Configure

Knowledge base

The knowledge base is what turns a generic chatbot into an expert on your product. Quincer AI accepts three input types: URLs, uploaded documents, and manually written Q&A pairs. All three end up in the same retrieval store.

Where to edit

Open Knowledge in the dashboard sidebar. You'll see three tabs:

• Items — everything the agent can cite: imported URLs, uploaded documents, and hand-written Q&A pairs.
• Playbook — your sales playbook: value props, ROI data, objection handlers, competitor comparisons, pricing, and case studies. Covered in Sales automation.
• Audit — search what the agent would retrieve for a question and trace an answer back to the exact item, with one-click edit. Useful when the agent says something wrong.

Importing from a URL

Click Import URL at the top of the page to open the Import from URL dialog shown below. Paste the page address into the URL field, then tick Crawl sub-pages to let Quincer follow the site's sitemap and links — the first few pages import immediately and the rest queue up on a background crawler. Imported pages count toward your plan's knowledge-item cap (see the per-plan limits table below). The Personas row scopes the source (leave all off to share it with every persona), the Sync schedule menu keeps it fresh automatically (No auto-sync, every 5 or 15 minutes, hourly, or daily), and Content volatility (Static, Dynamic, or Live) tells the AI how much to trust the cached copy when answering. Click Crawl & import (or Import & process for a single page) to start.

If the URL is blocked by a firewall (403)

Some sites sit behind a Web Application Firewall (Cloudflare, Sucuri, Akamai, AWS WAF) that blocks automated requests before they reach the origin. When that happens, Quincer AI returns a 403 Forbidden and the import fails with a message like “Blocked by the site's firewall.”

The fix is to allowlist Quincer AI's crawler on the site. You don't need to share any IP address — our crawler announces itself with a stable User-Agent string:

Mozilla/5.0 (compatible; QuincerBot/1.0; +https://chat.quincer.com)

Cloudflare (most common)

1. In the Cloudflare dashboard, select the site, then go to Security → WAF → Custom rules.
2. Click Create rule and name it Allow QuincerBot.
3. Under When incoming requests match, set:
   • Field: User Agent
   • Operator: contains
   • Value: QuincerBot
4. Under Then take action, choose Skip and check:
   • All remaining custom rules
   • Rate limiting rules
   • Managed rules
   • All Super Bot Fight Mode rules (if enabled)
5. Click Deploy. Changes take effect within ~30 seconds.

Reference: Cloudflare — Skip actions for custom rules. If Super Bot Fight Mode is on, also see Verified bots.

Sucuri

Dashboard → Firewall (WAF) → Access Control → Whitelist User-Agent. Add QuincerBot.

AWS WAF / CloudFront

Add a Custom rule with a String match on the User-Agent header containing QuincerBot, action Allow, priority above any blocking rules.

Akamai

In App & API Protector, create a Client Reputation or Bot Manager exception matching User-Agent QuincerBot, action Allow.

If the WAF rule didn't help: the web server is blocking us

A site can sit behind Cloudflare and still refuse us at its own web server. Cloudflare passes the request through, the origin returns 403, and the response still carries Cloudflare's headers — so it looks like a firewall block when it isn't one. No WAF rule and no bypass token can fix this, because nothing at the firewall is saying no.

Quincer AI tells these apart and says so in the error: an origin block reads “the site's own web server refused the request” rather than naming a firewall vendor. The most common cause is a rule that blocks any user agent containing the substring bot — which catches QuincerBot along with everything else. We automatically retry once with a user agent that still identifies us but omits that word, so many of these now resolve without anyone changing anything.

When the retry is refused too, the change has to happen wherever the site is hosted:

• IIS — Request Filtering, or a URL Rewrite rule matching HTTP_USER_AGENT.
• Apache — a mod_rewrite or mod_security rule on User-Agent.
• nginx — an if ($http_user_agent ~* ...) block.
• Managed CMS / website builder — the platform's own bot-blocking setting. You may need to ask the vendor.

Allowing a user agent containing Quincer, or the x-quincer-bypass-token header, is enough at any of them.

Uploading documents

Click Upload file to open the dialog below. The PDF or Markdown file picker accepts a .pdf or .md file up to 10 MB; its text is extracted, chunked, and embedded for AI search. As with a URL import, the Personas row scopes the document to one or more personas (leave all off to share it everywhere). Click Upload & process and the dialog reports back how many chunks it created.

Import inventory (CSV catalog)

If you sell from a product list that isn’t on Shopify — a car dealership’s inventory, a rental fleet, an equipment list — use Import inventory on the Knowledge base toolbar instead of pasting the raw spreadsheet as one item. A whole CSV in a single knowledge item doesn’t retrieve well; Import inventory turns each row into its own item so the agent can find the exact match and show it as a card — with its photo.

