Kewton · Kewton · Mar 25, 2026 · Mar 25, 2026
diff --git a/README.md b/README.md
@@ -53,8 +53,8 @@ commandindex diff src/auth/jwt.rs src/auth/middleware.rs --format json
 
 | モデル | 次元数 | 特徴 |
 |---|---|---|
-| `nomic-embed-text` | 768 | デフォルト。英語中心 |
-| `qllama/bge-m3:q8_0` | 1024 | 多言語対応（日本語に強い） |
+| `qllama/bge-m3:q8_0` | 1024 | デフォルト。多言語対応（日本語に強い） |
+| `nomic-embed-text` | 768 | 英語中心 |
 
 ### 前提条件
 
@@ -63,10 +63,10 @@ commandindex diff src/auth/jwt.rs src/auth/middleware.rs --format json
 
 ```bash
 # デフォルトモデル
-ollama pull nomic-embed-text
-
-# 日本語対応モデル
 ollama pull qllama/bge-m3:q8_0
+
+# 英語中心モデル
+ollama pull nomic-embed-text
 ```
 
 ### モデル変更手順
@@ -75,11 +75,19 @@ ollama pull qllama/bge-m3:q8_0
 
 ```toml
 [embedding]
-model = "qllama/bge-m3:q8_0"
+model = "nomic-embed-text"
 ```
 
 > **注意:** モデル変更後の再生成にはファイル数に応じた時間がかかります。
 
+### v0.x.x からの移行
+
+v0.x.x 以前からアップグレードした場合、デフォルトモデルが `nomic-embed-text` から `qllama/bge-m3:q8_0` に変更されています。
+
+1. 新しいデフォルトモデルをインストール: `ollama pull qllama/bge-m3:q8_0`
+2. `commandindex embed` または `commandindex index --with-embedding` を実行すると、旧モデルの embedding は自動的に削除され、新モデルで再生成されます。
+3. 旧モデルを引き続き使用する場合は、`commandindex.toml` に `[embedding]` セクションで `model = "nomic-embed-text"` を指定してください。
+
 ## 開発
 
 ### 前提条件

diff --git a/src/cli/embed.rs b/src/cli/embed.rs
@@ -12,7 +12,7 @@ use std::time::{Duration, Instant};
 
 use crate::config::{ConfigError, load_config};
 use crate::embedding::store::{EmbeddingStore, EmbeddingStoreError};
-use crate::embedding::{EmbeddingError, create_provider};
+use crate::embedding::{EmbeddingError, create_provider, model_not_found_hint};
 use crate::indexer::manifest::{Manifest, ManifestError};
 use crate::indexer::reader::{IndexReaderWrapper, ReaderError};
 
@@ -141,17 +141,13 @@ pub fn run(path: &Path, commandindex_dir: &Path) -> Result<EmbedSummary, EmbedEr
     let tantivy_dir = crate::indexer::index_dir(commandindex_dir);
     let reader = IndexReaderWrapper::open(&tantivy_dir)?;
 
-    // 6.5. Delete stale embeddings from previous model
     let model_name = provider.model_name();
-    let stale_deleted = store.delete_stale_model_embeddings(model_name)?;
-    if stale_deleted > 0 {
-        eprintln!("Info: Deleted {stale_deleted} stale embeddings from previous model.");
-    }
 
     let mut total_sections: u64 = 0;
     let mut generated: u64 = 0;
     let mut cached: u64 = 0;
     let mut failed: u64 = 0;
+    let mut stale_deleted = false;
 
     // 7. Process each file entry
     for entry in &manifest.files {
@@ -185,6 +181,15 @@ pub fn run(path: &Path, commandindex_dir: &Path) -> Result<EmbedSummary, EmbedEr
         // Generate embeddings
         match provider.embed(&texts) {
             Ok(embeddings) => {
+                // Delete stale embeddings once after first successful embed
+                if !stale_deleted {
+                    let count = store.delete_stale_model_embeddings(model_name)?;
+                    if count > 0 {
+                        eprintln!("Info: Deleted {count} stale embeddings from previous model.");
+                    }
+                    stale_deleted = true;
+                }
+
                 if sections.len() != embeddings.len() {
                     eprintln!(
                         "Warning: section/embedding count mismatch for {}: {} sections, {} embeddings",
@@ -221,6 +226,16 @@ pub fn run(path: &Path, commandindex_dir: &Path) -> Result<EmbedSummary, EmbedEr
                     }
                 }
             }
+            Err(EmbeddingError::ModelNotFound(model)) => {
+                eprintln!("{}", model_not_found_hint(&model));
+                return Ok(EmbedSummary {
+                    total_sections,
+                    generated,
+                    cached,
+                    failed,
+                    duration: start.elapsed(),
+                });
+            }
             Err(e) => {
                 eprintln!(
                     "Warning: embedding generation failed for {}: {e}",

