ai-saas / saas-search

Commit 200fdddf9b8cac307bb67bd9eac4bef1eb9a0e22

Authored by tangwang
1 parent 654f20d1

embed norm

Showing 10 changed files with 128 additions and 44 deletions   Show diff stats
Inline Side-by-side
embeddings/README.md
  Show/Hide comments   View file @200fddd
... ... @@ -14,7 +14,7 @@
14 14 - **clip-as-service 客户端**：`clip_as_service_encoder.py`（图片向量，推荐）
15 15 - **向量化服务（FastAPI）**：`server.py`
16 16 - **统一配置**：`config.py`
17   -- **接口契约**：`protocols.ImageEncoderProtocol`（图片编码统一为 `encode_image_urls(urls, batch_size)`，本地 CN-CLIP 与 clip-as-service 均实现该接口）
  17 +- **接口契约**：`protocols.ImageEncoderProtocol`（图片编码统一为 `encode_image_urls(urls, batch_size, normalize_embeddings)`，本地 CN-CLIP 与 clip-as-service 均实现该接口）
18 18  
19 19 说明：历史上的云端 embedding 试验实现（DashScope）已从主仓库移除，当前仅维护 6005 这条统一向量服务链路。
20 20  
... ... @@ -29,10 +29,12 @@
29 29  
30 30 - `POST /embed/text`
31 31 - 入参：`["文本1", "文本2", ...]`
  32 + - 可选 query 参数：`normalize=true|false`（不传则使用服务端默认）
32 33 - 出参：`[[...], [...], ...]`（与输入按 index 对齐，失败直接报错）
33 34  
34 35 - `POST /embed/image`
35 36 - 入参：`["url或本地路径1", ...]`
  37 + - 可选 query 参数：`normalize=true|false`（不传则使用服务端默认）
36 38 - 出参：`[[...], [...], ...]`（与输入按 index 对齐，失败直接报错）
37 39  
38 40 ### 图片向量：clip-as-service（推荐）
... ... @@ -77,6 +79,7 @@ TEI_USE_GPU=0 ./scripts/start_tei_service.sh
77 79  
78 80 - `PORT`: 服务端口（默认 6005）
79 81 - `TEXT_MODEL_ID`, `TEXT_DEVICE`, `TEXT_BATCH_SIZE`, `TEXT_NORMALIZE_EMBEDDINGS`
  82 +- `IMAGE_NORMALIZE_EMBEDDINGS`（默认 true）
80 83 - `USE_CLIP_AS_SERVICE`, `CLIP_AS_SERVICE_SERVER`：图片向量（clip-as-service）
81 84 - `IMAGE_MODEL_NAME`, `IMAGE_DEVICE`：本地 CN-CLIP（当 `USE_CLIP_AS_SERVICE=false` 时）
82 85 - TEI 相关：`TEI_USE_GPU`、`TEI_VERSION`、`TEI_MAX_BATCH_TOKENS`、`TEI_MAX_CLIENT_BATCH_SIZE`、`TEI_HEALTH_TIMEOUT_SEC`
... ...
embeddings/clip_as_service_encoder.py
  Show/Hide comments   View file @200fddd
... ... @@ -79,6 +79,7 @@ class ClipAsServiceImageEncoder:
79 79 self,
80 80 urls: List[str],
81 81 batch_size: Optional[int] = None,
  82 + normalize_embeddings: bool = True,
82 83 ) -> List[np.ndarray]:
83 84 """
84 85 Encode a list of image URLs to vectors.