• Pick your CSV. We detect the columns automatically and show a preview with a few product cards so you can confirm before importing.
• Map the columns. Confirm which column is the title (or let us combine year/make/model), the price, the image URL, the link to the item’s page, a category, and a unique id (VIN or stock #). Every other column is kept as a searchable spec.
• Filter out sold/test rows. If your export has a status column, list the values that mean “available” (e.g. available) and everything else is skipped — so sold or internal rows never reach customers.
• Photos. Map a column of image URLs and each item’s photo appears on its card. (Some image hosts block off-site display; if a photo doesn’t show, that host is blocking it.)
• Keeping it current. Re-import anytime. Leave “this is the full current inventory” checked and items not in the new file (sold, removed) are dropped; items are matched by the unique id so prices and photos update in place.
• Per-location stock. On a multi-location brand, run Import inventory from a location’s knowledge page instead and the catalog belongs to that branch: its shoppers see its stock plus anything imported brand-wide, never another branch’s, and each branch’s re-import (including the “full current inventory” cleanup) only touches its own items.

Once imported, shoppers get a visual, filterable panel (“show me SUVs under $40k”) and the agent can answer questions about a specific item in text or on a call — the same import powers both. Structured number ranges (price) filter exactly; other specs are matched leniently.

Real-estate listings (automatic)

If your website is a real-estate site, you don’t need to export anything. When Quincer AI crawls your site and finds a page of property listings, it reads them as individual properties rather than as page text — each one becoming its own card with its photo, price and address, exactly like the CSV inventory above. Beds and baths are captured too and are used to narrow a search (“three-bedroom”), though the card itself shows the photo, address and price. Nothing to switch on.
Extra detail held on each property’s own page — property type, square footage, taxes and the full description — is not collected by the automatic sync, because that means visiting every listing individually. Ask us to run a full import if you want those too.

This matters more than it sounds. Read as ordinary page text, a page of twenty listings becomes one block in which a price can end up next to the wrong address, and two things go wrong that are easy to miss:

• Rentals stop being confused with sales. A $1,999/month lease and a $975,000 sale look the same in flattened text. Each listing is tagged sale or lease, so “how much is the two-bedroom?” can’t be answered with a monthly rent for a property that’s for sale.
• Sold properties are never offered as available. Sold listings are usually built exactly like active ones — only the page they sit on tells them apart. Anything shown as sold is never offered to a customer as available, and when a property you had listed turns up on your sold page, it is removed from what the assistant can offer on the next sync.
  This depends on your site having a sold or recently-sold page that we can read: if yours doesn’t, a property that simply disappears from your active listings may still be offered until the page it came from is re-synced. Tell us if that’s your setup and we’ll sort it out.

Ask for “three-bedroom homes in Barrie under $900k” and shoppers get the same visual, filterable panel as any other catalog. On a phone call there is no screen, so the assistant answers about specific properties in conversation instead.

Writing Q&A pairs

A hand-written Q&A is the fastest way to handle “what do we say about X”-type questions with a canned answer. Click Add item to open the Add knowledge item dialog below. Put the question the way a customer would ask it in the Title field (Quincer matches semantically, so you don't need every phrasing), write the on-brand answer in the Content box (Markdown links are supported), and use the Personas chips to limit which personas cite it — leaving them all off makes the pair available to every persona. Click Create to save.

Retrieval ranking

Every time the agent answers, it retrieves the top N chunks from your knowledge base. The ranking uses both semantic similarity (via embeddings) and recency. Sources you've pinned are weighted more heavily. Pin frequently-cited docs from the three-dot menu on each source row.

Per-persona knowledge scoping

When a widget runs multiple personas, each knowledge item can be assigned to a specific persona or left shared (unassigned). Scoping happens before ranking — it decides which items a persona is even allowed to retrieve.

The rule: a persona answers only from its own assigned knowledge plus shared (unassigned) knowledge. It never retrieves knowledge assigned to a different persona. Shared, corporate-wide knowledge — anything you leave unassigned — is visible to every persona.

This is what lets one widget run several personas without cross-persona bleed. A Sales persona and a Support persona, or one persona per location or brand, can share your brand-wide facts while each one cites only the knowledge that belongs to it.

How to assign. Set a knowledge item's persona when you add or edit it (the Personas chips in the Add knowledge item and Upload file dialogs). Add an item from a persona's Knowledge tab and it is scoped to that persona automatically; on the main knowledge page you can filter by persona and by Shared. Leave the persona empty to make an item shared across all personas.

Keeping content fresh

Quincer AI auto-refreshes URL-based sources on a rolling 24-hour schedule. You can:

• Click Refresh now on any source to re-crawl immediately.
• Toggle a source Disabled temporarily without deleting it.
• Delete a source to purge its chunks completely.

Debugging retrieval

If the agent gives a wrong answer, check the Chunks tab and use the search bar. Type the visitor's question and you'll see the top-ranked chunks. If the right chunk is ranked low:

• Add a Q&A pair with the exact phrasing.
• Pin the relevant source.
• Rewrite the chunk's description for clarity.

Plan limits