Choose a source set that has an owner
Dify describes retrieval-augmented generation as a workflow that prepares source material, retrieves relevant content, and supplies it to a model. The practical value depends on the quality and scope of those sources. Define what the assistant is allowed to answer before importing a large folder.
For an illustrative SaaS help assistant, begin with a few current articles about workspace imports. Assign someone to maintain them. Include the supported file types, size constraints if applicable, and recovery steps for known errors. Keep unrelated product versions separate enough that a correct passage can be recognized during testing.
Connect retrieval to the answer deliberately
Dify's Knowledge Retrieval node accepts a query, searches selected knowledge bases, and returns content that can be supplied to an LLM node as context. Retrieval options include settings that affect which passages are returned. Check the current node configuration rather than assuming a knowledge base is used merely because it exists in the workspace.
Draw the import-help flow: user question, retrieval, model response, and displayed answer. Confirm which variable carries the retrieved material into the prompt. Test with a distinctive fact from a sample source. If the model can only answer generic import questions, inspect whether the intended context was actually passed through.
Write questions with an expected evidence record
Create a small evaluation sheet before adjusting retrieval settings. For every question, identify the source passage that should support the answer and the facts the answer must preserve. Add questions that the source set cannot answer so the assistant's limits are observable.
The sample import assistant should distinguish a supported file format from an unrelated export format. It should also handle a question about a feature the product does not document. Expected evidence makes a failed test specific: the passage was missing, the wrong version was retrieved, or the answer changed a supported fact.
| Question type | Expected behavior | Failure to investigate |
|---|---|---|
| Documented task | Use the current import instructions | Relevant passage absent |
| Different wording | Find the same instructions despite phrasing | Only exact keywords succeed |
| Version-sensitive detail | Use the intended current source | Outdated document outranks current guidance |
| Undocumented capability | Explain the limit or request clarification | Answer invents an unsupported feature |
Inspect retrieval separately from the final answer
Dify provides a retrieval testing page for simulating queries and comparing results. Its documentation states that adjustments made there are temporary for the test session. When a setting improves the evaluation set, apply it to the intended production configuration and verify the application again.
If the expected import passage is missing, investigate the source extraction, chunk boundaries, selected knowledge base, and retrieval settings before rewriting the answer prompt. Change one relevant setting at a time and retain the same questions. A test that changes both the source material and every query cannot clearly show why the results improved.
Check that each citation supports the claim beside it
Dify supports using knowledge in applications and configuring citations in relevant app workflows. Review how sources appear in the interface you publish. A visible source reference helps a reader investigate an answer, but the cited passage still needs to support the specific claim.
For the sample assistant, compare every instruction with the retrieved source, including the order of steps and any stated limits. Test whether links open the intended material. If the response mentions an unsupported file type but cites a page about another format, the citation has made the error traceable without making it correct.
Keep an evaluation record as the sources change
Record the source revision, retrieval configuration, model choice, answer quality, and response time for the fixed questions. Use synthetic or approved sample content during early experiments. When instructions change, update the expected answer and rerun affected cases before treating the assistant as current.
Dify is a candidate when visual orchestration and knowledge workflows fit the team's needs. Judge the first version by whether it answers the supported import questions accurately and handles missing evidence clearly. Expand the source set after the team can diagnose failures in this smaller workflow.
Sources and editorial method
Official sources linked in this guide were checked on October 27, 2026. Features, interfaces, and commercial limits can change. The examples are proposed workflows and illustrative calculations, not customer results or a hands-on product benchmark. Confirm the current documentation before implementing the workflow.