... ... @@ -117,10 +118,15 @@ class ClipAsServiceImageEncoder:
117 118 vec = np.asarray(row, dtype=np.float32)
118 119 if vec.ndim != 1 or vec.size == 0 or not np.isfinite(vec).all():
119 120 raise RuntimeError("clip-as-service returned invalid embedding vector")
  121 + if normalize_embeddings:
  122 + norm = float(np.linalg.norm(vec))
  123 + if not np.isfinite(norm) or norm <= 0.0:
  124 + raise RuntimeError("clip-as-service returned zero/invalid norm vector")
  125 + vec = vec / norm
120 126 out.append(vec)
121 127 return out
122 128  
123   - def encode_image_from_url(self, url: str) -> Optional[np.ndarray]:
  129 + def encode_image_from_url(self, url: str, normalize_embeddings: bool = True) -> Optional[np.ndarray]:
124 130 """Encode a single image URL. Returns 1024-dim vector or None."""
125   - results = self.encode_image_urls([url], batch_size=1)
  131 + results = self.encode_image_urls([url], batch_size=1, normalize_embeddings=normalize_embeddings)
126 132 return results[0] if results else None
... ...
embeddings/clip_model.py
  Show/Hide comments   View file @200fddd
... ... @@ -76,25 +76,27 @@ class ClipImageModel(object):
76 76 text_features /= text_features.norm(dim=-1, keepdim=True)
77 77 return text_features
78 78  
79   - def encode_image(self, image: Image.Image) -> Optional[np.ndarray]:
  79 + def encode_image(self, image: Image.Image, normalize_embeddings: bool = True) -> Optional[np.ndarray]:
80 80 if not isinstance(image, Image.Image):
81 81 raise ValueError("ClipImageModel.encode_image input must be a PIL.Image")
82 82 infer_data = self.preprocess(image).unsqueeze(0).to(self.device)
83 83 with torch.no_grad():
84 84 image_features = self.model.encode_image(infer_data)
85   - image_features /= image_features.norm(dim=-1, keepdim=True)
  85 + if normalize_embeddings:
  86 + image_features /= image_features.norm(dim=-1, keepdim=True)
86 87 return image_features.cpu().numpy().astype("float32")[0]
87 88  
88   - def encode_image_from_url(self, url: str) -> Optional[np.ndarray]:
  89 + def encode_image_from_url(self, url: str, normalize_embeddings: bool = True) -> Optional[np.ndarray]:
89 90 image_data = self.download_image(url)
90 91 image = self.validate_image(image_data)
91 92 image = self.preprocess_image(image)
92   - return self.encode_image(image)
  93 + return self.encode_image(image, normalize_embeddings=normalize_embeddings)
93 94  
94 95 def encode_image_urls(
95 96 self,
96 97 urls: List[str],
97 98 batch_size: Optional[int] = None,
  99 + normalize_embeddings: bool = True,
98 100 ) -> List[Optional[np.ndarray]]:
99 101 """
100 102 Encode a list of image URLs to vectors. Same interface as ClipAsServiceImageEncoder.
... ... @@ -106,19 +108,27 @@ class ClipImageModel(object):
106 108 Returns:
107 109 List of vectors (or None for failed items), same length as urls.
108 110 """
109   - return self.encode_batch(urls, batch_size=batch_size or 8)
  111 + return self.encode_batch(
  112 + urls,
  113 + batch_size=batch_size or 8,
  114 + normalize_embeddings=normalize_embeddings,
  115 + )
110 116  
111   - def encode_batch(self, images: List[Union[str, Image.Image]], batch_size: int = 8) -> List[Optional[np.ndarray]]:
  117 + def encode_batch(
  118 + self,
  119 + images: List[Union[str, Image.Image]],
  120 + batch_size: int = 8,
  121 + normalize_embeddings: bool = True,
  122 + ) -> List[Optional[np.ndarray]]:
112 123 results: List[Optional[np.ndarray]] = []
113 124 for i in range(0, len(images), batch_size):
114 125 batch = images[i : i + batch_size]
115 126 for img in batch:
116 127 if isinstance(img, str):
117   - results.append(self.encode_image_from_url(img))
  128 + results.append(self.encode_image_from_url(img, normalize_embeddings=normalize_embeddings))
118 129 elif isinstance(img, Image.Image):
119   - results.append(self.encode_image(img))
  130 + results.append(self.encode_image(img, normalize_embeddings=normalize_embeddings))
120 131 else:
121 132 results.append(None)
122 133 return results
123 134  
124   -
... ...
