_id = cli.collection_id;

        let clips_created = write_highlights(
            &db,
            media_node_id,
            &valid_highlights,
            created_by,
            collection_id,
        )
        .await?;

        // Logg AI-ressursforbruk
        log_resource_usage(
            &db,
            media_node_id,
            media.created_by,
            &model_id,
            tokens_in,
            tokens_out,
            requested_by,
        )
        .await;

        serde_json::json!({
            "status": "completed",
            "media_node_id": media_node_id.to_string(),
            "highlights_found": highlights.len(),
            "highlights_valid": valid_highlights.len(),
            "clips_created": clips_created,
            "model": model_id,
            "tokens_in": tokens_in,
            "tokens_out": tokens_out,
        })
    } else {
        let highlight_data: Vec<&Highlight> = valid_highlights.iter().copied().collect();
        serde_json::json!({
            "status": "completed",
            "media_node_id": media_node_id.to_string(),
            "highlights_found": highlights.len(),
            "highlights_valid": valid_highlights.len(),
            "highlights": highlight_data,
            "model": model_id,
            "tokens_in": tokens_in,
            "tokens_out": tokens_out,
        })
    };

    println!(
        "{}",
        serde_json::to_string_pretty(&result)
            .map_err(|e| format!("JSON-serialisering feilet: {e}"))?
    );

    Ok(())
}

/// Bygg en tidsstemplet transkripsjon for AI-analyse.
/// Format: [12000-15000ms] Tekst her...
fn build_timestamped_transcript(segments: &[TranscriptSegment]) -> String {
    let mut lines = Vec::with_capacity(segments.len());
    for seg in segments {
        lines.push(format!("[{}-{}ms] {}", seg.start_ms, seg.end_ms, seg.content));
    }
    lines.join("\n")
}

/// Kall LiteLLM for highlight-analyse.
async fn call_llm(
    user_content: &str,
) -> Result<(HighlightResponse, Option<UsageInfo>, Option<String>), String> {
    let gateway_url =
        std::env::var("AI_GATEWAY_URL").unwrap_or_else(|_| "http://localhost:4000".to_string());
    let api_key = std::env::var("LITELLM_MASTER_KEY").unwrap_or_default();
    let model =
        std::env::var("AI_HIGHLIGHT_MODEL").unwrap_or_else(|_| "synops/high".to_string());

    let request = ChatRequest {
        model,
        messages: vec![
            ChatMessage {
                role: "system".to_string(),
                content: SYSTEM_PROMPT.to_string(),
            },
            ChatMessage {
                role: "user".to_string(),
                content: user_content.to_string(),
            },
        ],
        temperature: 0.4,
        response_format: ResponseFormat {
            r#type: "json_object".to_string(),
        },
    };

    let client = reqwest::Client::new();
    let url = format!("{gateway_url}/v1/chat/completions");

    let resp = client
        .post(&url)
        .header("Authorization", format!("Bearer {api_key}"))
        .header("Content-Type", "application/json")
        .json(&request)
        .timeout(std::time::Duration::from_secs(120))
        .send()
        .await
        .map_err(|e| format!("LiteLLM-kall feilet: {e}"))?;

    if !resp.status().is_success() {
        let status = resp.status();
        let body = resp.text().await.unwrap_or_default();
        return Err(format!("LiteLLM returnerte {status}: {body}"));
    }

    let chat_resp: ChatResponse = resp
        .json()
        .await
        .map_err(|e| format!("Kunne ikke parse LiteLLM-respons: {e}"))?;

    let content = chat_resp
        .choices
        .first()
        .and_then(|c| c.message.content.as_deref())
        .ok_or("LiteLLM returnerte ingen content")?;

    let highlights: HighlightResponse = serde_json::from_str(content)
        .map_err(|e| format!("Kunne ikke parse LLM JSON: {e}. Rå output: {content}"))?;

