feat: add prompt caching support for Claude and Nova models

Add comprehensive prompt caching support with flexible control options:

Features:
- ENV variable control (ENABLE_PROMPT_CACHING, default: false)
- Per-request control via extra_body.prompt_caching
- Pattern-based model detection (Claude, Nova)
- Token limit warnings (Nova 20K limit)
- OpenAI-compatible response format (prompt_tokens_details.cached_tokens)

Supported models:
- Claude 3+ models (anthropic.claude-*)
- Nova models (amazon.nova-*)
- Auto-detection prevents breaking unsupported models

Implementation:
- System prompts caching via extra_body.prompt_caching.system
- Messages caching via extra_body.prompt_caching.messages
- Non-streaming and streaming modes
- Compatible with reasoning, thinking, and tool calls

README.md

@@ -29,6 +29,7 @@ If you find this GitHub repository useful, please consider giving it a free star
 - [x] Support Application Inference Profiles (**new**)
 - [x] Support Reasoning (**new**)
 - [x] Support Interleaved thinking (**new**)
+- [x] Support Prompt Caching (**new**)
 
 Please check [Usage Guide](./docs/Usage.md) for more details about how to use the new APIs.
 
@@ -221,6 +222,78 @@ print(completion.choices[0].message.content)
 
 For more information about creating and managing application inference profiles, see the [Amazon Bedrock User Guide](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html).
 
+### Prompt Caching
+
+This proxy now supports **Prompt Caching** for Claude and Nova models, which can reduce costs by up to 90% and latency by up to 85% for workloads with repeated prompts.
+
+**Supported Models:**
+- Claude 3+ models (Claude 3.5 Haiku, Claude 3.7 Sonnet, Claude 4, Claude 4.5, etc.)
+- Nova models (Nova Micro, Nova Lite, Nova Pro, Nova Premier)
+
+**Enabling Prompt Caching:**
+
+You can enable prompt caching in two ways:
+
+1. **Globally via Environment Variable** (set in ECS Task Definition or Lambda):
+```bash
+ENABLE_PROMPT_CACHING=true
+```
+
+2. **Per-request via `extra_body`** :
+
+**Python SDK:**
+```python
+from openai import OpenAI
+
+client = OpenAI()
+
+# Cache system prompts
+response = client.chat.completions.create(
+    model="us.anthropic.claude-3-7-sonnet-20250219-v1:0",
+    messages=[
+        {"role": "system", "content": "You are an expert assistant with knowledge of..."},
+        {"role": "user", "content": "Help me with this task"}
+    ],
+    extra_body={
+        "prompt_caching": {"system": True}
+    }
+)
+
+# Check cache hit
+if response.usage.prompt_tokens_details:
+    cached_tokens = response.usage.prompt_tokens_details.cached_tokens
+    print(f"Cached tokens: {cached_tokens}")
+```
+
+**cURL:**
+```bash
+curl $OPENAI_BASE_URL/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $OPENAI_API_KEY" \
+  -d '{
+    "model": "us.anthropic.claude-3-7-sonnet-20250219-v1:0",
+    "messages": [
+      {"role": "system", "content": "Long system prompt..."},
+      {"role": "user", "content": "Question"}
+    ],
+    "extra_body": {
+      "prompt_caching": {"system": true}
+    }
+  }'
+```
+
+**Cache Options:**
+- `"prompt_caching": {"system": true}` - Cache system prompts
+- `"prompt_caching": {"messages": true}` - Cache user messages
+- `"prompt_caching": {"system": true, "messages": true}` - Cache both
+
+**Requirements:**
+- Prompt must be ≥1,024 tokens to enable caching
+- Cache TTL is 5 minutes (resets on each cache hit)
+- Nova models have a 20,000 token caching limit
+
+For more information, see the [Amazon Bedrock Prompt Caching Guide](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html).
+
 ## Other Examples
 
