fix(agent): increase max_tokens for merge requests to avoid truncated responses

normalize: 512 tokens (sufficient for single track metadata)
merge: 4096 tokens (needed for artists with many albums)

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -2,6 +2,12 @@
 # It is not intended for manual editing.
 version = 4
 
+[[package]]
+name = "adler2"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "320119579fcad9c21884f5c4861d16174d0e06250625266f50fe6898340abefa"
+
 [[package]]
 name = "aho-corasick"
 version = "1.1.4"
@@ -572,6 +578,15 @@ version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "19d374276b40fb8bbdee95aef7c7fa6b5316ec764510eb64b8dd0e2ed0d7e7f5"
 
+[[package]]
+name = "crc32fast"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9481c1c90cbf2ac953f07c8d4a58aa3945c425b7185c9154d67a65e4230da511"
+dependencies = [
+ "cfg-if",
+]
+
 [[package]]
 name = "crossbeam-channel"
 version = "0.5.15"
@@ -969,6 +984,16 @@ version = "0.5.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1d674e81391d1e1ab681a28d99df07927c6d4aa5b027d7da16ba32d1d21ecd99"
 
+[[package]]
+name = "flate2"
+version = "1.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "843fba2746e448b37e26a819579957415c8cef339bf08564fe8b7ddbd959573c"
+dependencies = [
+ "crc32fast",
+ "miniz_oxide",
+]
+
 [[package]]
 name = "flume"
 version = "0.11.1"
@@ -1017,6 +1042,7 @@ dependencies = [
  "chrono",
  "clap",
  "encoding_rs",
+ "id3",
  "reqwest 0.12.28",
  "serde",
  "serde_json",
@@ -1165,6 +1191,7 @@ dependencies = [
  "tokio",
  "tokio-util",
  "tower 0.4.13",
+ "tower-http",
  "tracing",
  "tracing-subscriber",
  "urlencoding",
@@ -1748,6 +1775,17 @@ version = "2.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
 
+[[package]]
+name = "id3"
+version = "1.16.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "965c5e6a62a241f2f673df956ea5f52c27780bc1031855890a551ed9b869e2d1"
+dependencies = [
+ "bitflags 2.11.0",
+ "byteorder",
+ "flate2",
+]
+
 [[package]]
 name = "ident_case"
 version = "1.0.1"
@@ -2021,6 +2059,16 @@ version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "68354c5c6bd36d73ff3feceb05efa59b6acb7626617f4962be322a825e61f79a"
 
+[[package]]
+name = "miniz_oxide"
+version = "0.8.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fa76a2c86f704bdb222d66965fb3d63269ce38518b83cb0575fca855ebb6316"
+dependencies = [
+ "adler2",
+ "simd-adler32",
+]
+
 [[package]]
 name = "mio"
 version = "1.1.1"
@@ -3412,6 +3460,12 @@ dependencies = [
  "rand_core 0.6.4",
 ]
 
+[[package]]
+name = "simd-adler32"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "703d5c7ef118737c72f1af64ad2f6f8c5e1921f818cdcb97b8fe6fc69bf66214"
+
 [[package]]
 name = "simple_asn1"
 version = "0.6.4"

furumi-agent/Cargo.toml

@@ -14,6 +14,7 @@ serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 sqlx = { version = "0.8", features = ["runtime-tokio-rustls", "postgres", "chrono", "uuid", "migrate"] }
 symphonia = { version = "0.5", default-features = false, features = ["mp3", "aac", "flac", "vorbis", "wav", "alac", "adpcm", "pcm", "mpa", "isomp4", "ogg", "aiff", "mkv"] }
+id3 = "1"
 thiserror = "2.0"
 tokio = { version = "1.50", features = ["full"] }
 tracing = "0.1"

furumi-agent/src/ingest/metadata.rs

@@ -19,9 +19,25 @@ pub struct RawMetadata {
     pub duration_secs: Option<f64>,
 }
 
-/// Extract metadata from an audio file using Symphonia.
+/// Extract metadata from an audio file.
+/// For MP3, falls back to the `id3` crate when Symphonia cannot probe the file
+/// (e.g., ID3 tag with large embedded cover art exceeds Symphonia's 1 MB probe limit).
 /// Must be called from a blocking context (spawn_blocking).
 pub fn extract(path: &Path) -> anyhow::Result<RawMetadata> {
+    match extract_via_symphonia(path) {
+        Ok(meta) => return Ok(meta),
+        Err(e) => {
+            let is_mp3 = path.extension().and_then(|e| e.to_str()).map(|e| e.eq_ignore_ascii_case("mp3")).unwrap_or(false);
+            if is_mp3 {
+                tracing::debug!(error = %e, "Symphonia failed on MP3, falling back to id3 crate");
+                return extract_mp3_via_id3(path);
+            }
+            return Err(e);
+        }
+    }
+}
+
+fn extract_via_symphonia(path: &Path) -> anyhow::Result<RawMetadata> {
     let file = std::fs::File::open(path)?;
     let mss = MediaSourceStream::new(Box::new(file), Default::default());
 
