The purpose and current product context
This guide shows how to use Perplexity to produce a transparent result list with source metadata. Start with a query, retrieval constraints, and a display policy. The central question is: How will the application distinguish retrieved evidence from generated interpretation? A useful workflow has a recognizable finish line and a way to distinguish a supported result from an attractive but incomplete response.
Search API returns ranked structured retrieval results. It is distinct from an Agent API answer-generation workflow. Check current developer documentation for filters, limits, and request schemas. Check the current official guide.
A worked example
A developer retrieves product documentation and displays titles, links, and relevant metadata. A downstream summary is labeled separately. When results are empty, the application explains that no matching evidence was retrieved instead of inventing a general answer that looks sourced.
The example illustrates a review decision, not a reported product test. Before applying it, write down which information in your own situation is confirmed and which is provisional. If you cannot explain the difference, resolve that first. Otherwise, a summary may appear to solve the problem while merely making an uncertain assumption easier to overlook.
Choose the right approach
Keep retrieval and synthesis as distinct stages. Filters can improve relevance but do not establish source quality. Preserve enough metadata to let a reviewer inspect the source, and avoid treating rank as a truth score.
Step-by-step workflow
- Define retrieval constraints. Begin by defining the scope of this step and the evidence it needs. Keep the question visible while collecting context. If a detail is missing, mark it rather than substituting a likely value. This makes later review specific and helps you avoid correcting an answer to a different question.
- Request current structured results. Check the available product controls and the source material before proceeding. Confirm that the intended account and permitted information are being used. Test a small known detail where possible. A setup or access problem should be resolved as such, rather than mistaken for a failure of your research method.
- Preserve source metadata. Request a bounded output and inspect the first result against the requirements. Focus follow-up instructions on a concrete mismatch. Name the omitted condition, unclear relationship, or unsupported statement. Preserve correct evidence during revision so improving presentation does not accidentally alter the factual meaning.
- Handle empty and irrelevant results. Apply the review criteria before using the result elsewhere. Keep consequential claims traceable and unresolved points separate. If the answer needs a decision from another person, record that decision explicitly. Finish with an output someone can understand and review without reconstructing the entire conversation.
A reusable prompt
Use the prompt below as a brief. Replace bracketed fields with approved information and remove requirements that do not apply. Keep source text distinguishable from your instructions. The prompt is useful only when its assumptions match the task; pasting it unchanged does not establish that the resulting answer meets your needs.
Task: How will the application distinguish retrieved evidence from generated interpretation?
Inputs: [provide a query, retrieval constraints, and a display policy]
Audience: [intended reader]
Required result: a transparent result list with source metadata.
Use approved sources and preserve dates, definitions, conditions, and exceptions. Separate confirmed facts, interpretation, and unresolved questions. Do not invent missing details. Show the evidence needed to review consequential claims. Ask one focused question if a missing requirement blocks the result.
Acceptance checks:
- Links remain usable
- Filters are documented
- Rank is not called confidence
- Synthesis is distinguishableReview checklist and evidence table
- Links remain usable
- Filters are documented
- Rank is not called confidence
- Synthesis is distinguishable
Review substance before presentation. Ask whether the output preserves the actual requirement, important conditions, and the status of each statement. Then check readability and navigation. A polished layout can make an unsupported conclusion more persuasive, so visual improvement should follow factual review rather than replace it.
| Check | What to retain | What to do if it fails |
|---|---|---|
| Links remain usable | Source passage, observed result, or explicit decision | Correct the mismatch and mark remaining uncertainty |
| Filters are documented | Source passage, observed result, or explicit decision | Correct the mismatch and mark remaining uncertainty |
| Rank is not called confidence | Source passage, observed result, or explicit decision | Correct the mismatch and mark remaining uncertainty |
| Synthesis is distinguishable | Source passage, observed result, or explicit decision | Correct the mismatch and mark remaining uncertainty |
Common mistakes and repairs
A top-ranked result may still be outdated or off-topic. Review relevance to the query.
Diagnose the failing layer before repeating the entire task. If the source is wrong, repair the input. If the question is ambiguous, narrow it. If the output changes the meaning, correct the specific claim. Keep a record of repeated defects so the reusable instruction or review stage can be improved, rather than patching the same final sentence every time.
Practice the workflow
Design a result card and an empty-result message for a fictional documentation search.
Complete the practice once, record the reviewed outcome, and then change one important condition. Explain what should change in the answer before requesting another response. This small variation tests whether your method handles the task rather than simply reproducing a familiar example. Use the checklist to identify the exact error if your expectation and the response differ.
Save a useful result
Save a transparent result list with source metadata together with its source references and review date. Preserve links remain usable as a visible acceptance condition. Include unresolved questions and the person or evidence needed to settle them. Remove private details from reusable prompts so the method can be shared without exposing unnecessary context.
Download the prompt and checklist for your practice record.
Frequently asked questions
Does Search API itself produce a final researched answer?
The overview describes structured retrieval; distinguish that from Agent API synthesis.
Does rank mean truth?
No. Ranking and evidentiary quality require separate review.
What should an empty result show?
Explain the missing evidence and offer a narrower query, without inventing support.
What should I prepare before starting?
Prepare a query, retrieval constraints, and a display policy. Remove material unrelated to the question. Name the outcome and decide how you will check it. This preparation gives you a basis for judging the answer rather than accepting any polished response as useful.
How do I adapt this workflow to my own situation?
Replace the fictional scenario with approved information and keep the underlying question explicit: How will the application distinguish retrieved evidence from generated interpretation? Adjust the audience and output format. Preserve conditions and source details that affect the result instead of copying the example mechanically.
What if the first answer is too broad?
Return to the intended result: a transparent result list with source metadata. Name the missing requirement and ask for a focused correction. Adding unrelated instructions often makes the task harder to review. Compare the revision with your original question and retain unresolved details.
How should I handle missing information?
Keep it explicitly unresolved. Explain which source, permission, or decision would settle it. A plausible guess can be harder to detect than an obvious gap. For this workflow, completeness means knowing the limits of the answer as well as the confirmed result.
How can I check whether the workflow helped?
Compare the reviewed output with a baseline you understand. Apply the checklist in this guide and record what needed repair. Judge relevance, accuracy, and usability rather than response length. Keep one accepted example to make later reviews consistent.
What should I keep after finishing?
Retain the reviewed result, source versions, prompt, and unresolved questions in the place where work continues. Remove unnecessary private details from reusable examples. Another reader should understand the result without replaying every chat turn or relying on assumptions you never wrote down.
When should I recheck the product instructions?
The official references on this page were checked on 5 October 2026. Revisit them when your interface, account plan, source behavior, or permissions differ. Product controls can change; the practical review method remains useful, but documented availability should always be checked for your own account.
Official sources and related guides
The sources above support product details. The examples, review decisions, and practice exercises are original editorial guidance. When product documentation and your account differ, check the applicable account and platform guidance before following a step. Keep evidence beside consequential conclusions instead of treating links as decorative proof.
Continue with Perplexity File Creation: Review Reports and Slides, Perplexity for Beginners: A Source-First Workflow, NotebookLM for Languages: Learn from Your Sources.