 ### LangChain

deployment/BedrockProxy.template

@@ -11,6 +11,13 @@ Parameters:
     Type: String
     Default: anthropic.claude-3-sonnet-20240229-v1:0
     Description: The default model ID, please make sure the model ID is supported in the current region
+  EnablePromptCaching:
+    Type: String
+    Default: "false"
+    AllowedValues:
+      - "true"
+      - "false"
+    Description: Enable prompt caching for supported models (Claude, Nova). When enabled, adds cachePoint to system prompts and messages for cost savings.
 Resources:
   VPCB9E5F0B4:
     Type: AWS::EC2::VPC
@@ -184,6 +191,8 @@ Resources:
           DEFAULT_EMBEDDING_MODEL: cohere.embed-multilingual-v3
           ENABLE_CROSS_REGION_INFERENCE: "true"
           ENABLE_APPLICATION_INFERENCE_PROFILES: "true"
+          ENABLE_PROMPT_CACHING:
+            Ref: EnablePromptCaching
       MemorySize: 1024
       PackageType: Image
       Role:

deployment/BedrockProxyFargate.template

@@ -11,6 +11,13 @@ Parameters:
     Type: String
     Default: anthropic.claude-3-sonnet-20240229-v1:0
     Description: The default model ID, please make sure the model ID is supported in the current region
+  EnablePromptCaching:
+    Type: String
+    Default: "false"
+    AllowedValues:
+      - "true"
+      - "false"
+    Description: Enable prompt caching for supported models (Claude, Nova). When enabled, adds cachePoint to system prompts and messages for cost savings.
 Resources:
   VPCB9E5F0B4:
     Type: AWS::EC2::VPC
@@ -251,6 +258,9 @@ Resources:
               Value: "true"
             - Name: ENABLE_APPLICATION_INFERENCE_PROFILES
               Value: "true"
+            - Name: ENABLE_PROMPT_CACHING
+              Value:
+                Ref: EnablePromptCaching
           Essential: true
           Image:
             Ref: ContainerImageUri

src/api/models/bedrock.py

@@ -24,6 +24,7 @@ from api.schema import (
     ChatStreamResponse,
     Choice,
     ChoiceDelta,
+    CompletionTokensDetails,
     Embedding,
     EmbeddingsRequest,
     EmbeddingsResponse,
@@ -32,6 +33,7 @@ from api.schema import (
     ErrorMessage,
     Function,
     ImageContent,
+    PromptTokensDetails,
     ResponseFunction,
     TextContent,
     ToolCall,
@@ -46,6 +48,7 @@ from api.setting import (
     DEFAULT_MODEL,
     ENABLE_CROSS_REGION_INFERENCE,
     ENABLE_APPLICATION_INFERENCE_PROFILES,
+    ENABLE_PROMPT_CACHING,
 )
 
 logger = logging.getLogger(__name__)
@@ -203,6 +206,64 @@ class BedrockModel(BaseChatModel):
                 detail=error,
             )
 
+    @staticmethod
+    def _supports_prompt_caching(model_id: str) -> bool:
+        """
+        Check if model supports prompt caching based on model ID pattern.
+
+        Uses pattern matching instead of hardcoded whitelist for better maintainability.
+        Automatically supports new models following the naming convention.
+
+        Supported models:
+        - Claude: anthropic.claude-* (excluding very old versions)
+        - Nova: amazon.nova-*
+
+        Returns:
+            bool: True if model supports prompt caching
+        """
+        model_lower = model_id.lower()
+
+        # Claude models pattern matching
+        if "anthropic.claude" in model_lower or ".anthropic.claude" in model_lower:
+            # Exclude very old models that don't support caching
+            excluded_patterns = ["claude-instant", "claude-v1", "claude-v2"]
+            if any(pattern in model_lower for pattern in excluded_patterns):
+                return False
+            return True
+
+        # Nova models pattern matching
+        if "amazon.nova" in model_lower or ".amazon.nova" in model_lower:
+            return True
+
+        # Future providers can be added here
+        # Example: if "provider.model-name" in model_lower: return True
+
+        return False
+
+    @staticmethod
+    def _get_max_cache_tokens(model_id: str) -> int | None:
+        """
+        Get maximum cacheable tokens limit for the model.
+
+        Different models have different caching limits:
+        - Claude: No explicit limit mentioned in docs
+        - Nova: 20,000 tokens max
+
+        Returns:
+            int | None: Max tokens, or None if unlimited
+        """
+        model_lower = model_id.lower()
+
+        # Nova models have 20K limit
+        if "amazon.nova" in model_lower or ".amazon.nova" in model_lower:
+            return 20_000
+
+        # Claude: No explicit limit
+        if "anthropic.claude" in model_lower or ".anthropic.claude" in model_lower:
+            return None
+
+        return None
+
     async def _invoke_bedrock(self, chat_request: ChatRequest, stream=False):
         """Common logic for invoke bedrock models"""
         if DEBUG:
@@ -240,17 +301,32 @@ class BedrockModel(BaseChatModel):
         response = await self._invoke_bedrock(chat_request)
 