    Ok((highlights, chat_resp.usage, chat_resp.model))
}

/// Opprett klipp-noder og edges i PG for godkjente highlights.
/// Returnerer antall klipp opprettet.
async fn write_highlights(
    db: &sqlx::PgPool,
    media_node_id: Uuid,
    highlights: &[&Highlight],
    created_by: Uuid,
    collection_id: Option<Uuid>,
) -> Result<u32, String> {
    let mut clips_created = 0u32;

    for highlight in highlights {
        let clip_id = Uuid::now_v7();

        let clip_metadata = serde_json::json!({
            "ai_generated": true,
            "clip": {
                "start_ms": highlight.start_ms,
                "end_ms": highlight.end_ms,
                "score": highlight.score,
                "reason": highlight.reason,
                "suggested_caption": highlight.suggested_caption,
                "quote": highlight.quote,
                "hashtags": highlight.hashtags,
                "source_type": "highlight_reel",
            }
        });

        // Opprett klipp-node
        sqlx::query(
            r#"INSERT INTO nodes (id, node_kind, title, content, visibility, metadata, created_by)
               VALUES ($1, 'content', $2, $3, 'hidden', $4, $5)
               ON CONFLICT (id) DO NOTHING"#,
        )
        .bind(clip_id)
        .bind(&highlight.suggested_caption)
        .bind(&highlight.quote)
        .bind(&clip_metadata)
        .bind(created_by)
        .execute(db)
        .await
        .map_err(|e| format!("PG insert clip-node feilet: {e}"))?;

        // Edge: klipp → episode (derived_from)
        let edge_id = Uuid::now_v7();
        sqlx::query(
            r#"INSERT INTO edges (id, source_id, target_id, edge_type, metadata, system, created_by)
               VALUES ($1, $2, $3, 'derived_from', '{"origin": "highlight_reel"}'::jsonb, true, $4)
               ON CONFLICT (source_id, target_id, edge_type) DO NOTHING"#,
        )
        .bind(edge_id)
        .bind(clip_id)
        .bind(media_node_id)
        .bind(created_by)
        .execute(db)
        .await
        .map_err(|e| format!("PG insert derived_from-edge feilet: {e}"))?;

        // Edge: klipp → samling (belongs_to) hvis collection_id er gitt
        if let Some(coll_id) = collection_id {
            let edge_id = Uuid::now_v7();
            sqlx::query(
                r#"INSERT INTO edges (id, source_id, target_id, edge_type, metadata, system, created_by)
                   VALUES ($1, $2, $3, 'belongs_to', '{"origin": "highlight_reel"}'::jsonb, true, $4)
                   ON CONFLICT (source_id, target_id, edge_type) DO NOTHING"#,
            )
            .bind(edge_id)
            .bind(clip_id)
            .bind(coll_id)
            .bind(created_by)
            .execute(db)
            .await
            .map_err(|e| format!("PG insert belongs_to-edge feilet: {e}"))?;
        }

        clips_created += 1;

        tracing::info!(
            clip_id = %clip_id,
            start_ms = highlight.start_ms,
            end_ms = highlight.end_ms,
            score = highlight.score,
            reason = %highlight.reason,
            "Highlight-klipp opprettet"
        );
    }

    tracing::info!(
        media_node_id = %media_node_id,
        clips_created = clips_created,
        "Highlights skrevet til database"
    );

    Ok(clips_created)
}

/// Logg AI-ressursforbruk til resource_usage_log.
async fn log_resource_usage(
    db: &sqlx::PgPool,
    node_id: Uuid,
    _created_by: Option<Uuid>,
    model_id: &str,
    tokens_in: i64,
    tokens_out: i64,
    requested_by: Uuid,
) {
    let collection_id: Option<Uuid> = sqlx::query_scalar(
        "SELECT e.target_id FROM edges e
         JOIN nodes n ON n.id = e.target_id
         WHERE e.source_id = $1 AND e.edge_type = 'belongs_to' AND n.node_kind = 'collection'
         LIMIT 1",
    )
    .bind(node_id)
    .fetch_optional(db)
    .await
    .ok()
    .flatten();

    if let Err(e) = sqlx::query(
        "INSERT INTO resource_usage_log (target_node_id, triggered_by, collection_id, resource_type, detail)
         VALUES ($1, $2, $3, $4, $5)",
    )
    .bind(node_id)
    .bind(Some(requested_by))
    .bind(collection_id)
    .bind("ai")
    .bind(serde_json::json!({
        "model_level": "standard",
        "model_id": model_id,
        "tokens_in": tokens_in,
        "tokens_out": tokens_out,
        "job_type": "highlight_extract"
    }))
    .execute(db)
    .await
    {
        tracing::warn!(error = %e, "Kunne ikke logge AI-ressursforbruk");
    }
}