Academy
4. Testing Retrieval Quality
Use the built-in search field to verify that your knowledge base returns relevant passages and learn how to improve results.
Every knowledge base page in OUPI includes a built-in search field. This lets you type a question and instantly see the passages the AI would retrieve if that base were used in a conversation, agent, or workflow. Think of it as a preview of what the AI "sees" before it generates an answer. If the returned passages are relevant and complete, your base is well-configured. If they miss the mark, it's a signal to improve your documents or chunking settings.
When you test retrieval, focus on realistic queries — the kind of questions your end users or colleagues would actually ask. Avoid single keywords; use natural-language questions instead. For example, rather than typing "refund", try "What is the refund policy for annual subscriptions?". This mirrors how the AI will search the base in ChatPro, OUPI Studio, Deep Research, or a workflow RAG step, giving you an accurate picture of result quality.
Keep one knowledge base per topic. Retrieval is sharper when a base covers a focused subject rather than being a catch-all repository for every document you own.
If the search field returns irrelevant or incomplete passages, you have two main levers to improve results:
-
Add or refine documents. Upload cleaner, more targeted files that directly address the topics you're testing. Remove outdated or off-topic documents that add noise.
-
Adjust chunking settings. Documents are split into chunks before indexing. Larger chunks preserve more context per result, while smaller chunks work better for short-answer content like FAQs. The overlap setting prevents ideas from being cut at chunk boundaries. The defaults work for most documents, so only change them when test results clearly suggest a mismatch.
If a document failed to process, it won't appear in search results at all. Check for failed statuses on the base page. Re-upload a cleaner version, or ask your administrator to reprocess failed documents from the admin panel.
Open one of your knowledge bases. In the built-in search field, type a real question your users would ask. Review the returned passages: are they relevant and sufficient? If not, try uploading an additional document that covers the gap, wait for processing to complete, then search again.
To verify retrieval quality, use the search field on any knowledge base page with realistic, natural-language questions. If results are off, improve them by adding better documents or adjusting chunk size and overlap. Keep bases focused on a single topic for sharper retrieval, and make sure all documents have processed successfully. Once you're satisfied with the search results, you can confidently use that base across ChatPro, OUPI Studio, Deep Research, and workflows.