Data Sources & Knowledge Base

How the Bot Gets Its Knowledge

For the chatbot to answer customer questions correctly, it needs content: product descriptions, shipping info, return policies, FAQs and everything else you have on your site. Alpheris gathers this knowledge in two ways, which you can combine freely:

Both sources land in the same knowledge base. From the text, Alpheris creates searchable knowledge chunks that the bot draws on when answering. The bot can't invent sources that aren't in the knowledge base — that's exactly what grounded means.

Website Crawl

When you set up your account, Alpheris crawls your domain automatically and builds the first version of the knowledge base from it. The crawler works in a structured way:

• It follows your sitemap (including the WordPress-specific wp-sitemap.xml) to find pages reliably.
• It extracts product and page metadata (e.g. Open Graph and product meta), not just the plain body text.
• It detects the platform your site runs on and uses matching patterns to extract content cleanly.

The scope of the initial crawl is tuned per plan (on the Basic plan, for example, up to around 220 URLs are captured initially). For most shops that's enough to cover the most important pages right away.

Updating the Crawl

When your site changes — new products, updated shipping terms — you start a re-crawl in the dashboard so the knowledge base stays current. The crawl runs in the background; you can keep using the console while it does.

File Upload

Not everything is on the website — price lists, data sheets, internal FAQs or manuals often exist only as documents. You upload such content directly as a file in the console's Knowledge section. Alpheris reads out the text, breaks it into knowledge chunks and adds them to the knowledge base — just as if they came from the crawl.

Allowed File Types

Common document, text and spreadsheet formats are supported, as well as images (from which text is recognized via OCR):

From images and scanned documents, the contained text is extracted via optical character recognition (OCR), so even a photographed data sheet becomes usable.

Blocked File Types

For security reasons, script-capable and executable files are generally not allowed. They are rejected on upload — regardless of whether they show up via the file extension or the detected content type.

• .html and .svg are not allowed. Both can contain active code (scripts) and are therefore not on the allow list. If you want to include the content of an HTML page, use the Website Crawl instead, or copy the text into a .txt or .md file.
• Executable formats such as .exe, .msi, .bat, .cmd, .ps1, .sh, .js, .jar and others are blocked entirely.
• Macro-enabled Office files (.docm, .xlsm, .pptm and others) are also blocked — upload the macro-free variant (e.g. .xlsx) instead.

Embedding & Retrieval

Crawl and uploads produce raw text — the interesting part is how the bot gets from that to the right answer. To do so, every piece of content goes through two stages.

1. Indexing (Embedding)

Each knowledge chunk is made searchable in two forms:

• Full-text index (FTS): for literal searches of specific terms — such as a product name or article number.
• Vector embeddings: Each chunk is additionally translated into a numeric meaning representation (via the mistral-embed model). This way the bot also finds content that matches the question in meaning, even if the customer uses different words than your texts.

2. Retrieval

When a customer asks a question, Alpheris searches hybrid — full text and vector together — for the most fitting chunks. The best candidates are then sorted once more by relevance through a reranking step, so that only the most relevant content flows into the answer. Only on this basis does the AI formulate its response.

Knowledge Base Quality

The bot's answer quality depends directly on what's in the knowledge base. A few rules of thumb help:

• Completeness: Make sure the most common questions are covered — shipping, returns, payment methods, delivery times, contact. If a piece of info is missing from the website, add it as a file.
• Freshness: After changes to products or terms, start a re-crawl. You can delete outdated chunks individually in the dashboard.
• Clarity: Short, unambiguous texts are retrieved better than long prose covering many topics. A clearly structured FAQ is worth its weight in gold.
• No duplicates: Don't upload the same content twice (e.g. once as a PDF and identically as a website page); it dilutes the result list.

In the console's Knowledge section you can also create, edit or remove individual entries — for example to store an important answer by hand or correct a poorly extracted section.

Domains & Whitelist

Alpheris only crawls and answers content from your own, verified domain — not arbitrary third-party sites. The domain stored in the dashboard is the permitted source for the crawl, and your install ID is bound server-side to exactly this domain.

• During onboarding, your main domain is registered as the verified source; for Shopify and WordPress installations this happens automatically.
• The www. variant is created automatically for Shopify/WordPress, so both spellings are covered.
• Strict domain mode is active by default: widget requests are checked against the host of the embedding page, and foreign domains are rejected.

Which domains are allowed, how you add more, and how verification works is described in detail under Domains.

Privacy & Storage Location

All content in your knowledge base is stored in Germany (hosting at Hetzner, data centers in Nuremberg/Falkenstein). To generate answers, we pass the relevant excerpts to the AI providers Anthropic or OpenAI (USA, EU Standard Contractual Clauses) and Mistral AI (EU); the search vectors are created by Mistral AI in the EU. The data is strictly separated per tenant — other shops have no access to your knowledge base.

You'll find the data processing agreement (DPA) at alpheris.com/avv. Details on upload security, data location and subprocessors are under Security & Privacy.