Response streaming

The Problem with Standard Requests

• User messages sent to Claude can take 10-30 seconds to receive a response
• Showing only a spinner during this wait creates poor user experience
• Users expect to see response text appear almost immediately

How Streaming Works

• Server sends initial user message to Claude
• Claude immediately sends back an initial response (no text content yet) to signal it's starting generation
• Server then receives a stream of events containing chunks of the generated response
• Each event contains a small piece of the overall message, not necessarily just one word
• Server can extract text from each event and send it to the client immediately
• Result: text appears chunk-by-chunk on screen as it's generated

Event Types in Streaming

• Multiple event types are sent back: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop
• content_block_delta is the most important event type—it contains the actual generated text
• Standard sequence: message_start → content_block_start → multiple content_block_deltas → content_block_stop → message_delta → message_stop

Using the Anthropic SDK for Streaming

• Use client. messages. stream() wrapped in a with block instead of client. messages. create() with stream=true
• Access text chunks via stream. text_stream() to get only the text content without processing raw events
• This simplifies code by automatically extracting text from content_block_delta events

Collecting Full Messages

• After streaming completes, use stream. get_final_message() to assemble all chunks into one complete message
• Allows storing the entire conversation in a database while still streaming chunks to the user in real-time

Takeaways

• Streaming sends response text chunk-by-chunk to the client as Claude generates it, eliminating long wait times and improving user experience
• The content_block_delta event type contains the actual generated text chunks that should be displayed to users
• Use client. messages. stream() with stream. text_stream() in the Anthropic SDK to easily extract and display text chunks without handling raw events
• After streaming completes, call stream. get_final_message() to get the complete assembled message for storage or further processing