diff --git a/src/cli/index.rs b/src/cli/index.rs
@@ -25,7 +25,7 @@ use chrono::{DateTime, Utc};
 
 use crate::config::{ConfigError, load_config};
 use crate::embedding::store::{EmbeddingStore, EmbeddingStoreError};
-use crate::embedding::{EmbeddingError, create_provider};
+use crate::embedding::{EmbeddingError, create_provider, model_not_found_hint};
 use crate::indexer::diff::{DiffError, detect_changes, scan_files};
 use crate::indexer::manifest::{
     self, FileEntry, FileType, Manifest, ManifestError, to_relative_path_string,
@@ -929,18 +929,15 @@ fn generate_embeddings_for_manifest(
     let store = EmbeddingStore::open(&db_path)?;
     store.create_tables()?;
 
-    // Delete stale embeddings from previous model
     let model_name = provider.model_name();
-    let stale_deleted = store.delete_stale_model_embeddings(model_name)?;
-    if stale_deleted > 0 {
-        eprintln!("Info: Deleted {stale_deleted} stale embeddings from previous model.");
-    }
 
     let tantivy_dir = crate::indexer::index_dir(commandindex_dir);
     let reader = IndexReaderWrapper::open(&tantivy_dir).map_err(|e| {
         IndexError::IndexCorrupted(format!("Failed to open tantivy for embedding: {e}"))
     })?;
 
+    let mut stale_deleted = false;
+
     for entry in &manifest.files {
         if store.has_current_embedding(&entry.path, &entry.hash, model_name)? {
             continue;
@@ -966,6 +963,15 @@ fn generate_embeddings_for_manifest(
 
         match provider.embed(&texts) {
             Ok(embeddings) => {
+                // Delete stale embeddings once after first successful embed
+                if !stale_deleted {
+                    let count = store.delete_stale_model_embeddings(model_name)?;
+                    if count > 0 {
+                        eprintln!("Info: Deleted {count} stale embeddings from previous model.");
+                    }
+                    stale_deleted = true;
+                }
+
                 let dimension = provider.dimension();
                 let model = provider.model_name();
                 if sections.len() != embeddings.len() {
@@ -995,6 +1001,10 @@ fn generate_embeddings_for_manifest(
                     );
                 }
             }
+            Err(EmbeddingError::ModelNotFound(model)) => {
+                eprintln!("{}", model_not_found_hint(&model));
+                return Ok(());
+            }
             Err(e) => {
                 eprintln!(
                     "Warning: embedding generation failed for {}: {e}",

diff --git a/src/config/mod.rs b/src/config/mod.rs
@@ -458,7 +458,7 @@ fn resolve_config(raw: RawConfig, sources: Vec<ConfigSource>) -> AppConfig {
     let embedding = if let Some(emb) = raw.embedding {
         EmbeddingConfig {
             provider: emb.provider.unwrap_or_default(),
-            model: emb.model.unwrap_or_else(|| "nomic-embed-text".to_string()),
+            model: emb.model.unwrap_or_else(crate::embedding::default_model),
             endpoint: emb
                 .endpoint
                 .unwrap_or_else(|| "http://localhost:11434".to_string()),
@@ -687,7 +687,7 @@ mod tests {
         assert_eq!(config.search.snippet_lines, 2);
         assert_eq!(config.search.snippet_chars, 120);
         assert_eq!(config.embedding.provider, ProviderType::Ollama);
-        assert_eq!(config.embedding.model, "nomic-embed-text");
+        assert_eq!(config.embedding.model, "qllama/bge-m3:q8_0");
         assert_eq!(config.embedding.endpoint, "http://localhost:11434");
         assert!(config.embedding.api_key.is_none());
         assert_eq!(config.rerank.model, "llama3");
@@ -1015,7 +1015,7 @@ timeout_secs = 60
             },
             embedding: EmbeddingConfig {
                 provider: ProviderType::Ollama,
-                model: "nomic-embed-text".to_string(),
+                model: "qllama/bge-m3:q8_0".to_string(),
                 endpoint: "http://localhost:11434".to_string(),
                 api_key: Some("secret".to_string()),
             },

diff --git a/src/embedding/mod.rs b/src/embedding/mod.rs
@@ -79,8 +79,8 @@ pub enum ProviderType {
 // EmbeddingConfig
 // ---------------------------------------------------------------------------
 
