The problem this solves
A professional firm—legal, accounting, financial, or similar advisory work—usually has real expertise scattered across dozens of articles, service pages, and PDFs. A visitor with a specific question has no easy way to find the one piece of content that actually answers it, and ends up either giving up or calling in with a question the website already covered.
The approach, in plain terms
Two components make the site’s existing content easier to find without pretending to be a professional adviser: SignalSearch, a smarter search that understands the meaning of a question rather than just matching keywords, and AnswerMap, which points a visitor to the specific source page instead of a vague summary. Every answer is tied to a citation the visitor can click through and verify. If a question falls outside what the approved content actually covers, the system says so and offers a contact path—it is built to refuse and hand off rather than to improvise a legal, medical, or financial conclusion.
How it is actually built
Approved public content is indexed through a provider-neutral embedding adapter (a translation layer that turns text into a searchable format, built so the underlying AI provider can be swapped without rebuilding the feature). Semantic vector search is combined with conventional keyword search, so results stay useful even if the AI layer is unavailable. Any generated text is optional, restricted to the firm’s own approved sources, rate-limited so it cannot be spammed into runaway cost, and sits behind a kill switch the firm can flip off at any time.
What it costs to run, and what happens if a provider goes down
Cloudflare Vectorize (the search index) and Workers AI (the inference layer) can meaningfully lower the cost of running this at launch under light traffic—see the launch allowances for the specific free-tier numbers this scenario stays inside. Conventional keyword search is the stable fallback if either provider is unavailable, and the underlying AI provider can be replaced through the adapter layer without redesigning the public interface or rebuilding the entire content system—the replacement is evaluated before traffic is moved to it.
What this scenario excludes
Generating regulated advice of any kind—legal, medical, or financial conclusions—is explicitly out of scope by design, not an oversight to fix later. Client-portal document exchange and case-management integration are also not part of this scenario.
What would need to be true to accept this as built
Every unsupported question would need to trigger the refusal-and-contact path in testing, every generated answer would need a working citation back to an approved source, and the firm’s own subject-matter reviewer would need to sign off on the approved-content set before launch.