Imagine an afternoon-shift employee trying to answer a question about receiving stock. The procedure is in a PDF manual, but the file is a scan, so searching for the relevant phrase returns nothing. The employee calls the same shift supervisor again, even though that person is already handling another issue.
The problem is often not that people fail to document their work. Procedures sit in different folders, a catalogue is mostly images, or the latest exception lives in the memory of someone who has done the job for years. Searchable knowledge can help a new employee get started without asking the same person every time, and give a team a source document to check when work changes hands.
On October 1, 2026, Cloudflare made AI Search generally available and announced image search and optical character recognition (OCR) for scanned PDFs. OCR extracts text from a scanned file so it can be indexed for search. Cloudflare says PDFs with OCR enabled can be up to 10 MiB, compared with the previous 4 MiB limit. Cloudflare's announcement is relevant to organizations with scanned manuals, work sheets, or product documents. Turning on OCR alone does not prepare a company knowledge base for use.
Start with the document people ask for
Choose one recurring question, such as how to receive a shipment, maintain a machine, or find a product code. Trace it back to the source document. Identify its owner and the version staff are expected to use.
At minimum, document metadata should say which team owns a file, who is accountable for it, when it was last reviewed, and who is allowed to read it. When a new manual replaces an old one, the search index needs to favor the current version and a process must remove expired documents. Naming files manual_final_really_final_v7 is not version control that a team can reliably audit.
Cloudflare AI Search can index documents from sources such as uploaded files, websites, and R2. Its product documentation describes keyword, semantic, and hybrid search. Metadata can also filter results by attributes such as category, version, or language, as explained in the metadata documentation. Those filters are design tools. Do not assume AI Search automatically inherits permissions from SharePoint, an ERP, or an existing document system.
If Accounting and Warehouse use different document sets, design authentication and search boundaries to match the existing permissions before indexing the files. An endpoint that lets everyone in the company search every document can broaden access by accident.
Check what OCR actually reads
OCR turns text in an image into searchable content, but the result depends on the source file and conversion. Test it on your own Thai documents, especially files with skewed pages, tables, numbers, stamps, or handwritten notes. Ask the process owner to check whether important terms, numbers, and exceptions were read correctly.
Begin with a small, lower-risk set such as public manuals or internal SOPs without personal data. Use questions staff actually ask and compare search results with the original documents. Does the right version appear near the top? Can a person open the source and check the surrounding context? If a result concerns machine safety, finance, or an operating instruction, route it to the person responsible before anyone acts on it.
AI Search is a retrieval layer. It does not decide who should approve an SOP or which document has legal authority. If you add answer generation, display the source document, version, review date, and location of the supporting text. People should check the original before relying on a result that could affect safety, money, or a customer commitment.
Estimate the bill before indexing the whole company
Cloudflare says billing for AI Search begins on November 1, 2026. Charges are based on ingested content, stored data, and queries. OCR is billed as image-processing ingestion tokens. The Cloudflare limits and pricing page lists monthly included usage of 5 million ingestion tokens, 10 GB-month of storage, 1,000 semantic queries, and 1,000 full-text queries. Usage beyond those amounts is priced by type: base ingestion, image processing, storage, and each query class.
Before enabling the service, estimate how much text the first document set contains, how many pages need OCR, how long the index will be retained, and how many searches staff may run. Re-indexing after documents change or chunking settings are adjusted can also add ingestion usage. Check the current pricing and plan limits before setting a budget, because provider rates and terms can change.
As of October 12, Cloudflare lists rates above the included usage of $0.75 per million base ingestion tokens, an additional $0.50 per million image-processing tokens, $2 per GB-month of storage, $0.75 per 1,000 semantic, vector, or hybrid queries, and $0.10 per 1,000 full-text queries. Image-processing tokens share the 5-million monthly ingestion allowance. Check the pricing page again before rollout.
Before indexing, classify the information. Check whether it contains employee, customer, or supplier data, or material covered by a contract. Define the purpose, who can search it, how long it will be kept, and how it will be deleted. Review the data-processing terms and any relevant transfers. Choosing a cloud region in Thailand does not, by itself, establish PDPA compliance.
A searchable knowledge base helps when staff can tell which document is authoritative and follow a result back to its source. Start with a small set, test OCR and permissions, then expand based on questions people actually need answered. That gives teams a practical way to share procedures and onboard colleagues without making one experienced employee the only route to an answer.
Talk with Enersys about business systems and team workflows
References
