Skip to main content

Debugging Chains

Debugging chains can be challenging since they often involve preprocessing inputs and postprocessing outputs. Setting verbose=True when creating a chain prints out internal states of the chain during execution, providing visibility into what's happening inside.

Enabling Verbose Mode​

Here's how to enable verbose mode for different chain types:

from langchain.chains import LLMChain

llm_chain = LLMChain(llm=llm, prompt=prompt, verbose=True)
llm_chain.run("Hello")
from langchain.chains import HumanChain

human_chain = HumanChain(llm=llm, prompt=prompt, verbose=True)  
human_chain.run("What is 2 + 2?")
from langchain.chains import ConversationChain

conversation = ConversationChain(
  llm=chat,
  memory=ConversationBufferMemory(),
  verbose=True
)
conversation.run("What is ChatGPT?")

Interpreting Verbose Output​

The verbose output contains internal chain states during execution. Here's what to look for:

  • Formatted prompt: The prompt after input formatting. Checks if prompt is formatted as expected.

  • Memory variables: Any memory variables injected into the prompt. Checks if memory is populated and injected correctly.

  • Conversation context: For conversation chains, the current context at each turn. Checks if context is maintained properly across turns.

  • LLM requests: The prompts sent at each step. Check for mismatches with expected prompts.

  • Outputs: The raw LLM outputs before postprocessing. Verify outputs match prompts.

Troubleshooting Tips​

Here are some common issues you can debug with verbose mode:

  • Infinite loops: The verbose logs will show the chain stuck in a loop. Inspect the requests and outputs to identify causes.

  • Prompt mismatches: Verbose logs highlight mismatches between expected and actual prompts. Fix formatting issues.

  • Memory errors: Errors loading memory variables indicate issues with the memory component. Inspect verbose memory logs.

  • Context mismatches: For conversations, mismatches between verbose context logs and requests indicate context maintenance issues.

  • Output errors: Raw output errors usually indicate an LLM problem rather than a chain issue. Try modifying prompts.

Enabling verbose mode provides visibility into these issues. Use the logs and guidance above to identify and debug chain problems.