Anthropic transform

Parameter Mappings & Transformations

Basic Parameter Mappings

• model → direct mapping (default: ‘claude-2.1’)
• max_tokens → direct mapping to max_tokens
• temperature → direct mapping (constrained: 0-1)
• top_p → direct mapping (default: -1)
• stream → direct mapping (default: false)
• user → mapped to metadata.user_id
• stop → mapped to stop_sequences
• max_completion_tokens → mapped to max_tokens

Complex Transformations

Messages Transformation

1. System Messages:

   • Extracted from messages array where role === 'system'
   • Transformed into AnthropicMessageContentItem[]
   • Handles both string content and object content with text
   • Preserves cache control metadata if present

2. Assistant Messages (transformAssistantMessage):

   • Transforms content into Anthropic’s content array format
   • Handles tool calls by converting them into Anthropic’s tool_use format
   • Text content is wrapped in {type: 'text', text: content}
   • Tool calls are transformed into {type: 'tool_use', name: function.name, id: toolCall.id, input: parsed_arguments}

3. Tool Messages (transformToolMessage):

   • Converted to user role with tool_result type
   • Preserves tool_call_id as tool_use_id
   • Content wrapped in specific format: {type: 'tool_result', tool_use_id: id, content: string}

4. User Messages with Images:

   • Handles base64 encoded images in content
   • Transforms them into Anthropic’s image format with proper media type
   • Preserves cache control metadata

Tools Transformation

• Converts OpenAI-style function definitions to Anthropic tool format
• Maps function parameters to input_schema
• Preserves cache control metadata
• Structure transformation:
  OpenAI: {function: {name, description, parameters}}
  ↓
  Anthropic: {name, description, input_schema: {type, properties, required}}
  

Tool Choice Transformation

• ‘required’ → {type: 'any'}
• ‘auto’ → {type: 'auto'}
• Function specification → {type: 'tool', name: function.name}

Response Transformations

Regular Response

1. Content Processing:

   • Extracts text content from first content item if type is ‘text’
   • Processes tool_use items into OpenAI tool_calls format
   • Preserves tool IDs and function names

2. Usage Statistics:

   • Maps input_tokens → prompt_tokens
   • Maps output_tokens → completion_tokens
   • Calculates total_tokens
   • Preserves cache-related tokens if present

Streaming Response

1. Event Handling:

   • Filters out ‘ping’ and ‘content_block_stop’ events
   • Converts ‘message_stop’ to ‘[DONE]’
   • Handles multiple event types: content_block_delta, content_block_start, message_delta, message_start

2. Special States:

   • Tracks chain of thought messages
   • Maintains usage statistics across stream
   • Handles tool streaming differently based on message context

Edge Cases & Special Handling

1. Image Content:

   • Special handling for base64 encoded images
   • Parses media type from data URL
   • Validates image URL format

2. Tool Streaming:

   • Handles partial JSON in tool arguments
   • Manages tool indices differently when chain-of-thought messages are present
   • Separates tool name and arguments into different stream chunks

3. Cache Control:

   • Preserves ephemeral cache control metadata throughout transformations
   • Handles cache usage statistics in both regular and streaming responses

4. Error Handling:

   • Transforms Anthropic-specific error format to universal format
   • Preserves error types and messages
   • Handles non-200 response status codes

5. Empty/Null Handling:

   • Safely handles missing usage statistics
   • Manages undefined tool calls
   • Handles empty content arrays