CLEARCRAFT INSIGHTS · 2026-10-23

An AI knowledge base for Hong Kong SMEs: from repeated questions to source-backed search

When sales, support and the owner give different answers to the same service question, the first opportunity is a shared, current source of truth. Using Cloudflare AI Search as an example, this guide sets out a small pilot that solves one recurring problem before expanding.

Responsible unit: CLEARCRAFT LAB · Content owner: WilsonVersion: 1.0

Enterprise AI search is not a Google AI ranking service

Cloudflare AI Search indexes content selected by a business, including a website, R2 storage or uploaded documents, and can generate answers from retrieved material. It addresses how to find answers in your information; it does not buy recommendations on Google, Gemini or other platforms. [1]

GEO concerns discovery, interpretation and citation by external search and AI platforms. A business knowledge base concerns correct use of its own information. They can share approved public service facts, but customer contracts, internal quotes and personal data should not be mixed into public content as a result.

What recent product changes mean for an SME

Cloudflare's 1 October 2026 release note marks AI Search as generally available, with hybrid search enabled by default for new instances. Usage-based billing is scheduled to begin on 1 November 2026. Evaluate current pricing and intended use alongside the features. [2] [3]

Our recommendation is to model document volume, update frequency, monthly queries, answer generation and associated model or service charges. Set a pilot budget and stopping conditions before adding more users. This article does not quote a fixed monthly cost or claim that ClearCraft has enabled the product for a customer.

Start with one recurring question, not the whole company

Good starting points have a clear answer owner and recur frequently: what to prepare for an initial consultation, what a service includes, or where to find the current delivery checklist. The following is a proposed pilot, not customer performance data.

  • Appoint a content owner and select a small set of approved documents. Record the version, effective date and public or internal classification of each.
  • Separate superseded prices and terms from current versions. Resolve contradictory information with a person before asking AI to choose an answer.
  • Begin with internal search or reply drafting. Keep explicit confirmation steps for external sending, price commitments and booking changes.

Test Traditional Chinese and Cantonese phrasing with real questions

Cloudflare's keyword-search documentation warns that the default porter tokenizer is unsuitable for Chinese, Japanese and Korean. It recommends considering trigram for these languages; terms shorter than three characters cannot match independently, and changing the tokenizer triggers reindexing. This is a product retrieval setting, not a Google ranking technique. [4]

For a Hong Kong team, test formal Traditional Chinese, Cantonese wording, mixed Chinese and English, and abbreviations. Put expressions such as 服務費用, 幾錢 and quotation into the test set and check the retrieved document. Short names and two-character abbreviations need separate testing; a settings change is not proof of correct retrieval.

Show what an answer is based on

Cloudflare documents a pattern for displaying the retrieved chunks, document identifiers and links used in an answer. Available citations do not prove that every generated sentence is supported; the answer still needs to agree with its cited material. [5]

A useful interface puts a short answer, source document, version date and responsible contact together. When the information does not answer the question, show a confirmation or escalation route instead of inventing a price, deadline or service commitment.

Separate public, internal and customer information

Cloudflare documents protecting an AI Search endpoint through a custom domain and Access, including disabling the default hostname that would otherwise allow anonymous queries to bypass that protection. Authentication alone does not establish which documents each user should see: the implementation must also test document permissions and separation between customers. [6]

Keep public FAQs separate from internal document collections and begin acceptance tests with non-sensitive material. A hidden link is not an access control. Retrieved third-party content should not become an instruction to execute actions.

A practical two-week pilot

This is ClearCraft's proposed evaluation schedule, not a guaranteed delivery period. Confirm ownership, data scope and test questions first; sensitive information or complex permissions require separate assessment.

  • Days 1–3: organise approved information, versions and permissions; prepare 20 representative questions and the expected supporting sources.
  • Days 4–7: build a limited prototype and test Chinese and Cantonese phrasing, citations, missing answers and denied access.
  • Days 8–10: ask intended users to review each question. Record correct-source retrieval, incorrect answers, escalations and actual usage costs.
  • Days 11–14: compare the time needed for similar tasks before and during the pilot. Fix gaps and decide whether to stop, extend testing or expand.

The next step is usable information

If people spend time finding documents or repeating answers every day, list the ten most frequent questions and the official source for each answer. That list helps you choose an AI tool and reveals missing explanations on your public website.

ClearCraft's starting point is a set of reviewable decisions: which information may be public, which answers need review, which actions need confirmation, and how to measure whether the workflow genuinely saves time.

OFFICIAL SOURCES · CHECKED 2026-10-23

First-party documents

  1. Cloudflare: How AI Search works
  2. Cloudflare: AI Search release note (1 October 2026)
  3. Cloudflare: Limits & pricing
  4. Cloudflare: Keyword search
  5. Cloudflare: Show source citations in responses
  6. Cloudflare: Access for public endpoints

Revision record

1.0 · 2026-10-23 · Initial publication in Traditional Chinese and English.

Submit verifiable corrections through the editorial and corrections policy.