embeddings/config.py
  Show/Hide comments   View file @200fddd
... ... @@ -37,6 +37,7 @@ class EmbeddingConfig(object):
37 37  
38 38 # Service behavior
39 39 IMAGE_BATCH_SIZE = 8
  40 + IMAGE_NORMALIZE_EMBEDDINGS = os.getenv("IMAGE_NORMALIZE_EMBEDDINGS", "true").lower() in ("1", "true", "yes")
40 41  
41 42  
42 43 CONFIG = EmbeddingConfig()
... ...
embeddings/image_encoder.py
  Show/Hide comments   View file @200fddd
... ... @@ -26,7 +26,7 @@ class CLIPImageEncoder:
26 26 self.endpoint = f"{self.service_url}/embed/image"
27 27 logger.info("Creating CLIPImageEncoder instance with service URL: %s", self.service_url)
28 28  
29   - def _call_service(self, request_data: List[str]) -> List[Any]:
  29 + def _call_service(self, request_data: List[str], normalize_embeddings: bool = True) -> List[Any]:
30 30 """
31 31 Call the embedding service API.
32 32  
... ... @@ -39,6 +39,7 @@ class CLIPImageEncoder:
39 39 try:
40 40 response = requests.post(
41 41 self.endpoint,
  42 + params={"normalize": "true" if normalize_embeddings else "false"},
42 43 json=request_data,
43 44 timeout=60
44 45 )
... ... @@ -56,7 +57,7 @@ class CLIPImageEncoder:
56 57 """
57 58 raise NotImplementedError("encode_image with PIL Image is not supported by embedding service")
58 59  
59   - def encode_image_from_url(self, url: str) -> np.ndarray:
  60 + def encode_image_from_url(self, url: str, normalize_embeddings: bool = True) -> np.ndarray:
60 61 """
61 62 Generate image embedding via network service using URL.
62 63  
... ... @@ -66,7 +67,7 @@ class CLIPImageEncoder:
66 67 Returns:
67 68 Embedding vector
68 69 """
69   - response_data = self._call_service([url])
  70 + response_data = self._call_service([url], normalize_embeddings=normalize_embeddings)
70 71 if not response_data or len(response_data) != 1 or response_data[0] is None:
71 72 raise RuntimeError(f"No image embedding returned for URL: {url}")
72 73 vec = np.array(response_data[0], dtype=np.float32)
... ... @@ -77,7 +78,8 @@ class CLIPImageEncoder:
77 78 def encode_batch(
78 79 self,
79 80 images: List[Union[str, Image.Image]],
80   - batch_size: int = 8
  81 + batch_size: int = 8,
  82 + normalize_embeddings: bool = True,
81 83 ) -> List[np.ndarray]:
82 84 """
83 85 Encode a batch of images efficiently via network service.
... ... @@ -98,7 +100,7 @@ class CLIPImageEncoder:
98 100 results: List[np.ndarray] = []
99 101 for i in range(0, len(images), batch_size):
100 102 batch_urls = [str(u).strip() for u in images[i:i + batch_size]]
101   - response_data = self._call_service(batch_urls)
  103 + response_data = self._call_service(batch_urls, normalize_embeddings=normalize_embeddings)
102 104 if not response_data or len(response_data) != len(batch_urls):
103 105 raise RuntimeError(
104 106 f"Image embedding response length mismatch: expected {len(batch_urls)}, "
... ... @@ -119,6 +121,7 @@ class CLIPImageEncoder:
119 121 self,
120 122 urls: List[str],
121 123 batch_size: Optional[int] = None,
  124 + normalize_embeddings: bool = True,
122 125 ) -> List[np.ndarray]:
123 126 """
124 127 与 ClipImageModel / ClipAsServiceImageEncoder 一致的接口，供索引器 document_transformer 调用。
... ... @@ -130,4 +133,8 @@ class CLIPImageEncoder:
130 133 Returns:
131 134 与 urls 等长的向量列表
132 135 """
133   - return self.encode_batch(urls, batch_size=batch_size or 8)
  136 + return self.encode_batch(
  137 + urls,
  138 + batch_size=batch_size or 8,
  139 + normalize_embeddings=normalize_embeddings,
  140 + )
... ...
