Debugging Chains: The Detective Work

In Module 4 (Chains), we learned how to pipe many steps together. But as your pipes grow longer, tiny errors propagate.

• Step 1 creates a slight formatting error.
• Step 3 misinterprets that error as a "Math" command.
• Step 5 crashes because it received a string instead of a number.

Debugging AI is harder than debugging code because the "Input" is unpredictable text.

1. The 3 Levels of Debugging

1. Print Debugging: Manually printing the output of every step. (Slow, messy).
2. set_debug(True): A global LangChain flag that prints everything to the console.
3. Observability Platforms: Professional dashboards like LangSmith.

---

2. Using set_debug

If your script isn't working and you want to see the "Raw" data passing through the pipes, run this at the very top:

from langchain.globals import set_debug

set_debug(True)

# Now every invoke() will show the full JSON prompt and response
chain.invoke({"input": "..."})

---

3. Visualizing Propagation Error

graph TD
    S1[Step 1: Write JSON] -->|Missing Comma| S2[Step 2: Parse JSON]
    S2 -->|Fail| S3[Step 3: Correct Error]
    S3 -->|Hallucination| S4[Step 4: Execute Incorrect Tool]
    S4 --> Crash[System failure]

---

4. Common Chain Failure Modes

• Context Stuffing: You retrieved too many documents and hit the token limit.
• Variable Mismatch: You asked for {question} in your prompt, but you sent {"input": "..."} in your dictionary.
• Prompt Drift: The model's tone changes significantly in step 4 compared to step 1.

---

5. Engineering Tip: Atomic Testing

Don't debug the whole chain at once. Test each Runnable individually.

# Test only the prompt
print(prompt.invoke({"input": "test"}))

# Test only the model
print(model.invoke("test"))

---

Key Takeaways

• AI Debugging requires seeing the intermediate data between steps.
• set_debug(True) is the fastest way to see the raw "Guts" of your chain.
• Variable Names must match exactly between the Prompt and the Invoke Dict.
• Testing in isolation is the best defense against complex chain failure.