tei

tangwang
1 parent c7e80cc2
Showing 3 changed files with 56 additions and 19 deletions Show diff stats
config/config.yaml
embeddings/server.py
embeddings/tei_model.py
@@ -168,7 +168,7 @@ services:
       http:
         base_url: "http://127.0.0.1:6005"
     # 服务内文本后端（embedding 进程启动时读取）
-    backend: "tei"  # tei | local_st
+    backend: "local_st"  # tei | local_st
     backends:
       tei:
         base_url: "http://127.0.0.1:8080"
@@ -146,13 +146,20 @@ def embed_text(texts: List[str]) -&gt; List[Optional[List[float]]]:
             raise HTTPException(status_code=400, detail=f"Invalid text at index {i}: empty string")
         normalized.append(s)
-    with _text_encode_lock:
-        embs = _text_model.encode_batch(
-            normalized,
-            batch_size=int(CONFIG.TEXT_BATCH_SIZE),
-            device=CONFIG.TEXT_DEVICE,
-            normalize_embeddings=bool(CONFIG.TEXT_NORMALIZE_EMBEDDINGS),
-        )
+    try:
+        with _text_encode_lock:
+            embs = _text_model.encode_batch(
+                normalized,
+                batch_size=int(CONFIG.TEXT_BATCH_SIZE),
+                device=CONFIG.TEXT_DEVICE,
+                normalize_embeddings=bool(CONFIG.TEXT_NORMALIZE_EMBEDDINGS),
+            )
+    except Exception as e:
+        logger.error("Text embedding backend failure: %s", e, exc_info=True)
+        raise HTTPException(
+            status_code=502,
+            detail=f"Text embedding backend failure: {e}",
+        ) from e
     if embs is None or len(embs) != len(normalized):
         raise RuntimeError(
             f"Text model response length mismatch: expected {len(normalized)}, "
@@ -2,7 +2,7 @@
 from __future__ import annotations
-from typing import List, Union
+from typing import Any, List, Union
 import numpy as np
 import requests
@@ -30,6 +30,15 @@ class TEITextModel:
         health_url = f"{self.base_url}/health"
         response = requests.get(health_url, timeout=5)
         response.raise_for_status()
+        # Probe one tiny embedding at startup so runtime requests do not fail later
+        # with opaque "Invalid TEI embedding" errors.
+        probe_resp = requests.post(
+            self.endpoint,
+            json={"inputs": ["health check"]},
+            timeout=min(self.timeout_sec, 15),
+        )
+        probe_resp.raise_for_status()
+        self._parse_payload(probe_resp.json(), expected_len=1)
     @staticmethod
     def _normalize(embedding: np.ndarray) -> np.ndarray:
@@ -77,20 +86,41 @@ class TEITextModel:
         )
         response.raise_for_status()
         payload = response.json()
+        vectors = self._parse_payload(payload, expected_len=len(texts))
+        if normalize_embeddings:
+            vectors = [self._normalize(vec) for vec in vectors]
+        return np.array(vectors, dtype=object)
-        if not isinstance(payload, list) or len(payload) != len(texts):
+    def _parse_payload(self, payload: Any, expected_len: int) -> List[np.ndarray]:
+        if not isinstance(payload, list) or len(payload) != expected_len:
+            got = 0 if payload is None else (len(payload) if isinstance(payload, list) else "non-list")
             raise RuntimeError(
-                f"TEI response length mismatch: expected {len(texts)}, "
-                f"got {0 if payload is None else len(payload)}"
+                f"TEI response length mismatch: expected {expected_len}, got {got}. "
+                f"Response type={type(payload).__name__}"
             )
         vectors: List[np.ndarray] = []
-        for i, emb in enumerate(payload):
-            vec = np.asarray(emb, dtype=np.float32)
-            if vec.ndim != 1 or vec.size == 0 or not np.isfinite(vec).all():
-                raise RuntimeError(f"Invalid TEI embedding at index {i}")
-            if normalize_embeddings:
-                vec = self._normalize(vec)
+        for i, item in enumerate(payload):
+            emb = item.get("embedding") if isinstance(item, dict) else item
+            try:
+                vec = np.asarray(emb, dtype=np.float32)
+            except (TypeError, ValueError) as exc:
+                raise RuntimeError(
+                    f"Invalid TEI embedding at index {i}: cannot convert to float array "
+                    f"(item_type={type(item).__name__})"
+                ) from exc
+
+            if vec.ndim != 1 or vec.size == 0:
+                raise RuntimeError(
+                    f"Invalid TEI embedding at index {i}: shape={vec.shape}, size={vec.size}"
+                )
+            if not np.isfinite(vec).all():
+                preview = vec[:8].tolist()
+                raise RuntimeError(
+                    f"Invalid TEI embedding at index {i}: contains non-finite values, "
+                    f"preview={preview}. This often indicates TEI backend/model runtime issues "
+                    f"(for example an incompatible dtype or model config)."
+                )
             vectors.append(vec)
-        return np.array(vectors, dtype=object)
+        return vectors