         output_message = response["output"]["message"]
-        input_tokens = response["usage"]["inputTokens"]
+        usage = response["usage"]
-        output_tokens = response["usage"]["outputTokens"]
+
+        # Extract all token counts
+        output_tokens = usage["outputTokens"]
+        total_tokens = usage["totalTokens"]
         finish_reason = response["stopReason"]
 
+        # Extract prompt caching metrics if available
+        cache_read_tokens = usage.get("cacheReadInputTokens", 0)
+        cache_creation_tokens = usage.get("cacheCreationInputTokens", 0)
+
+        # Calculate actual prompt tokens
+        # Bedrock's totalTokens includes all: inputTokens + cacheRead + cacheWrite + outputTokens
+        # So: prompt_tokens = totalTokens - outputTokens
+        actual_prompt_tokens = total_tokens - output_tokens
+
         chat_response = self._create_response(
             model=chat_request.model,
             message_id=message_id,
             content=output_message["content"],
             finish_reason=finish_reason,
-            input_tokens=input_tokens,
+            input_tokens=actual_prompt_tokens,
             output_tokens=output_tokens,
+            total_tokens=total_tokens,
+            cache_read_tokens=cache_read_tokens,
+            cache_creation_tokens=cache_creation_tokens,
         )
         if DEBUG:
             logger.info("Proxy response :" + chat_response.model_dump_json())
@@ -296,24 +372,68 @@ class BedrockModel(BaseChatModel):
             yield self.stream_response_to_bytes(error_event)
 
     def _parse_system_prompts(self, chat_request: ChatRequest) -> list[dict[str, str]]:
-        """Create system prompts.
+        """Create system prompts with optional prompt caching support.
-        Note that not all models support system prompts.
 
-        example output: [{"text" : system_prompt}]
+        Prompt caching can be enabled via:
+        1. ENABLE_PROMPT_CACHING environment variable (global default)
+        2. extra_body.prompt_caching.system = True/False (per-request override)
 
-        See example:
+        Only adds cachePoint if:
-        https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html#message-inference-examples
+        - Model supports caching (Claude, Nova)
+        - Caching is enabled (ENV or extra_body)
+        - System prompts exist and meet minimum token requirements
+
+        Example output: [{"text" : system_prompt}, {"cachePoint": {"type": "default"}}]
+
+        See: https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
         """
-
         system_prompts = []
         for message in chat_request.messages:
             if message.role != "system":
-                # ignore system messages here
                 continue
             if not isinstance(message.content, str):
                 raise TypeError(f"System message content must be a string, got {type(message.content).__name__}")
             system_prompts.append({"text": message.content})
 
+        if not system_prompts:
+            return system_prompts
+
+        # Check if model supports prompt caching
+        if not self._supports_prompt_caching(chat_request.model):
+            return system_prompts
+
+        # Determine if caching should be enabled
+        cache_enabled = ENABLE_PROMPT_CACHING  # Default from ENV
+
+        # Check for extra_body override
+        if chat_request.extra_body and isinstance(chat_request.extra_body, dict):
+            prompt_caching = chat_request.extra_body.get("prompt_caching", {})
+            if "system" in prompt_caching:
+                # extra_body explicitly controls caching
+                cache_enabled = prompt_caching.get("system") is True
+
+        if not cache_enabled:
+            return system_prompts
+
+        # Estimate total tokens for limit check
+        total_text = " ".join(p.get("text", "") for p in system_prompts)
+        estimated_tokens = len(total_text.split()) * 1.3  # Rough estimate
+
+        # Check token limits (Nova has 20K limit)
+        max_tokens = self._get_max_cache_tokens(chat_request.model)
+        if max_tokens and estimated_tokens > max_tokens:
+            logger.warning(
+                f"System prompts (~{estimated_tokens:.0f} tokens) exceed model cache limit ({max_tokens} tokens). "
+                f"Caching will still be attempted but may not work optimally."
+            )
+            # Still add cachePoint - let Bedrock handle the limit
+
+        # Add cache checkpoint after system prompts
+        system_prompts.append({"cachePoint": {"type": "default"}})
+
+        if DEBUG:
+            logger.info(f"Added cachePoint to system prompts for model {chat_request.model}")
+
         return system_prompts
 