@@ -66,6 +82,25 @@ pub fn extract(path: &Path) -> anyhow::Result<RawMetadata> {
     Ok(meta)
 }
 
+/// Read MP3 tags via the `id3` crate. Duration is not available this way.
+fn extract_mp3_via_id3(path: &Path) -> anyhow::Result<RawMetadata> {
+    use id3::TagLike;
+
+    let tag = id3::Tag::read_from_path(path)
+        .map_err(|e| anyhow::anyhow!("id3 read failed: {}", e))?;
+
+    let mut meta = RawMetadata::default();
+    meta.title = tag.title().map(|s| fix_encoding(s.to_owned()));
+    meta.artist = tag.artist().map(|s| fix_encoding(s.to_owned()));
+    meta.album = tag.album().map(|s| fix_encoding(s.to_owned()));
+    meta.year = tag.year().and_then(|y| u32::try_from(y).ok());
+    meta.track_number = tag.track();
+    meta.genre = tag.genre().map(|s: &str| fix_encoding(s.to_owned()));
+    // duration_secs remains None — acceptable for large-cover files
+
+    Ok(meta)
+}
+
 fn extract_tags(tags: &[symphonia::core::meta::Tag], meta: &mut RawMetadata) {
     for tag in tags {
         let value = fix_encoding(tag.value.to_string());

furumi-agent/src/ingest/normalize.rs

@@ -25,16 +25,37 @@ pub async fn normalize(
 ) -> anyhow::Result<NormalizedFields> {
     let user_message = build_user_message(raw, hints, similar_artists, similar_albums, folder_ctx);
 
+    let schema = normalize_schema();
     let response = call_ollama(
         &state.config.ollama_url,
         &state.config.ollama_model,
         &state.system_prompt,
         &user_message,
         state.config.ollama_auth.as_deref(),
+        0.5,
+        512,
+        Some(("normalized_metadata", schema.clone())),
     )
     .await?;
 
-    parse_response(&response)
+    match parse_response(&response) {
+        Ok(fields) => Ok(fields),
+        Err(e) => {
+            tracing::warn!(error = %e, "LLM parse failed, retrying with higher frequency_penalty");
+            let response2 = call_ollama(
+                &state.config.ollama_url,
+                &state.config.ollama_model,
+                &state.system_prompt,
+                &user_message,
+                state.config.ollama_auth.as_deref(),
+                1.5,
+                512,
+                Some(("normalized_metadata", schema)),
+            )
+            .await?;
+            parse_response(&response2)
+        }
+    }
 }
 
 fn build_user_message(
@@ -113,32 +134,49 @@ fn build_user_message(
 }
 
 #[derive(Serialize)]
-struct OllamaRequest {
+struct ChatRequest {
     model: String,
-    messages: Vec<OllamaMessage>,
-    format: String,
+    messages: Vec<ChatMessage>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    response_format: Option<ChatResponseFormat>,
     stream: bool,
-    options: OllamaOptions,
+    temperature: f64,
+    max_tokens: u32,
+    frequency_penalty: f64,
 }
 
 #[derive(Serialize)]
-struct OllamaMessage {
+struct ChatMessage {
     role: String,
     content: String,
 }
 
 #[derive(Serialize)]
-struct OllamaOptions {
-    temperature: f64,
+struct ChatResponseFormat {
+    #[serde(rename = "type")]
+    kind: String,
+    json_schema: JsonSchemaWrapper,
+}
+
+#[derive(Serialize)]
+struct JsonSchemaWrapper {
+    name: String,
+    strict: bool,
+    schema: serde_json::Value,
 }
 
 #[derive(Deserialize)]
-struct OllamaResponse {
-    message: OllamaResponseMessage,
+struct ChatResponse {
+    choices: Vec<ChatChoice>,
 }
 
 #[derive(Deserialize)]
-struct OllamaResponseMessage {
+struct ChatChoice {
+    message: ChatResponseMessage,
+}
+
+#[derive(Deserialize)]
+struct ChatResponseMessage {
     content: String,
 }
 
@@ -148,30 +186,40 @@ pub async fn call_ollama(
     system_prompt: &str,
     user_message: &str,
     auth: Option<&str>,
+    frequency_penalty: f64,
+    max_tokens: u32,
+    schema: Option<(&str, serde_json::Value)>,
 ) -> anyhow::Result<String> {
     let client = reqwest::Client::builder()
         .timeout(std::time::Duration::from_secs(120))
         .build()?;
 
