openai/chatkit-python
Publicmirrored from https://github.com/openai/chatkit-pythonAvailable
docs/guides/update-client-during-response.md
164lines · modeblame
df03dc0bJiwon Kim7 months ago | 1 | # Update the client during a response |
| 2 | | |
| 3 | Keep your UI responsive while the server is working: stream progress text, trigger client-side effects, and run client tools mid-response without blocking everything else. | |
| 4 | | |
| 5 | This guide covers three patterns: | |
| 6 | | |
| 7 | - Progress updates for lightweight status while tools run | |
| 8 | - Client effects for fire-and-forget UI behavior | |
| 9 | - Client tools for round-trips to the browser/app during inference | |
| 10 | | |
| 11 | ## Show progress while tools run | |
| 12 | | |
| 13 | Use `ProgressUpdateEvent` when you need lightweight, real-time status. These updates stream immediately to the client and disappear after the turn—they are not stored in the thread. | |
| 14 | | |
| 15 | ### From tools | |
| 16 | | |
| 17 | Inside a tool, use `AgentContext.stream` to enqueue progress events. They are delivered to the client immediately and are not persisted as thread items. | |
| 18 | | |
| 19 | ```python | |
| 20 | from agents import RunContextWrapper, function_tool | |
| 21 | from chatkit.agents import AgentContext | |
| 22 | from chatkit.types import ProgressUpdateEvent | |
| 23 | | |
| 24 | | |
| 25 | @function_tool() | |
| 26 | async def ingest_files(ctx: RunContextWrapper[AgentContext], paths: list[str]): | |
| 27 | await ctx.context.stream(ProgressUpdateEvent(icon="upload", text="Uploading...")) | |
| 28 | await upload(paths) | |
| 29 | | |
| 30 | await ctx.context.stream( | |
| 31 | ProgressUpdateEvent(icon="search", text="Indexing and chunking...") | |
| 32 | ) | |
| 33 | await index_files(paths) | |
| 34 | | |
| 35 | await ctx.context.stream(ProgressUpdateEvent(icon="check", text="Done")) | |
| 36 | ``` | |
| 37 | | |
| 38 | `stream_agent_response` will forward these events for you alongside any assistant text or tool call updates. | |
| 39 | | |
| 40 | ### From custom pipelines | |
| 41 | | |
| 42 | If you are not using the Agents SDK, yield `ProgressUpdateEvent` directly from your `respond` or `action` methods while your backend works: | |
| 43 | | |
| 44 | ```python | |
| 45 | async def respond(...): | |
| 46 | yield ProgressUpdateEvent(icon="search", text="Searching tickets...") | |
| 47 | results = await search_tickets() | |
| 48 | | |
| 49 | yield ProgressUpdateEvent(icon="code", text="Generating summary...") | |
| 50 | yield from await stream_summary(results) | |
| 51 | ``` | |
| 52 | | |
| 53 | Use short, action-oriented messages and throttle updates to meaningful stages instead of every percent to avoid noisy streams. | |
| 54 | | |
| 55 | ## Trigger client-side effects without blocking | |
| 56 | | |
| 57 | Send `ClientEffectEvent` to trigger fire-and-forget UI work (such as refreshing a view, opening a modal, or showing a toast) without creating thread items or pausing the model stream. | |
| 58 | | |
| 59 | Client effects are ephemeral: they stream immediately to ChatKit.js, trigger your registered effect handler, and are not persisted to the thread history. Use client tool calls instead when you need a round-trip response from the client. | |
| 60 | | |
| 61 | ### Stream a client effect from your server | |
| 62 | | |
| 63 | Yield client effects directly from the `respond` or `action` method: | |
| 64 | | |
| 65 | ```python | |
| 66 | async def respond(...): | |
| 67 | yield ClientEffectEvent( | |
| 68 | name="highlight_text", | |
| 69 | data={"index": 142, "length": 35}, | |
| 70 | ) | |
| 71 | ``` | |
| 72 | | |
| 73 | Or from tools, through `AgentContext`: | |
| 74 | | |
| 75 | ```python | |
| 76 | from agents import RunContextWrapper, function_tool | |
| 77 | from chatkit.agents import AgentContext | |
| 78 | from chatkit.types import ClientEffectEvent | |
| 79 | | |
| 80 | | |
| 81 | @function_tool() | |