     def _parse_messages(self, chat_request: ChatRequest) -> list[dict]:
@@ -402,7 +522,7 @@ class BedrockModel(BaseChatModel):
             else:
                 # ignore others, such as system messages
                 continue
-        return self._reframe_multi_payloard(messages)
+        return self._reframe_multi_payloard(messages, chat_request)
 
     def _extract_tool_content(self, content) -> str:
         """Extract text content from various OpenAI SDK tool message formats.
@@ -455,7 +575,7 @@ class BedrockModel(BaseChatModel):
             # Return a safe fallback
             return str(content) if content is not None else ""
 
-    def _reframe_multi_payloard(self, messages: list) -> list:
+    def _reframe_multi_payloard(self, messages: list, chat_request: ChatRequest = None) -> list:
         """Receive messages and reformat them to comply with the Claude format
 
         With OpenAI format requests, it's not a problem to repeatedly receive messages from the same role, but
@@ -510,6 +630,29 @@ class BedrockModel(BaseChatModel):
                 {"role": current_role, "content": current_content}
             )
 
+        # Add cachePoint to messages if enabled and supported
+        if chat_request and reformatted_messages:
+            if not self._supports_prompt_caching(chat_request.model):
+                return reformatted_messages
+
+            # Determine if messages caching should be enabled
+            cache_enabled = ENABLE_PROMPT_CACHING
+
+            if chat_request.extra_body and isinstance(chat_request.extra_body, dict):
+                prompt_caching = chat_request.extra_body.get("prompt_caching", {})
+                if "messages" in prompt_caching:
+                    cache_enabled = prompt_caching.get("messages") is True
+
+            if cache_enabled:
+                # Add cachePoint to the last user message content
+                for msg in reversed(reformatted_messages):
+                    if msg["role"] == "user" and msg.get("content"):
+                        # Add cachePoint at the end of user message content
+                        msg["content"].append({"cachePoint": {"type": "default"}})
+                        if DEBUG:
+                            logger.info(f"Added cachePoint to last user message for model {chat_request.model}")
+                        break
+
         return reformatted_messages
 
     def _parse_request(self, chat_request: ChatRequest) -> dict:
@@ -547,24 +690,39 @@ class BedrockModel(BaseChatModel):
             "inferenceConfig": inference_config,
         }
         if chat_request.reasoning_effort:
-            # From OpenAI api, the max_token is not supported in reasoning mode
+            # reasoning_effort is supported by Claude and DeepSeek v3
-            # Use max_completion_tokens if provided.
+            # Different models use different formats
+            model_lower = chat_request.model.lower()
 
-            max_tokens = (
+            if "anthropic.claude" in model_lower or ".anthropic.claude" in model_lower:
-                chat_request.max_completion_tokens
+                # Claude format: reasoning_config = object with budget_tokens
-                if chat_request.max_completion_tokens
+                max_tokens = (
-                else chat_request.max_tokens
+                    chat_request.max_completion_tokens
-            )
+                    if chat_request.max_completion_tokens
-            budget_tokens = self._calc_budget_tokens(
+                    else chat_request.max_tokens
-                max_tokens, chat_request.reasoning_effort
+                )
-            )
+                budget_tokens = self._calc_budget_tokens(
-            inference_config["maxTokens"] = max_tokens
+                    max_tokens, chat_request.reasoning_effort
-            # unset topP - Not supported
+                )
-            inference_config.pop("topP", None)
+                inference_config["maxTokens"] = max_tokens
+                # unset topP - Not supported
+                inference_config.pop("topP", None)
 