embeddings/protocols.py
  Show/Hide comments   View file @200fddd
... ... @@ -17,6 +17,7 @@ class ImageEncoderProtocol(Protocol):
17 17 self,
18 18 urls: List[str],
19 19 batch_size: Optional[int] = None,
  20 + normalize_embeddings: bool = True,
20 21 ) -> List[Optional[np.ndarray]]:
21 22 """
22 23 Encode a list of image URLs to vectors.
... ...
embeddings/server.py
  Show/Hide comments   View file @200fddd
... ... @@ -112,14 +112,24 @@ def load_models():
112 112 logger.info("All embedding models loaded successfully, service ready")
113 113  
114 114  
115   -def _as_list(embedding: Optional[np.ndarray]) -> Optional[List[float]]:
  115 +def _normalize_vector(vec: np.ndarray) -> np.ndarray:
  116 + norm = float(np.linalg.norm(vec))
  117 + if not np.isfinite(norm) or norm <= 0.0:
  118 + raise RuntimeError("Embedding vector has invalid norm (must be > 0)")
  119 + return vec / norm
  120 +
  121 +
  122 +def _as_list(embedding: Optional[np.ndarray], normalize: bool = False) -> Optional[List[float]]:
116 123 if embedding is None:
117 124 return None
118 125 if not isinstance(embedding, np.ndarray):
119 126 embedding = np.array(embedding, dtype=np.float32)
120 127 if embedding.ndim != 1:
121 128 embedding = embedding.reshape(-1)
122   - return embedding.astype(np.float32).tolist()
  129 + embedding = embedding.astype(np.float32, copy=False)
  130 + if normalize:
  131 + embedding = _normalize_vector(embedding).astype(np.float32, copy=False)
  132 + return embedding.tolist()
123 133  
124 134  
125 135 @app.get("/health")
... ... @@ -134,9 +144,10 @@ def health() -&gt; Dict[str, Any]:
134 144  
135 145  
136 146 @app.post("/embed/text")
137   -def embed_text(texts: List[str]) -> List[Optional[List[float]]]:
  147 +def embed_text(texts: List[str], normalize: Optional[bool] = None) -> List[Optional[List[float]]]:
138 148 if _text_model is None:
139 149 raise RuntimeError("Text model not loaded")
  150 + effective_normalize = bool(CONFIG.TEXT_NORMALIZE_EMBEDDINGS) if normalize is None else bool(normalize)
140 151 normalized: List[str] = []
141 152 for i, t in enumerate(texts):
142 153 if not isinstance(t, str):
... ... @@ -152,7 +163,7 @@ def embed_text(texts: List[str]) -&gt; List[Optional[List[float]]]:
152 163 normalized,
153 164 batch_size=int(CONFIG.TEXT_BATCH_SIZE),
154 165 device=CONFIG.TEXT_DEVICE,
155   - normalize_embeddings=bool(CONFIG.TEXT_NORMALIZE_EMBEDDINGS),
  166 + normalize_embeddings=effective_normalize,
156 167 )
157 168 except Exception as e:
158 169 logger.error("Text embedding backend failure: %s", e, exc_info=True)
