project_6/upstream_ref/xllm/xllm/proto/anthropic.proto

syntax = "proto3";

option go_package = "jd.com/jd-infer/xllm;xllm";
package xllm.proto;

import "google/protobuf/struct.proto";

// Anthropic API Protocol Definitions
// Reference: https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/anthropic/protocol.py

message AnthropicError {
  string type = 1;

  // Human-readable error message
  string message = 2;
}

message AnthropicErrorResponse {
  string type = 1;

  // The error details
  AnthropicError error = 2;
}

message AnthropicUsage {
  // Number of input tokens
  int32 input_tokens = 1;

  // Number of output tokens
  int32 output_tokens = 2;

  // Number of tokens used for cache creation (optional)
  optional int32 cache_creation_input_tokens = 3;

  // Number of tokens read from cache (optional)
  optional int32 cache_read_input_tokens = 4;
}

message AnthropicStructList {
  repeated google.protobuf.Struct items = 1;
}

message AnthropicContentBlock {
  // Type of content block: "text", "image", "tool_use", "tool_result"
  string type = 1;

  // Text content (for text type)
  optional string text = 2;

  // Image source (for image type) - dict[str, Any] in Python
  optional google.protobuf.Struct source = 3;

  // Tool use/result ID
  optional string id = 4;

  // Tool name (for tool_use type)
  optional string name = 5;

  // Tool input (for tool_use type) - JSON object
  optional google.protobuf.Struct input = 6;

  // Tool result content - can be string or complex structure
  // Using oneof to handle string or structured content
  oneof tool_result_content {
    string content_string = 7;
    AnthropicStructList content_list = 8;
  }

  // Whether tool result is an error (for tool_result type)
  optional bool is_error = 9;
}

message AnthropicContentBlockList {
  repeated AnthropicContentBlock blocks = 1;
}

message AnthropicMessage {
  // Role of the message author: "user" or "assistant"
  string role = 1;

  // Content can be a simple string or a list of content blocks
  // For simple string content, use content_string
  oneof message_content {
    string content_string = 2;
    AnthropicContentBlockList content_blocks = 3;
  }
}

message AnthropicTool {
  // Name of the tool
  string name = 1;

  // Description of the tool (optional)
  optional string description = 2;

  // JSON Schema for the tool's input parameters
  google.protobuf.Struct input_schema = 3;
}

message AnthropicToolChoice {
  // Type of tool choice: "auto", "any", or "tool"
  string type = 1;

  // Name of specific tool (required when type is "tool")
  optional string name = 2;
}

message AnthropicMessagesRequest {
  // ID of the model to use
  string model = 1;

  // A list of messages comprising the conversation so far
  repeated AnthropicMessage messages = 2;

  // The maximum number of tokens to generate
  int32 max_tokens = 3;

  // Metadata about the request (optional)
  optional google.protobuf.Struct metadata = 4;

  // Custom stop sequences (optional)
  repeated string stop_sequences = 5;

  // Whether to stream the response (optional, default: false)
  optional bool stream = 6;

  // System prompt - can be string or list of content blocks
  oneof system_prompt {
    string system_string = 7;
    AnthropicContentBlockList system_blocks = 8;
  }

  // Sampling temperature (optional)
  optional float temperature = 9;

  // Tool choice configuration (optional)
  optional AnthropicToolChoice tool_choice = 10;

  // List of tools available to the model (optional)
  repeated AnthropicTool tools = 11;

  // Top-K sampling (optional)
  optional int32 top_k = 12;

  // Top-P (nucleus) sampling (optional)
  optional float top_p = 13;
}

message AnthropicDelta {
  // Type of delta: "text_delta" or "input_json_delta"
  optional string type = 1;

  // Text content (for text_delta)
  optional string text = 2;

  // Partial JSON (for input_json_delta, used in tool calls)
  optional string partial_json = 3;

  // Stop reason (for message_delta)
  optional string stop_reason = 4;

  // Stop sequence that triggered the stop
  optional string stop_sequence = 5;
}

message AnthropicStreamEvent {
  // Type of stream event
  string type = 1;

  // Full message (for message_start)
  optional AnthropicMessagesResponse message = 2;

  // Delta content (for message_delta, content_block_delta)
  optional AnthropicDelta delta = 3;

  // Content block (for content_block_start)
  optional AnthropicContentBlock content_block = 4;

  // Index of the content block
  optional int32 index = 5;

  // Error details (for error events)
  optional AnthropicError error = 6;

  // Usage information (for message_delta with usage)
  optional AnthropicUsage usage = 7;
}

message AnthropicMessagesResponse {
  // Unique message ID
  string id = 1;

  // Object type, always "message"
  string type = 2;

  // Role, always "assistant"
  string role = 3;

  // List of content blocks in the response
  repeated AnthropicContentBlock content = 4;

  // The model that generated the response
  string model = 5;

  // Reason the model stopped generating
  optional string stop_reason = 6;

  // The stop sequence that triggered the stop (if applicable)
  optional string stop_sequence = 7;

  // Token usage statistics
  optional AnthropicUsage usage = 8;
}