Knowledge Base (FAQs) #

FAQs (Frequently Asked Questions) are one of the easiest and most effective ways to make your AI bot smarter. When a contact asks a question, the bot searches your FAQs to find the best answer and weaves it naturally into the conversation.

Think of FAQs as a cheat sheet for your bot. The better the cheat sheet, the better the bot’s answers.

Why FAQs Matter #

Without FAQs, your bot relies only on the general instructions and company info you provide. That works for basic conversations, but FAQs give the bot precise, verified answers to specific questions.

What well-maintained FAQs do for you:

• Your bot gives accurate, consistent answers every time
• The bot is far less likely to make up information
• You can cover detailed or specialized topics without overloading your bot instructions
• Your bot sounds more knowledgeable and trustworthy

How the Bot Finds the Right FAQ #

When a contact sends a message, the bot does not just search for matching words. The platform uses a hybrid smart search system that understands meaning, not just exact wording.

This meaning-based search only kicks in once an Agent has 3 or more FAQs linked. When you have that many, the bot can match by meaning - for example, a contact asking “How much does it cost?” will find your FAQ about “pricing” even if the word “cost” never appears in it.

If an Agent has only 1 or 2 FAQs, the bot falls back to keyword matching instead. In that case the contact’s message and your FAQ need to share actual words - the bot cannot match purely by meaning. So once you have a few FAQs in place, the smarter meaning-based search becomes available.

Here is how it works behind the scenes:

1. Vector search (meaning-based) - The system converts the contact’s message into a mathematical representation of its meaning, then finds FAQs with similar meaning, regardless of the exact words used
2. BM25 keyword matching - A proven keyword-ranking algorithm that catches exact word matches and ensures nothing obvious is missed

These two approaches work together as a hybrid search to find the top 7 most relevant FAQs for each conversation. Only these relevant FAQs are shared with the bot, keeping its responses focused and on-topic.

You do not need to understand the technical details. Just write clear, natural-language FAQs and the system handles the rest.

Where to Manage FAQs #

All your FAQs live in one place: the Knowledge Base page, under AI Studio in the main sidebar. It combines your FAQ library with the knowledge base sources they’re generated from.

The page has three parts:

• Knowledge base sources — the websites and documents you’ve imported. Each source shows its page count and how many FAQs it generated, with Refresh to re-scan it for changes and an Agent button to open the picker and link (or unlink) that source to an Agent.
• Knowledge groups — bundle FAQs together so you can apply the whole bundle to an Agent in one click, instead of linking FAQs one at a time.
• Frequently asked questions — your full FAQ library, searchable and filterable, with Find contradictions, Deduplicate, and Import / Export actions across the top. Every row carries a number so you can refer to a FAQ (“number 7 needs a better answer”) when reviewing the list with a colleague or client. The numbers follow the order on screen and keep counting across pages, so with the default Newest first sort a new FAQ shifts everything down by one — switch the sort to Oldest first or Alphabetical when you want numbers that stay put.

Adding a Single FAQ #

1. Click AI Studio → Knowledge Base in the main sidebar.
2. Click New FAQ at the top-right. A pop-up opens.

3. Type the Question in the first field.
4. Type the Answer in the second field.

5. Click Add FAQ at the bottom of the pop-up.

You can also add or review an Agent’s FAQs from inside its own editor: open the Agent and click its FAQs & Knowledge tab. It’s the same shared library — a FAQ you add there is the same FAQ that shows up on the Knowledge Base page, just filtered to the ones linked to that Agent.

What to Fill In #

There’s no Category field or per-FAQ Active toggle in the New FAQ pop-up — a FAQ is either in your library or it isn’t. To organize a growing library, use Knowledge groups (below) to bundle related FAQs together instead.

Tips for Writing Great FAQs #

• Write the question the way a customer would ask it. Use natural language. Instead of “Refund Policy,” write “Can I get a refund if I’m not satisfied?”
• Keep answers complete but concise. Include all the facts the bot needs. It will rephrase the answer naturally in conversation.
• One topic per FAQ. Do not combine multiple questions into one entry. Create separate FAQs for each question.
• Include specific details. Prices, timelines, dimensions, requirements - the bot cannot guess these.

Example FAQs #

Q: What are your shipping times?
A: Standard shipping takes 3-5 business days within the US. Express
   shipping (next day) is available for $12.99. International shipping
   takes 7-14 business days depending on the destination.

Q: Do you offer a free trial?
A: Yes, we offer a 14-day free trial with full access to all features.
   No credit card required. You can upgrade to a paid plan at any time.

Q: What happens if I cancel my subscription?
A: You can cancel at any time from your account settings. Your access
   continues until the end of your billing period. We do not offer
   prorated refunds for partial months.