-    let request = OllamaRequest {
+    let response_format = schema.map(|(name, schema)| ChatResponseFormat {
+        kind: "json_schema".to_owned(),
+        json_schema: JsonSchemaWrapper { name: name.to_owned(), strict: true, schema },
+    });
+
+    let request = ChatRequest {
         model: model.to_owned(),
         messages: vec![
-            OllamaMessage {
+            ChatMessage {
                 role: "system".to_owned(),
                 content: system_prompt.to_owned(),
             },
-            OllamaMessage {
+            ChatMessage {
                 role: "user".to_owned(),
                 content: user_message.to_owned(),
             },
         ],
-        format: "json".to_owned(),
+        response_format,
         stream: false,
-        options: OllamaOptions { temperature: 0.1 },
+        temperature: 0.1,
+        max_tokens,
+        frequency_penalty,
     };
 
-    let url = format!("{}/api/chat", base_url.trim_end_matches('/'));
-    tracing::info!(%url, model, prompt_len = user_message.len(), "Calling Ollama API...");
+    let url = format!("{}/v1/chat/completions", base_url.trim_end_matches('/'));
+    tracing::info!(%url, model, prompt_len = user_message.len(), "Calling LLM API...");
 
     let start = std::time::Instant::now();
     let mut req = client.post(&url).json(&request);
@@ -184,18 +232,45 @@ pub async fn call_ollama(
     if !resp.status().is_success() {
         let status = resp.status();
         let body = resp.text().await.unwrap_or_default();
-        tracing::error!(%status, body = &body[..body.len().min(500)], "Ollama API error");
-        anyhow::bail!("Ollama returned {}: {}", status, body);
+        tracing::error!(%status, body = &body[..body.len().min(500)], "LLM API error");
+        anyhow::bail!("LLM returned {}: {}", status, body);
     }
 
-    let ollama_resp: OllamaResponse = resp.json().await?;
+    let chat_resp: ChatResponse = resp.json().await?;
+    let content = chat_resp
+        .choices
+        .into_iter()
+        .next()
+        .ok_or_else(|| anyhow::anyhow!("LLM returned empty choices"))?
+        .message
+        .content;
     tracing::info!(
         elapsed_ms = elapsed.as_millis() as u64,
-        response_len = ollama_resp.message.content.len(),
-        "Ollama response received"
+        response_len = content.len(),
+        "LLM response received"
     );
-    tracing::debug!(raw_response = %ollama_resp.message.content, "LLM raw output");
-    Ok(ollama_resp.message.content)
+    tracing::debug!(raw_response = %content, "LLM raw output");
+    Ok(content)
+}
+
+fn normalize_schema() -> serde_json::Value {
+    serde_json::json!({
+        "type": "object",
+        "properties": {
+            "artist":           { "type": ["string", "null"] },
+            "album":            { "type": ["string", "null"] },
+            "title":            { "type": ["string", "null"] },
+            "year":             { "type": ["integer", "null"] },
+            "track_number":     { "type": ["integer", "null"] },
+            "genre":            { "type": ["string", "null"] },
+            "featured_artists": { "type": "array", "items": { "type": "string" } },
+            "release_kind":     { "type": ["string", "null"] },
+            "confidence":       { "type": ["number", "null"] },
+            "notes":            { "type": ["string", "null"] }
+        },
+        "required": ["artist", "album", "title", "year", "track_number", "genre", "featured_artists", "release_kind", "confidence", "notes"],
+        "additionalProperties": false
+    })
 }
 
 /// Parse the LLM JSON response into NormalizedFields.
@@ -222,6 +297,7 @@ fn parse_response(response: &str) -> anyhow::Result<NormalizedFields> {
         genre: Option<String>,
         #[serde(default)]
         featured_artists: Vec<String>,
+        #[serde(rename = "release_kind")]
         release_type: Option<String>,
         confidence: Option<f64>,
         notes: Option<String>,

furumi-agent/src/merge.rs

@@ -35,12 +35,39 @@ pub async fn propose_merge(state: &Arc<AppState>, merge_id: Uuid) -> anyhow::Res
 
     let user_message = build_merge_message(&artists_data);
 
+    let schema = serde_json::json!({
+        "type": "object",
+        "properties": {
+            "canonical_artist_name": { "type": "string" },
+            "winner_artist_id":      { "type": "integer" },
+            "album_mappings": {
+                "type": "array",
+                "items": {
+                    "type": "object",
+                    "properties": {
+                        "source_album_id":     { "type": "integer" },
+                        "canonical_name":      { "type": "string" },
+                        "merge_into_album_id": { "type": ["integer", "null"] }
+                    },
+                    "required": ["source_album_id", "canonical_name", "merge_into_album_id"],
+                    "additionalProperties": false
+                }
+            },
+            "notes": { "type": "string" }
+        },
+        "required": ["canonical_artist_name", "winner_artist_id", "album_mappings", "notes"],
+        "additionalProperties": false
+    });
+
     let response = call_ollama(
         &state.config.ollama_url,
         &state.config.ollama_model,
         &state.merge_prompt,
         &user_message,
         state.config.ollama_auth.as_deref(),
+        0.5,
+        4096,
+        Some(("artist_merge", schema)),
     ).await?;
 
     let proposal = parse_merge_response(&response)?;