| 82 | async def highlight_text(ctx: RunContextWrapper[AgentContext], index: int, length: int): | |
| 83 | await ctx.context.stream( | |
| 84 | ClientEffectEvent( | |
| 85 | name="highlight_text", | |
| 86 | data={"index": index, "length": length}, | |
| 87 | ) | |
| 88 | ) | |
| 89 | ``` | |
| 90 | | |
| 91 | ### Handle the client effect in ChatKit.js | |
| 92 | | |
| 93 | Register a client effect handler when initializing ChatKit on the client: | |
| 94 | | |
| 95 | ```ts | |
| 96 | const chatkit = useChatKit({ | |
| 97 | // ... | |
| 98 | onEffect: async ({name, data}) => { | |
| 99 | if (name === "highlight_text") { | |
| 100 | const {index, length} = data; | |
| 101 | highlightArticleText({index, length}); | |
| 102 | // No return value needed | |
| 103 | } | |
| 104 | }, | |
| 105 | }); | |
| 106 | ``` | |
| 107 | | |
| 108 | ## Call client tools mid-inference | |
| 109 | | |
| 110 | Client tool calls let the agent invoke browser/app callbacks mid-inference. Register the tool on both client and server; when triggered, ChatKit pauses the model, sends the tool request to the client, and resumes with the returned result. | |
| 111 | | |
| 112 | Use client effects instead when you do not need to wait for the client callback response for the rest of your response. | |
| 113 | | |
| 114 | ### Define a client tool in your agent | |
| 115 | | |
| 116 | Set `ctx.context.client_tool_call` inside a tool and configure the agent to stop at that tool. Only one client tool call can run per turn. Include client tools in `stop_at_tool_names` so the model pauses while the client callback runs and returns its result. | |
| 117 | | |
| 118 | ```python | |
| 119 | from agents import Agent, RunContextWrapper, StopAtTools, function_tool | |
| 120 | from chatkit.agents import AgentContext, ClientToolCall | |
| 121 | | |
| 122 | | |
| 123 | @function_tool(description_override="Read the user's current canvas selection.") | |
| 124 | async def get_selected_canvas_nodes(ctx: RunContextWrapper[AgentContext]) -> None: | |
| 125 | ctx.context.client_tool_call = ClientToolCall( | |
| 126 | name="get_selected_canvas_nodes", | |
| 127 | arguments={"project": my_project()}, | |
| 128 | ) | |
| 129 | | |
| 130 | | |
| 131 | assistant = Agent[AgentContext]( | |
| 132 | ..., | |
| 133 | tools=[get_selected_canvas_nodes], | |
| 134 | # StopAtTools pauses model generation so the pending client callback can run and resume the run. | |
| 135 | tool_use_behavior=StopAtTools(stop_at_tool_names=[get_selected_canvas_nodes.name]), | |
| 136 | ) | |
| 137 | ``` | |
| 138 | | |
| 139 | ### Register the client tool in ChatKit.js | |
| 140 | | |
| 141 | Provide a matching callback when initializing ChatKit on the client. The function name must match the `ClientToolCall.name`, and its return value is sent back to the server to resume the run. | |
| 142 | | |
| 143 | ```ts | |
| 144 | const chatkit = useChatKit({ | |
| 145 | // ... | |
| 146 | onClientTool: async ({name, params}) => { | |
| 147 | if (name === "get_selected_canvas_nodes") { | |
| 148 | const {project} = params; | |
| 149 | const nodes = myCanvas.getSelectedNodes(project); | |
| 150 | return { | |
| 151 | nodes: nodes.map((node) => ({id: node.id, kind: node.type})), | |
| 152 | }; | |
| 153 | } | |
| 154 | }, | |
| 155 | }); | |
| 156 | ``` | |
| 157 | | |
| 158 | ### Stream and resume | |
| 159 | | |
| 160 | In `respond`, stream via `stream_agent_response` as usual. ChatKit emits a pending client tool call item; the frontend runs your registered client tool, posts the output back, and the server continues the run. | |
| 161 | | |
| 162 | When the client posts the tool result, ChatKit stores it as a `ClientToolCallItem`. The continued inference after the client tool call handler returns the result feeds both the call and its output back to the model through `ThreadItemConverter.client_tool_call_to_input`, which emits a `function_call` plus matching `function_call_output` so the model sees the browser-provided context. | |
| 163 | | |
| 164 | |