-            args["additionalModelRequestFields"] = {
+                args["additionalModelRequestFields"] = {
-                "reasoning_config": {"type": "enabled", "budget_tokens": budget_tokens}
+                    "reasoning_config": {"type": "enabled", "budget_tokens": budget_tokens}
-            }
+                }
+            elif "deepseek.v3" in model_lower or "deepseek.deepseek-v3" in model_lower:
+                # DeepSeek v3 format: reasoning_config = string ('low', 'medium', 'high')
+                # From Bedrock Playground: {"reasoning_config": "high"}
+                args["additionalModelRequestFields"] = {
+                    "reasoning_config": chat_request.reasoning_effort  # Direct string: low/medium/high
+                }
+                if DEBUG:
+                    logger.info(f"Applied reasoning_config={chat_request.reasoning_effort} for DeepSeek v3")
+            else:
+                # For other models (Qwen, etc.), ignore reasoning_effort parameter
+                if DEBUG:
+                    logger.info(f"reasoning_effort parameter ignored for model {chat_request.model} (not supported)")
         # add tool config
         if chat_request.tools:
             tool_config = {"tools": [self._convert_tool_spec(t.function) for t in chat_request.tools]}
@@ -585,16 +743,42 @@ class BedrockModel(BaseChatModel):
                         raise ValueError("tool_choice must contain 'function' key when specifying a specific tool")
                     tool_config["toolChoice"] = {"tool": {"name": chat_request.tool_choice["function"].get("name", "")}}
             args["toolConfig"] = tool_config
-        # add Additional fields to enable extend thinking
+        # Add additional fields to enable extend thinking or other model-specific features
         if chat_request.extra_body:
-            # reasoning_config will not be used
+            # Filter out prompt_caching (our control field, not for Bedrock)
-            args["additionalModelRequestFields"] = chat_request.extra_body
+            additional_fields = {
-            # Extended thinking doesn't support both temperature and topP
+                k: v for k, v in chat_request.extra_body.items()
-            # Remove topP to avoid validation error
+                if k != "prompt_caching"
-            if "thinking" in chat_request.extra_body:
+            }
-                inference_config.pop("topP", None)
+
+            if additional_fields:
+                # Only set additionalModelRequestFields if there are actual fields to pass
+                args["additionalModelRequestFields"] = additional_fields
+
+                # Extended thinking doesn't support both temperature and topP
+                # Remove topP to avoid validation error
+                if "thinking" in additional_fields:
+                    inference_config.pop("topP", None)
+
         return args
 
+    def _estimate_reasoning_tokens(self, content: list[dict]) -> int:
+        """
+        Estimate reasoning tokens from reasoningContent blocks.
+
+        Bedrock doesn't separately report reasoning tokens, so we estimate
+        them using tiktoken to maintain OpenAI API compatibility.
+        """
+        reasoning_text = ""
+        for block in content:
+            if "reasoningContent" in block:
+                reasoning_text += block["reasoningContent"]["reasoningText"].get("text", "")
+
+        if reasoning_text:
+            # Use tiktoken to estimate token count
+            return len(ENCODER.encode(reasoning_text))
+        return 0
+
     def _create_response(
         self,
         model: str,
@@ -603,6 +787,9 @@ class BedrockModel(BaseChatModel):
         finish_reason: str | None = None,
         input_tokens: int = 0,
         output_tokens: int = 0,
+        total_tokens: int = 0,
+        cache_read_tokens: int = 0,
+        cache_creation_tokens: int = 0,
     ) -> ChatResponse:
         message = ChatResponseMessage(
             role="assistant",
@@ -642,6 +829,25 @@ class BedrockModel(BaseChatModel):
                 message.content = f"<think>{message.reasoning_content}</think>{message.content}"
                 message.reasoning_content = None
 