Organizing FAQs with Knowledge Groups #

The FAQ library doesn’t have a category or folder field — with a hybrid search that matches by meaning (see above), the bot doesn’t need one, and neither do you for finding a FAQ later (the search box on the Knowledge Base page searches both questions and answers).

Knowledge groups are the real organizing tool: a faster way to link many FAQs to an Agent at once. Bundle related FAQs into a group on the Knowledge Base page (+ New group), then link the whole group to an Agent in one click instead of linking each FAQ individually. Good groups to build: everything about pricing, everything about shipping, everything specific to one promotion.

When you add FAQs to a group, tick a FAQ to add it and untick it to take it back out — the checkmarks update as you click. To fill a group fast, type a search first and click Add all shown to add every matching FAQ in one go (with no search it adds your whole library). A group’s number badge counts the FAQs that are actually still in your library, so deleting or de-duplicating FAQs updates the count automatically.

Global FAQs vs. Agent-Specific FAQs #

Every FAQ lives in one shared library, but a FAQ is only used by a bot if that FAQ is linked to the Agent answering the conversation. The “Global” label just means the FAQ isn’t tied to a single Agent so you can reuse it — it does not mean the bot applies it everywhere automatically.

To use a library FAQ, link it from the FAQs & Knowledge step on the Agent — or link a whole knowledge source at once. There is no automatic “override on conflict” between global and specific FAQs — the bot simply uses whichever FAQs are linked to the Agent it’s answering on.

Best practice: Keep general company information (business hours, general policies, contact details) as reusable library FAQs and link them to every Agent that needs them. Put details unique to one offer, product, or promotion in FAQs you link only to that Agent. This avoids duplicating FAQs across Agents.

FAQs & Knowledge on the Agent #

Step 4 of the agent editor is where an Agent’s answers come from. Open the Agent and click FAQs & Knowledge to link FAQs and knowledge sources without leaving the editor — it’s the same shared library, just scoped to this one Agent.

A search box sits above the Knowledge sources list: type part of a page’s address to narrow the website, its folders and its pages at once. While a search is active, Unlink all becomes Unlink matching and only detaches the pages on screen.

Structured FAQs vs. Legacy text: Agents carried over from an older bot may hold a free-text FAQ blob instead of individual questions. Those Agents show a Structured FAQs / Legacy text switch at the top of the step; Agents built since the structured system arrived are structured-only and never show it.

• Structured FAQs — the default view. Individual, searchable, reusable questions and answers, same as everywhere else in this guide.
• Legacy text — the old free-text block, sent to the AI as-is. You can still edit it here, but it isn’t a set of individual FAQs — it’s one block of text.

If the free-text block is still filled in and no structured FAQs are linked yet, a banner offers a Migrate now button. Migrating converts the free-text block into individual structured FAQs automatically and switches the step to the Structured FAQs view — a one-way upgrade, not something you need to do question by question. #

Auto-Generate FAQs from Your Website #

When you add website URLs (like sales pages or product pages) while building an Agent, the platform can automatically generate FAQs from the page content.

Every FAQ generated this way remembers which page it came from, and your AI is given that page address alongside the answer. So when a customer asks to see something the knowledge describes — a product category, your pricing, a policy page — the AI shares the real link to that exact page on your site. The AI is only ever allowed to share links that genuinely exist in its knowledge or instructions; it will never construct or guess a web address.

How to do it:

1. Create a new Agent, or edit an existing one
2. Add your website URLs in the Sales Pages section
3. The platform reads the page content
4. AI analyzes the content and generates relevant FAQs
5. The generated FAQs appear on the Knowledge Base page (and on the Agent’s FAQs & Knowledge tab) for review

During onboarding: If you use the guided onboarding wizard, the platform can auto-generate FAQs from your website as part of the setup process. The onboarding wizard scans your website, analyzes its content, and creates an initial set of FAQs for you automatically - giving your bot a head start before you even finish setup.

Always review auto-generated FAQs. The AI does a good job, but you should verify accuracy, add missing details, and remove anything irrelevant.

Auto-Generate FAQs from Documents (Knowledge Base Sources) #

If you have existing documents like product guides, policy documents, or training materials, you can upload them and let AI generate FAQs automatically. This is a fast way to give your bot detailed knowledge without writing every FAQ by hand.

Supported file types: PDF, Word (.docx and legacy .doc), plain text (.txt), Markdown (.md), CSV, and Excel (.xlsx and legacy .xls). Maximum file size is 50 MB per file.

