Add LLM token

backend/app/infrastructure/vectorstore/cross_encoder_reranker.py

@@ -71,22 +71,48 @@ class OpenAICompatibleReranker(Reranker):
         return result
 
     def _call_reranker(self, query: str, texts: list[str]) -> list[float]:
-        """Call the reranker API and return a score per text."""
+        """Call the reranker API and return a score per text.
+
+        Tries TEI format first (POST /rerank with model+texts), then falls back
+        to Cohere/OpenAI format (POST /v1/rerank with model+documents).
+        Both formats now include the model name, which most gateways require.
+        """
         headers = {"Content-Type": "application/json"}
         if self._api_key:
             headers["Authorization"] = f"Bearer {self._api_key}"
 
-        # Try TEI format first: POST /rerank
-        payload = {"query": query, "texts": texts, "raw_scores": False, "return_text": False}
+        # TEI format: POST /rerank — include model name (required by gateway proxies)
+        payload = {
+            "model": self._model,
+            "query": query,
+            "texts": texts,
+            "raw_scores": False,
+            "return_text": False,
+        }
         url = f"{self._base_url}/rerank"
         resp = requests.post(url, json=payload, headers=headers, timeout=self._timeout)
 
-        if resp.status_code == 404:
-            # Fall back to Cohere / OpenAI-style: POST /v1/rerank
+        if resp.status_code in (404, 400):
+            # Gateway returned an error — try Cohere/OpenAI-style format as fallback.
+            logger.debug(
+                "TEI rerank returned {} — trying Cohere format. Body: {}",
+                resp.status_code,
+                resp.text[:200],
+            )
             payload_v1 = {"model": self._model, "query": query, "documents": texts}
             url = f"{self._base_url}/v1/rerank"
             resp = requests.post(url, json=payload_v1, headers=headers, timeout=self._timeout)
 
+        if not resp.ok:
+            # Surface a clear error message so callers can log it meaningfully.
+            try:
+                err_body = resp.json()
+                err_msg = err_body.get("error", {}).get("message", resp.text[:200])
+            except Exception:
+                err_msg = resp.text[:200]
+            resp.raise_for_status()  # raises HTTPError with status code
+            raise ValueError(err_msg)  # unreachable but satisfies type checker
+
         resp.raise_for_status()
         data = resp.json()