+        # Create prompt_tokens_details if cache metrics are available
+        prompt_tokens_details = None
+        if cache_read_tokens > 0 or cache_creation_tokens > 0:
+            # Map Bedrock cache metrics to OpenAI format
+            # cached_tokens represents tokens read from cache (cache hits)
+            prompt_tokens_details = PromptTokensDetails(
+                cached_tokens=cache_read_tokens,
+                audio_tokens=0,
+            )
+
+        # Create completion_tokens_details if reasoning content exists
+        completion_tokens_details = None
+        reasoning_tokens = self._estimate_reasoning_tokens(content) if content else 0
+        if reasoning_tokens > 0:
+            completion_tokens_details = CompletionTokensDetails(
+                reasoning_tokens=reasoning_tokens,
+                audio_tokens=0,
+            )
+
         response = ChatResponse(
             id=message_id,
             model=model,
@@ -656,7 +862,9 @@ class BedrockModel(BaseChatModel):
             usage=Usage(
                 prompt_tokens=input_tokens,
                 completion_tokens=output_tokens,
-                total_tokens=input_tokens + output_tokens,
+                total_tokens=total_tokens if total_tokens > 0 else input_tokens + output_tokens,
+                prompt_tokens_details=prompt_tokens_details,
+                completion_tokens_details=completion_tokens_details,
             ),
         )
         response.system_fingerprint = "fp"
@@ -744,14 +952,36 @@ class BedrockModel(BaseChatModel):
             metadata = chunk["metadata"]
             if "usage" in metadata:
                 # token usage
+                usage_data = metadata["usage"]
+
+                # Extract prompt caching metrics if available
+                cache_read_tokens = usage_data.get("cacheReadInputTokens", 0)
+                cache_creation_tokens = usage_data.get("cacheCreationInputTokens", 0)
+
+                # Create prompt_tokens_details if cache metrics are available
+                prompt_tokens_details = None
+                if cache_read_tokens > 0 or cache_creation_tokens > 0:
+                    prompt_tokens_details = PromptTokensDetails(
+                        cached_tokens=cache_read_tokens,
+                        audio_tokens=0,
+                    )
+
+                # Calculate actual prompt tokens
+                # Bedrock's totalTokens includes all tokens
+                # prompt_tokens = totalTokens - outputTokens
+                total_tokens = usage_data["totalTokens"]
+                output_tokens = usage_data["outputTokens"]
+                actual_prompt_tokens = total_tokens - output_tokens
+
                 return ChatStreamResponse(
                     id=message_id,
                     model=model_id,
                     choices=[],
                     usage=Usage(
-                        prompt_tokens=metadata["usage"]["inputTokens"],
+                        prompt_tokens=actual_prompt_tokens,
-                        completion_tokens=metadata["usage"]["outputTokens"],
+                        completion_tokens=output_tokens,
-                        total_tokens=metadata["usage"]["totalTokens"],
+                        total_tokens=total_tokens,
+                        prompt_tokens_details=prompt_tokens_details,
                     ),
                 )
         if message:

src/api/schema.py

@@ -110,10 +110,24 @@ class ChatRequest(BaseModel):
     extra_body: dict | None = None
 
 
+class PromptTokensDetails(BaseModel):
+    """Details about prompt tokens usage, following OpenAI API format."""
+    cached_tokens: int = 0
+    audio_tokens: int = 0
+
+
+class CompletionTokensDetails(BaseModel):
+    """Details about completion tokens usage, following OpenAI API format."""
+    reasoning_tokens: int = 0
+    audio_tokens: int = 0
+
+
 class Usage(BaseModel):
     prompt_tokens: int
     completion_tokens: int
     total_tokens: int
+    prompt_tokens_details: PromptTokensDetails | None = None
+    completion_tokens_details: CompletionTokensDetails | None = None
 
 
 class ChatResponseMessage(BaseModel):

src/api/setting.py

@@ -15,3 +15,4 @@ DEFAULT_MODEL = os.environ.get("DEFAULT_MODEL", "anthropic.claude-3-sonnet-20240
 DEFAULT_EMBEDDING_MODEL = os.environ.get("DEFAULT_EMBEDDING_MODEL", "cohere.embed-multilingual-v3")
 ENABLE_CROSS_REGION_INFERENCE = os.environ.get("ENABLE_CROSS_REGION_INFERENCE", "true").lower() != "false"
 ENABLE_APPLICATION_INFERENCE_PROFILES = os.environ.get("ENABLE_APPLICATION_INFERENCE_PROFILES", "true").lower() != "false"
+ENABLE_PROMPT_CACHING = os.environ.get("ENABLE_PROMPT_CACHING", "false").lower() != "false"