The dated announcement
On 1 October 2026, Cloudflare announced AI Search general availability, highlighting multimodal retrieval and OCR for scanned PDFs. This Aivan analysis was prepared on 20 October. The announcement describes the external service, not a capability claim about Aivan.
An editorial equipment-manual example
Use an equipment-panel image to find the corresponding manual among authorized documents. Include visually similar equipment, different model numbers and poor scans. Check that results identify the right device and page, rather than merely matching colors. For Persian documents, evaluate right-to-left text, tables and part numbers independently against the original page.
A practical acceptance plan
Compare text-based and multimodal retrieval on identical inputs. Measure relevant-document position, citation location and processing time. Attribute errors to extraction, retrieval or generation. Evaluate data destinations, permissions and reprocessing costs separately. Image retrieval does not itself guarantee correct interpretation of a chart or table.
Identify a workflow that needs visual retrieval
Investigate cases where staff begin with an equipment photo or scanned page and existing text cannot locate the right source. Missing documents and authorization failures require other remedies. In an editorial maintenance pilot, first evaluate finding the correct manual page; do not immediately turn that result into an automated repair instruction.
Run a paired acceptance exercise
Label clear, blurred, rotated and visually confusing inputs. Compare the existing text workflow and the proposed visual workflow on the same set, including held-out examples. Measure document rank, page accuracy and time to usable evidence. Check Persian table values against their row, column and unit. Asking for a better image may be the appropriate outcome when identification is ambiguous. These are proposed tests, not Aivan test results or vendor guarantees.
Include derived data and reprocessing in the decision
Trace the original file, page images, extracted text and retrieval representations through processing and retention. Test permission changes and deletion across derived outputs and caches. Obtain actual plan-specific reprocessing and price information rather than assuming it. Separate retrieval gains from interpretation errors in the final report. A limited visual-search deployment for one collection may be justified even when replacing the entire knowledge workflow is not.
A practical pilot brief for the equipment-manual scenario
Ask technicians to select confusing but representative examples: similar panels from different device models, low-quality plate photos and scanned tables. Label the exact manual version and page that would help with each request. Keep a separate set unseen during setup. The initial deliverable can be finding usable evidence, without automatically generating maintenance instructions or taking action.
Run the current and proposed workflows with identical inputs. Inspect whether a failure originates in reading the image, retrieving the right source or interpreting the retrieved material. A table result must preserve the relationship between value, unit, row and column. If visual retrieval improves but generated interpretation remains wrong, report those outcomes separately. Asking for a clearer image can be appropriate when the input cannot support identification. These steps are an editorial acceptance proposal, not a report that Aivan has tested or adopted the service.
Before selecting a processing path, trace original and derived data, permissions and deletion behavior. Obtain current terms for the exact option under review rather than assuming all plans share retention or reprocessing behavior. Budget initial processing, updated manuals, waiting time and review effort. Present gains, unchanged cases and new failure modes with concrete examples. A limited addition for one collection may be a stronger decision than a broad replacement. The general-availability announcement justifies investigation; local acceptance still depends on the organization's data and workflow evidence.