Рабочий вариант

src/rag_agent/index/embeddings.py

@@ -34,6 +34,8 @@ class StubEmbeddingClient:
 
 
 _GIGACHAT_BATCH_SIZE = 50
+# GigaChat embeddings: max 514 tokens per input; Russian/English ~3 chars/token → truncate to stay under
+_GIGACHAT_MAX_CHARS_PER_INPUT = 1200
 
 
 class GigaChatEmbeddingClient:
@@ -57,16 +59,43 @@ class GigaChatEmbeddingClient:
             return []
 
         result: list[list[float]] = []
-        for i in range(0, len(texts_list), _GIGACHAT_BATCH_SIZE):
-            batch = texts_list[i : i + _GIGACHAT_BATCH_SIZE]
-            with GigaChat(
-                credentials=self._credentials,
-                verify_ssl_certs=self._verify_ssl_certs,
-            ) as giga:
-                response = giga.embeddings(model=self._model, input=batch)
-            # Preserve order by index
-            by_index = {item.index: item.embedding for item in response.data}
-            result.extend(by_index[j] for j in range(len(batch)))
+        try:
+            for i in range(0, len(texts_list), _GIGACHAT_BATCH_SIZE):
+                raw_batch = texts_list[i : i + _GIGACHAT_BATCH_SIZE]
+                batch = [
+                    t[: _GIGACHAT_MAX_CHARS_PER_INPUT] if len(t) > _GIGACHAT_MAX_CHARS_PER_INPUT else t
+                    for t in raw_batch
+                ]
+                with GigaChat(
+                    credentials=self._credentials,
+                    model=self._model,
+                    verify_ssl_certs=self._verify_ssl_certs,
+                ) as giga:
+                    # API: embeddings(texts: list[str]) — single positional argument (gigachat 0.2+)
+                    response = giga.embeddings(batch)
+                # Preserve order by index
+                by_index = {item.index: item.embedding for item in response.data}
+                result.extend(by_index[j] for j in range(len(batch)))
+        except Exception as e:
+            from gigachat.exceptions import ResponseError, RequestEntityTooLargeError
+
+            is_402 = (
+                isinstance(e, ResponseError)
+                and (getattr(e, "status_code", None) == 402 or "402" in str(e) or "Payment Required" in str(e))
+            )
+            if is_402:
+                raise ValueError(
+                    "GigaChat: недостаточно средств (402 Payment Required). "
+                    "Пополните баланс в кабинете GigaChat или отключите GIGACHAT_CREDENTIALS для stub-режима."
+                )
+            if isinstance(e, RequestEntityTooLargeError) or (
+                isinstance(e, ResponseError) and getattr(e, "status_code", None) == 413
+            ):
+                raise ValueError(
+                    "GigaChat: превышен лимит токенов на один запрос (413). "
+                    "Уменьшите RAG_CHUNK_SIZE_LINES или RAG_CHUNK_SIZE в .env (текущий лимит ~514 токенов на чанк)."
+                )
+            raise
         return result