fix: improve context window error detection across providers

Add additional error message patterns to is_error_str_context_window_exceeded()
so that context limit errors from Mistral, Perplexity, Bedrock, Replicate, and
Anthropic are correctly mapped to ContextWindowExceededError instead of raising
a generic BadRequestError.

New patterns:
- "too large for model" (Mistral)
- "length of all messages is too long" (Perplexity)
- "too many tokens" / "too many input tokens" (Bedrock)
- "input is too long" (Replicate, Bedrock)
- "prompt is too long" (Anthropic, Bedrock)

Fixes #21558

litellm/litellm_core_utils/exception_mapping_utils.py

@@ -84,6 +84,12 @@ class ExceptionCheckers:
             "input tokens exceed the configured limit",
             "`inputs` tokens + `max_new_tokens` must be",
             "exceeds the maximum number of tokens allowed",  # Gemini
+            "too large for model",  # Mistral
+            "length of all messages is too long",  # Perplexity
+            "too many tokens",  # Bedrock
+            "too many input tokens",  # Bedrock
+            "input is too long",  # Replicate, Bedrock
+            "prompt is too long",  # Anthropic, Bedrock
         ]
         for substring in known_exception_substrings:
             if substring in _error_str_lowercase:

tests/test_litellm/litellm_core_utils/test_exception_mapping_utils.py

@@ -71,6 +71,34 @@ context_window_test_cases = [
         "CerebrasException - Please reduce the length of the messages or completion. Current length is 50000 while limit is 40000",
         True,
     ),
+    # Mistral context window error (issue #21558)
+    (
+        '{"object":"error","message":"Prompt contains 1487700 tokens and 0 draft tokens, too large for model with 131072 maximum context length","type":"invalid_request_invalid_args","param":null,"code":"3051"}',
+        True,
+    ),
+    # Perplexity context window error (issue #21558)
+    (
+        "The total length of all messages is too long.",
+        True,
+    ),
+    # Bedrock context window errors
+    (
+        "BedrockException: Context Window Error - too many tokens",
+        True,
+    ),
+    (
+        "Too many input tokens. Max input tokens: 200000, request input token count: 250000",
+        True,
+    ),
+    (
+        "Input is too long. Max input length is 100000 tokens.",
+        True,
+    ),
+    # Anthropic/Bedrock prompt too long
+    (
+        "prompt is too long: 150000 tokens > 100000 maximum",
+        True,
+    ),
     # Negative cases (should return False)
     ("A generic API error occurred.", False),
     ("Invalid API Key provided.", False),