chatbot

Chatbot with Tool Calls

A long-running conversational workflow with tool calls and update handlers. Demonstrates how LangSmith traces an agentic loop where the model calls tools across multiple turns.

See the parent README for prerequisites.

Running

# Terminal 1 — start the worker
uv run --group langsmith-tracing python -m langsmith_tracing.chatbot.worker

# Terminal 2 — interactive CLI
uv run --group langsmith-tracing python -m langsmith_tracing.chatbot.starter

Commands in the CLI:

• Type a message and press Enter to chat
• exit — end the session

Tools

The model has two tools, both implemented as @traceable methods on the workflow class:

• save_note(name, content) — Stores a note in the workflow's in-memory dict. The note is durable because workflow state survives crashes and restarts via Temporal's event history.
• read_note(name) — Reads a note from the workflow's in-memory dict.

Architecture

The main @workflow.run method runs a loop that processes user messages (calls _query_openai, which handles the tool-call loop). The message_from_user update handler coordinates: it hands the message to the main loop via a shared _pending_message field, then waits for the response.

This means:

• Activity calls and the tool loop happen inside the main workflow run
• The update handler's trace just shows the input/output of the coordination step

Trace Structure

add_temporal_runs=False (default)

Only @traceable and wrap_openai spans appear. The client-side @traceable wraps start_workflow and each execute_update, so both workflow and update traces nest under it via context propagation.

Chatbot Session a1b2c3d4              (@traceable, client-side)
├── Session Apr 17 10:30               (@traceable, workflow main loop)
│   ├── Request: hello                 (@traceable, per-message in main loop)
│   │   └── Call OpenAI                (@traceable, activity)
│   │       └── ChatOpenAI             (automatic via wrap_openai)
│   └── Request: save that as note 15  (@traceable, per-message in main loop)
│       ├── Call OpenAI                → function_call: save_note
│       ├── Save Note                  (@traceable, workflow method)
│       └── Call OpenAI                → text response
├── Update: hello                      (@traceable, update handler)
└── Update: save that as note 15       (@traceable, update handler)

add_temporal_runs=True

With --add-temporal-runs, Temporal operation spans are added. StartWorkflow/RunWorkflow and StartActivity/RunActivity appear as sibling pairs.

Chatbot Session a1b2c3d4                          (@traceable, client-side)
├── StartWorkflow:ChatbotWorkflow                 (automatic, Temporal plugin)
├── RunWorkflow:ChatbotWorkflow                   (automatic, Temporal plugin)
│   └── Session Apr 17 10:30                      (@traceable, workflow main loop)
│       └── Request: save that as note 15         (@traceable, per-message)
│           ├── StartActivity:call_openai         (automatic, Temporal plugin)
│           ├── RunActivity:call_openai           (automatic, Temporal plugin)
│           │   └── Call OpenAI                   (@traceable, activity)
│           │       └── ChatOpenAI                (automatic via wrap_openai)
│           ├── Save Note                         (@traceable, workflow method)
│           ├── StartActivity:call_openai         (automatic, Temporal plugin)
│           └── RunActivity:call_openai           (automatic, Temporal plugin)
│               └── Call OpenAI                   (@traceable, activity)
│                   └── ChatOpenAI                (automatic via wrap_openai)
├── StartWorkflowUpdate:message_from_user         (automatic, Temporal plugin)
│   └── HandleUpdate:message_from_user            (automatic, Temporal plugin)
│       └── Update: save that as note 15          (@traceable, update handler)
└── ...

Note that the Request: chain (with activity calls) lives under RunWorkflow (the main loop), while the Update: span lives under HandleUpdate (the update handler). They're connected by the shared workflow state but appear as separate subtrees.