cohere-streaming_skill

Overview

This skill documents a practical reference for Cohere streaming across real-time text generation, tool use events, and retrieval-augmented generation (RAG) citations. It consolidates event type handling, async streaming patterns, full-response collection, tool call flows, and error handling into concise examples. Use it to implement production-ready streaming chat UIs and agent tool integrations with Cohere APIs.

How this skill works

The skill demonstrates iterating over co.chat_stream (sync and async) to receive discrete stream events such as content-delta, message-start, and message-end. It shows how to assemble partial content chunks into a complete response, surface tool planning and tool-call events, and capture citation-generation events when providing RAG sources. Examples include safe streaming wrappers with API error handling and simple hooks for token-level UI updates.

When to use it

• Building a real-time chat interface that renders tokens as they arrive.
• Implementing agents that call external tools and must stream planning and tool-call activity.
• Capturing RAG citations in-stream and displaying source metadata alongside text.
• Collecting the complete generated response while showing partial output to users.
• Handling streaming in asynchronous Python applications or event loops.

Best practices

• Listen for all event types (message-start, content-delta, content-end, message-end) to maintain correct UI state and final response capture.
• Accumulate content-delta chunks in a list and join them at message-end to avoid lost tokens or ordering issues.
• Expose token-level callbacks (on_token) for low-latency UI updates and an on_complete callback for post-processing.
• Handle tool-plan-delta, tool-call-start/delta/end to show plan and arguments, and to execute tool calls deterministically.
• Wrap streaming in try/except to catch ApiError and unexpected exceptions and to provide graceful user feedback.

Example use cases

• A chat app that streams model responses token-by-token and shows a typing indicator during message-start.
• An agent that streams its instrumented plan, then makes tool calls while streaming tool-call events to a diagnostic pane.
• A RAG-enabled assistant that shows inline citations as citation-generation events arrive and lists sources at completion.
• An async web server endpoint that returns streamed SSE or websocket messages using async_co.chat_stream.
• A CLI tool that prints incremental output and recovers from stream interruptions with a safe_stream wrapper.