... ... @@ -167,7 +178,7 @@ def embed_text(texts: List[str]) -&gt; List[Optional[List[float]]]:
167 178 )
168 179 out: List[Optional[List[float]]] = []
169 180 for i, emb in enumerate(embs):
170   - vec = _as_list(emb)
  181 + vec = _as_list(emb, normalize=effective_normalize)
171 182 if vec is None:
172 183 raise RuntimeError(f"Text model returned empty embedding for index {i}")
173 184 out.append(vec)
... ... @@ -175,9 +186,10 @@ def embed_text(texts: List[str]) -&gt; List[Optional[List[float]]]:
175 186  
176 187  
177 188 @app.post("/embed/image")
178   -def embed_image(images: List[str]) -> List[Optional[List[float]]]:
  189 +def embed_image(images: List[str], normalize: Optional[bool] = None) -> List[Optional[List[float]]]:
179 190 if _image_model is None:
180 191 raise RuntimeError("Image model not loaded")
  192 + effective_normalize = bool(CONFIG.IMAGE_NORMALIZE_EMBEDDINGS) if normalize is None else bool(normalize)
181 193 urls: List[str] = []
182 194 for i, url_or_path in enumerate(images):
183 195 if not isinstance(url_or_path, str):
... ... @@ -188,7 +200,11 @@ def embed_image(images: List[str]) -&gt; List[Optional[List[float]]]:
188 200 urls.append(s)
189 201  
190 202 with _image_encode_lock:
191   - vectors = _image_model.encode_image_urls(urls, batch_size=CONFIG.IMAGE_BATCH_SIZE)
  203 + vectors = _image_model.encode_image_urls(
  204 + urls,
  205 + batch_size=CONFIG.IMAGE_BATCH_SIZE,
  206 + normalize_embeddings=effective_normalize,
  207 + )
192 208 if vectors is None or len(vectors) != len(urls):
193 209 raise RuntimeError(
194 210 f"Image model response length mismatch: expected {len(urls)}, "
... ... @@ -196,7 +212,7 @@ def embed_image(images: List[str]) -&gt; List[Optional[List[float]]]:
196 212 )
197 213 out: List[Optional[List[float]]] = []
198 214 for i, vec in enumerate(vectors):
199   - out_vec = _as_list(vec)
  215 + out_vec = _as_list(vec, normalize=effective_normalize)
200 216 if out_vec is None:
201 217 raise RuntimeError(f"Image model returned empty embedding for index {i}")
202 218 out.append(out_vec)
... ...
embeddings/text_encoder.py
  Show/Hide comments   View file @200fddd
... ... @@ -50,7 +50,7 @@ class TextEmbeddingEncoder:
50 50 logger.warning("Failed to initialize Redis cache for embeddings: %s, continuing without cache", e)
51 51 self.redis_client = None
52 52  
53   - def _call_service(self, request_data: List[str]) -> List[Any]:
  53 + def _call_service(self, request_data: List[str], normalize_embeddings: bool = True) -> List[Any]:
54 54 """
55 55 Call the embedding service API.
