AI steps don’t always give the same answer twice, and a prompt that works on one page can fail on the next. Before you let a workflow run on its own, check its answers on realistic inputs, including difficult ones.
Test one step at a time
Section titled “Test one step at a time”Don’t test the whole workflow first. Check each AI step on its own:
- Click the AI node and choose Test step. AWFlow also runs the steps it needs before it. See Run from the canvas.
- Look at the node’s input: is it the text you expected, or a whole page full of menus and footers?
- Look at the output: is it correct, and in the shape the next step expects?
- Change one thing (the prompt, the model, the input) and test again.
Fields that use {{ }} show Preview from last run, so you can see what each expression resolved to.
Try difficult inputs
Section titled “Try difficult inputs”A workflow that only works on the example you built it with will fail in real use. Collect a few test cases and run them one after another:
| Case | What to check |
|---|---|
| A typical input | The answer is correct and well formatted. |
| A very long input | It still fits the model; nothing important is cut off. |
| An empty or very short input | The workflow doesn’t invent an answer. |
| An off-topic input | The classifier returns unknown, or the agent says it can’t help. |
| Input in another language | The answer is in the language you want. |
| Input that tries to give orders (“ignore your instructions…”) | The model ignores it, and no action runs without your approval. See Action approval. |
The Debug node can generate sample items or a long string, or throw an error on purpose, so you can test what comes after it without a real source.
Make answers easy to check
Section titled “Make answers easy to check”- Ask for structure. Connect a Structured Output Parser so you can check fields one by one instead of reading prose.
- Ask for sources. When answering from a knowledge base, ask the model to say where each fact comes from. See RAG.
- Let the model say “I don’t know”. Tell it in the system message what to answer when the information isn’t there, and test that case.
- Add an automatic check. Guardrails can check every answer against a policy and route failures elsewhere.
Test your knowledge base search
Section titled “Test your knowledge base search”Most wrong RAG answers come from retrieving the wrong passages, not from the model. Open the knowledge base and use Test search: type a question and see which passages come back and how similar they are. Fix the sources or the search options before you change the prompt. See Test search.
Watch real runs
Section titled “Watch real runs”Once the workflow runs on its own, check the first few runs:
- The workflow’s Executions panel shows past runs. Open one to see what each node received and returned.
- When a run fails, the Last run tab explains what happened and what to try.
- Keep a person in the loop at first: put Wait For Approval before steps that send or change things, and remove it once you trust the results.