When using model-based decision-making non-deterministic systems (e.g., LLM-driven agents), it is useful to closely examine their decision-making processes:
🤔 Understanding Reasoning: Analyzing the steps that lead to successful outcomes.
🐞 Debugging Errors: Identifying where and why errors occur.
🔍 Exploring Alternatives: Testing different paths to discover better solutions.
We refer to these debugging techniques as time travel, consisting of two key operations:
The replay function allows us to revisit and reproduce the past operations of the agent. This can be done from the current state of the graph (or a checkpoint) or from a specific checkpoint. To replay from the current state, simply pass None as the input along with the thread id:
thread = {"configurable": {"thread_id": "1"}}for event in graph.stream(None, thread, stream_mode="values"): print(event)
To replay operations from a specific checkpoint, first retrieve all checkpoints of the thread:
all_checkpoints = []for state in graph.get_state_history(thread): all_checkpoints.append(state)
Each checkpoint has a unique ID. After identifying the desired checkpoint (e.g., xyz), include its ID in the configuration:
config = {'configurable': {'thread_id': '1', 'checkpoint_id': 'xyz'}}for event in graph.stream(None, config, stream_mode="values"): print(event)
The graph effectively replays previously executed nodes rather than re-executing them, leveraging its awareness of previous checkpoint executions.
Forking allows re-accessing past operations of the agent and exploring alternative paths in the graph.
To edit a specific checkpoint (e.g., xyz), provide its checkpoint_id when updating the graph state:
config = {"configurable": {"thread_id": "1", "checkpoint_id": "xyz"}}graph.update_state(config, {"state": "updated state"})
This will create a new fork checkpoint xyz-fork, from which we can continue running the graph:
config = {'configurable': {'thread_id': '1', 'checkpoint_id': 'xyz-fork'}}for event in graph.stream(None, config, stream_mode="values"): print(event)
