Short answer: Treat an AI workflow as connected parts: model input, instructions, tools, retrieved context, and final output. Test each boundary with a small prompt and a known document.

What should the model input contain?

Pass the user request and only the context needed for the task. Keep system instructions separate from data supplied by a user or external document. This makes prompt changes easier to review.

Normalize empty fields before the model node. A missing customer name or blank document section should be explicit rather than an accidental expression result.

How should tools be connected?

Give each tool one clear purpose. A lookup tool should return the fields needed for the answer. A write tool should require the fields needed for the change. Do not expose a broad database operation when a narrow query is enough.

Log the tool name and outcome, not secret arguments. Test a tool failure separately from a model failure so the alert identifies the correct boundary.

Where does a vector store fit?

A vector store can retrieve text that is semantically related to a query. The retrieved chunks still need source metadata, sensible chunk size, and a relevance check. Retrieval does not guarantee that a passage answers the question.

Return the source identifier with each chunk. During testing, inspect the retrieved text before asking the model to summarize it.

How should an AI node be evaluated?

Keep a small set of representative prompts. Check whether the output follows the required format, cites the expected source, and declines unsupported requests. Test an empty retrieval result and a tool timeout.

Use the vector store integration guide for retrieval design. For deterministic transformations around the model, see the Code node examples.