-fn default_model() -> String {
-    "nomic-embed-text".to_string()
+pub(crate) fn default_model() -> String {
+    "qllama/bge-m3:q8_0".to_string()
 }
 
 fn default_endpoint() -> String {
@@ -133,6 +133,13 @@ impl EmbeddingConfig {
 // Shared utilities for providers
 // ---------------------------------------------------------------------------
 
+/// ModelNotFound時のインストール案内メッセージを返す。
+pub(crate) fn model_not_found_hint(model: &str) -> String {
+    format!(
+        "Model '{model}' not found. Install it with:\n  ollama pull {model}\nThen retry the command."
+    )
+}
+
 /// HTTPレスポンスのステータスコードをEmbeddingErrorに変換する。
 /// OllamaProvider・OpenAiProvider共通のエラーハンドリング。
 pub(crate) fn map_status_to_error(
@@ -246,7 +253,7 @@ mod tests {
     fn test_embedding_config_default() {
         let config = EmbeddingConfig::default();
         assert_eq!(config.provider, ProviderType::Ollama);
-        assert_eq!(config.model, "nomic-embed-text");
+        assert_eq!(config.model, "qllama/bge-m3:q8_0");
         assert_eq!(config.endpoint, "http://localhost:11434");
         assert!(config.api_key.is_none());
     }
@@ -295,7 +302,7 @@ provider = "ollama"
 "#;
         let emb: EmbeddingConfig = toml::from_str(toml_str).unwrap();
         assert_eq!(emb.provider, ProviderType::Ollama);
-        assert_eq!(emb.model, "nomic-embed-text");
+        assert_eq!(emb.model, "qllama/bge-m3:q8_0");
         assert_eq!(emb.endpoint, "http://localhost:11434");
         assert!(emb.api_key.is_none());
     }
@@ -419,4 +426,36 @@ provider = "ollama"
         let provider = create_provider(&config).unwrap();
         assert_eq!(provider.provider_name(), "openai");
     }
+
+    // --- model_not_found_hint ---
+
+    #[test]
+    fn test_model_not_found_hint_contains_model_name() {
+        let hint = model_not_found_hint("qllama/bge-m3:q8_0");
+        assert!(hint.contains("qllama/bge-m3:q8_0"));
+        assert!(hint.contains("ollama pull"));
+    }
+
+    #[test]
+    fn test_model_not_found_hint_contains_retry_instruction() {
+        let hint = model_not_found_hint("some-model");
+        assert!(hint.contains("some-model"));
+        assert!(hint.contains("retry"));
+    }
+
+    // --- default_model returns bge-m3 ---
+
+    #[test]
+    fn test_default_model_is_bge_m3() {
+        assert_eq!(default_model(), "qllama/bge-m3:q8_0");
+    }
+
+    // --- default config dimension is 1024 (bge-m3) ---
+
+    #[test]
+    fn test_default_config_dimension_is_1024() {
+        let config = EmbeddingConfig::default();
+        let provider = create_provider(&config).unwrap();
+        assert_eq!(provider.dimension(), 1024);
+    }
 }
diff --git a/tests/e2e_semantic_hybrid.rs b/tests/e2e_semantic_hybrid.rs
@@ -337,6 +337,13 @@ fn test_embed_without_ollama_fails() {
     let (dir, _commandindex_dir) =
         setup_semantic_test_dir().expect("test_embed_without_ollama_fails: setup");
 
+    // Force embed to fail by pointing to unreachable endpoint
+    fs::write(
+        dir.path().join("commandindex.toml"),
+        "[embedding]\nendpoint = \"http://127.0.0.1:19999\"\n",
+    )
+    .expect("test_embed_without_ollama_fails: write config");
+
     // Running embed without Ollama available exits successfully but reports
     // failures in stderr warnings and "Failed: N" in stdout.
     let output = common::cmd()