56 56  
... ... @@ -63,6 +63,7 @@ class TextEmbeddingEncoder:
63 63 try:
64 64 response = requests.post(
65 65 self.endpoint,
  66 + params={"normalize": "true" if normalize_embeddings else "false"},
66 67 json=request_data,
67 68 timeout=60
68 69 )
... ... @@ -84,7 +85,7 @@ class TextEmbeddingEncoder:
84 85  
85 86 Args:
86 87 sentences: Single string or list of strings to encode
87   - normalize_embeddings: Whether to normalize embeddings (ignored for service)
  88 + normalize_embeddings: Whether to request normalized embeddings from service
88 89 device: Device parameter ignored for service compatibility
89 90 batch_size: Batch size for processing (used for service requests)
90 91  
... ... @@ -103,7 +104,7 @@ class TextEmbeddingEncoder:
103 104 embeddings: List[Optional[np.ndarray]] = [None] * len(sentences)
104 105  
105 106 for i, text in enumerate(sentences):
106   - cached = self._get_cached_embedding(text, "generic")
  107 + cached = self._get_cached_embedding(text, "generic", normalize_embeddings)
107 108 if cached is not None:
108 109 embeddings[i] = cached
109 110 else:
... ... @@ -115,7 +116,7 @@ class TextEmbeddingEncoder:
115 116  
116 117 # If there are uncached texts, call service
117 118 if uncached_texts:
118   - response_data = self._call_service(request_data)
  119 + response_data = self._call_service(request_data, normalize_embeddings=normalize_embeddings)
119 120  
120 121 # Process response
121 122 for i, text in enumerate(uncached_texts):
... ... @@ -129,7 +130,7 @@ class TextEmbeddingEncoder:
129 130 embedding_array = np.array(embedding, dtype=np.float32)
130 131 if self._is_valid_embedding(embedding_array):
131 132 embeddings[original_idx] = embedding_array
132   - self._set_cached_embedding(text, "generic", embedding_array)
  133 + self._set_cached_embedding(text, "generic", embedding_array, normalize_embeddings)
133 134 else:
134 135 raise ValueError(
135 136 f"Invalid embedding returned from service for text index {original_idx}"
... ... @@ -144,7 +145,8 @@ class TextEmbeddingEncoder:
144 145 self,
145 146 texts: List[str],
146 147 batch_size: int = 32,
147   - device: str = 'cpu'
  148 + device: str = 'cpu',
  149 + normalize_embeddings: bool = True,
148 150 ) -> np.ndarray:
149 151 """
150 152 Encode a batch of texts efficiently via network service.
... ... @@ -157,11 +159,17 @@ class TextEmbeddingEncoder:
157 159 Returns:
158 160 numpy array of embeddings
159 161 """
160   - return self.encode(texts, batch_size=batch_size, device=device)
  162 + return self.encode(
  163 + texts,
  164 + batch_size=batch_size,
  165 + device=device,
  166 + normalize_embeddings=normalize_embeddings,
  167 + )
161 168  
162   - def _get_cache_key(self, query: str, language: str) -> str:
  169 + def _get_cache_key(self, query: str, language: str, normalize_embeddings: bool = True) -> str:
163 170 """Generate a cache key for the query"""
164   - return f"embedding:{language}:{query}"
  171 + norm_flag = "norm1" if normalize_embeddings else "norm0"
  172 + return f"embedding:{language}:{norm_flag}:{query}"
165 173  
166 174 def _is_valid_embedding(self, embedding: np.ndarray) -> bool:
167 175 """
... ... @@ -184,13 +192,18 @@ class TextEmbeddingEncoder:
184 192 return False
185 193 return True
186 194  
187   - def _get_cached_embedding(self, query: str, language: str) -> Optional[np.ndarray]:
  195 + def _get_cached_embedding(
  196 + self,
  197 + query: str,
  198 + language: str,
  199 + normalize_embeddings: bool = True,
  200 + ) -> Optional[np.ndarray]:
188 201 """Get embedding from cache if exists (with sliding expiration)"""
189 202 if not self.redis_client:
190 203 return None
191 204  
192 205 try:
193   - cache_key = self._get_cache_key(query, language)
  206 + cache_key = self._get_cache_key(query, language, normalize_embeddings)
194 207 cached_data = self.redis_client.get(cache_key)
195 208 if cached_data:
196 209 embedding = pickle.loads(cached_data)
... ... @@ -216,13 +229,19 @@ class TextEmbeddingEncoder:
216 229 logger.error(f"Error retrieving embedding from cache: {e}")
217 230 return None
218 231  
219   - def _set_cached_embedding(self, query: str, language: str, embedding: np.ndarray) -> bool:
  232 + def _set_cached_embedding(
  233 + self,
  234 + query: str,
  235 + language: str,
  236 + embedding: np.ndarray,
  237 + normalize_embeddings: bool = True,
  238 + ) -> bool:
220 239 """Store embedding in cache"""
221 240 if not self.redis_client:
222 241 return False
223 242  
224 243 try:
225   - cache_key = self._get_cache_key(query, language)
  244 + cache_key = self._get_cache_key(query, language, normalize_embeddings)
226 245 serialized_data = pickle.dumps(embedding)
227 246 self.redis_client.setex(
228 247 cache_key,
... ...
