Website, PDF, product feed or API: which sources should your chatbot use?

Use website pages and documents for explanations and policies, a product catalog for searchable product information, and a configured API action for current records such as a particular order. The choice depends on what the question needs and how quickly the answer can change.

Uploading everything into one knowledge base is tempting. It also makes it easy to mix last year's price list, today's delivery conditions and private customer records. Give each kind of information a clear home before adding sources.

A source map for common questions

An API is a connection mechanism, not a guarantee of freshness. A connected system may itself update only periodically. Likewise, a catalog imported this morning is a useful source, but it is not proof that the last unit is still on the shelf now.

Website pages: good for information your customers already use

A website is a practical starting point when your policies and service descriptions are maintained there. Import the pages that answer recurring questions: delivery, returns, contact details, service scope and product guidance.

After processing, inspect the extracted text. A visitor may understand a price table from its visual layout while the imported text loses the relationship between a price and its column heading. Collapsed sections, cookie notices and repeated navigation can also affect what ends up in the source.

Start with a few pages whose content you can verify. Test one question per important condition. Then expand. If a page cannot be imported, inspect the source status and follow the website import troubleshooting guide. Repeating the question in chat does not repair an unsuccessful crawl.

In Breezaro, websites and uploaded documents become knowledge sources for the selected bot. The knowledge-base documentation describes importing and reviewing them. A second bot needs its own relevant sources; a matching business name or website does not mean it shares the first bot's knowledge.

PDFs and documents: good for controlled explanations

Documents are useful for a service guide, assembly instructions or a policy that does not yet have a suitable webpage. Use clear headings, complete sentences and explicit conditions. A small document about one topic is often easier to maintain than a large brochure containing several unrelated offers.

Consider this fictional rule:

Collection-point delivery in Ireland is free for a goods total of at least €50 after discounts. The delivery fee does not count toward that amount. Oversized products and delivery to other countries are excluded.

Every part answers a plausible follow-up question. “Free delivery from €50” leaves those conditions unstated.

Check whether text can actually be extracted from the file. A scanned page or a diagram should not be assumed to become usable knowledge merely because the upload succeeded. Read the processed result, including numbers and headings. Keep private customer lists and account records out of general answer material.

Choose one owner for each document and record when it needs review. When a replacement becomes current, remove or update conflicting copies. The process matters more than the file extension.

Product feeds: good for a structured, searchable catalog

A product feed separates information into fields such as title, price, currency, availability and product URL. This makes it suitable for maintaining a catalog that the assistant can search. Breezaro supports sources including Shopify's public product feed, supported XML feeds and product detection from website pages. The product-source documentation explains those options.

Review a small sample after the first import. Include an item with variants, an unavailable item and a discounted item. Check that the price and URL correspond to the same product or variant. A plausible card with a mismatched price can be more misleading than an explicit missing result.

Descriptions need to contain the properties customers search for. A feed with only “Model X, €69” cannot supply a reliable explanation of waterproofing, fit or compatibility. Add those facts to the owning product data and check what reaches the catalog.

Also test what is absent. A search that returns several backpacks does not establish that every matching backpack has been found. Avoid turning a limited result set into an absolute statement about your entire range.

API actions: good for a specific live lookup

An order status, account entitlement or appointment belongs to an operational system. Copying its records into a general knowledge base creates stale information and an inappropriate way to handle private data.

A configured Custom Action can query an HTTPS endpoint during the conversation. Decide which records the visitor may access and which response fields the assistant needs. Verification and ownership are separate: proving an email address does not automatically authorise access to every order number.

Separate reads from changes. Checking available times is different from making a booking. In Breezaro, data-changing actions use a visitor confirmation card in the web widget; those write actions are unavailable on the connected messaging channels. Your endpoint must still enforce its own rules and return a meaningful result. The worked action examples show how the pieces fit together.

Keep a small source register

You do not need a complex content project. A simple table maintained by your team is enough:

After a change, check the original source, the imported result and a few answers. Keep the questions so the next update is easier to verify. The answer-repair guide follows that process in detail.

To start in Breezaro, choose one reliable policy page and one representative catalog source. Add a live action when you have a concrete question that needs it. You will have a smaller set of sources to maintain and a clear explanation for where each answer should come from.