restart.sh
  Show/Hide comments   View file @200fddd
... ... @@ -4,4 +4,7 @@
4 4  
5 5 cd "$(dirname "$0")"
6 6  
7   -./scripts/service_ctl.sh restart
  7 +START_EMBEDDING=1 START_TRANSLATOR=1 START_RERANKER=1 START_TEI=1 CNCLIP_DEVICE=cuda TEI_USE_GPU=1 ./scripts/service_ctl.sh restart
  8 +
  9 +# ./scripts/service_ctl.sh restart
  10 +
... ...
scripts/service_ctl.sh
  Show/Hide comments   View file @200fddd
... ... @@ -190,7 +190,14 @@ start_one() {
190 190 nohup bash -lc "${cmd}" > "${lf}" 2>&1 &
191 191 local pid=$!
192 192 echo "${pid}" > "${pf}"
193   - if wait_for_health "${service}"; then
  193 + # Some services (notably reranker with vLLM backend) can take longer
  194 + # to load models / compile graphs on first start. Give them a longer
  195 + # health check window to avoid false negatives.
  196 + local retries=30
  197 + if [ "${service}" = "reranker" ]; then
  198 + retries=90
  199 + fi
  200 + if wait_for_health "${service}" "${retries}"; then
194 201 echo "[ok] ${service} healthy (pid=${pid}, log=${lf})"
195 202 else
196 203 echo "[error] ${service} health check timeout, inspect ${lf}" >&2
... ... @@ -320,9 +327,14 @@ resolve_targets() {
320 327 if [ "${START_RERANKER:-0}" = "1" ]; then targets+=("reranker"); fi
321 328 echo "${targets[@]}"
322 329 ;;
323   - stop|restart|status)
  330 + stop|status)
324 331 echo "$(all_services)"
325 332 ;;
  333 + restart)
  334 + # Restart with no explicit services should preserve start-order dependency
  335 + # behavior (e.g. tei/cnclip before embedding).
  336 + echo "$(resolve_targets start)"
  337 + ;;
326 338 *)
327 339 echo ""
328 340 ;;
... ... @@ -340,7 +352,7 @@ Usage:
340 352 Default target set (when no service provided):
341 353 start -> backend indexer frontend (+ optional by env flags)
342 354 stop -> all known services
343   - restart -> all known services
  355 + restart -> stop all known services, then start with start targets
344 356 status -> all known services
345 357  
346 358 Optional startup flags:
... ... @@ -361,7 +373,13 @@ main() {
361 373 shift || true
362 374  
363 375 load_env_file "${PROJECT_ROOT}/.env"
  376 + local stop_targets=""
364 377 local targets
  378 + # For restart without explicit services, stop everything first, then start
  379 + # with dependency-aware start targets.
  380 + if [ "${action}" = "restart" ] && [ "$#" -eq 0 ]; then
  381 + stop_targets="$(resolve_targets stop)"
  382 + fi
365 383 targets="$(resolve_targets "${action}" "$@")"
366 384 if [ -z "${targets}" ]; then
367 385 usage
... ... @@ -380,7 +398,7 @@ main() {
380 398 done
381 399 ;;
382 400 restart)
383   - for svc in ${targets}; do
  401 + for svc in ${stop_targets:-${targets}}; do
384 402 stop_one "${svc}"
385 403 done
386 404 for svc in ${targets}; do
... ...