How to do it:

1. Go to AI Studio → Knowledge Base.
2. Scroll to the Knowledge base sources section.
3. Click the green Add source button.
4. A pop-up titled Add knowledge source opens with two tabs: Website (to scrape a website — see URL sources below for how multi-page crawling works) and Document (to upload a file).

5. Click the Document tab.

6. Click Click to choose a document (or drag and drop a file onto the pop-up) to pick one document.
7. Click Add document. Wait for the button’s progress state (“Uploading…” then “Reading document…”) to finish — this can take a minute for larger files.
8. The generated FAQs appear in the FAQ list above with the document as their source.

To add more documents, repeat from step 3 — one source per modal opening.

Updating a document you already uploaded: the Refresh button only applies to website sources. To update a file, delete the old source and upload the new version of the file — the platform reads it the same way it read the first one.

Uploading a product catalogue or price list? Use a spreadsheet or CSV, with one product per row and a header row naming each column (product name, price, product URL, availability, and so on). Those files are read row by row: every product becomes its own knowledge entry with all of its columns kept exactly as you wrote them, so the bot can quote a price and share the right link. This also copes with real-world files: title rows above the header, a data sheet next to a notes sheet, and other messy layouts are detected and handled automatically. Two things to keep in mind: a sheet is read row by row up to 2,000 rows (split a bigger catalogue across several files), and the header row needs to sit within the first few rows with a plain, un-merged cell for each column name. A catalogue inside a PDF or Word document is read as prose instead, which is much less reliable for exact numbers and links, so spot-check a few products afterwards if you go that route.

URL Sources and Sitemap Crawling #

URL sources are not limited to a single page. When you add a URL like help.example.com, the platform reads the site’s sitemap (a list of all the pages on that website) and pulls in the pages it finds as part of that one source. Each page becomes its own entry under the source, but they’re managed together as one “domain.”

The platform crawls the site’s sitemap up to 100 pages per import or refresh. Very large sites are capped at this limit — if your site has more pages, add the most important sections as separate sources so each stays under the cap.

This makes external knowledge bases (your help docs, your product wiki, anything with a sitemap) ideal sources: keep one external doc up-to-date and the platform mirrors it into your FAQs. One exception worth knowing: a changelog page is usually a poor source, even though it looks information-rich — its entries describe what was true when each was written, so they go stale by design and can end up contradicting the pages that document your features today. The page picker flags changelogs with a Skip (suggested) label for exactly this reason.

Watch out for anything time-bound on the page you import. A sentence like “we’re fully booked until September”, “closed for the holidays” or “offer ends Friday” is read as a plain fact and becomes an FAQ that keeps saying it long after you’ve changed the page. Nothing re-checks it for you: an imported FAQ is a snapshot of the page on the day you imported it, and it stays that way until you refresh the source or edit the FAQ. This bites hardest on a booking or contact page, where a stale line about availability makes your bot turn people away who could have booked. Either keep those sentences off the pages you import, or make a habit of refreshing the source whenever you change them. If your bot ever says something about your availability that you don’t recognise, search your FAQ library for it — it is almost always sitting there in writing.

When you pick pages to import, each one carries the same worth-adding guidance as the domain-refresh picker: the most useful pages sort to the top, Recommended marks your core pages (homepage, pricing, products, FAQ and the like), and pages that rarely belong — legal text, login screens, search results, tag listings, changelogs — say so with a plain-language reason. A Select recommended button ticks the recommended set in one click, and you can always tick anything else yourself.

To see the pages the platform discovered, expand the source row in the Knowledge Base Sources list — every crawled URL is listed there.

Finding and cleaning up sources in bulk. The sources list has a search box that looks inside every site: type part of a page’s address or name (say, “changelog”) and the matching pages surface with a “3 of 227 pages match” count on their site. Every page and file row has a checkbox, and a site’s own checkbox selects all of its pages at once — tick what you want gone and click Delete selected; the confirmation states exactly how many pages and FAQs that removes, and the deletion runs in the background with an email when it finishes. While a search is active, a site’s checkbox selects only the pages shown, so searching and deleting can never quietly remove pages you weren’t looking at. You can also tick several whole websites and click Refresh selected to re-scan them all in one go — each site’s newly found pages are offered one site at a time, with the same worth-adding guidance as a single refresh.

Stopping an Import #

A large import runs in the background, page by page, and you can stop it at any point. While pages are still waiting, the domain’s row shows a Stop import button with the number still queued. Stopping drops those pages and costs nothing — a page that hasn’t been read yet has never been charged for. Pages already imported stay exactly as they are, and pages that were being read at the moment you pressed stop will finish (those are charged, and the confirmation tells you how many).

Stopped pages stay in the list marked Stopped so you can see what was dropped. Delete them, or import them again later if you change your mind.

You can also delete pages while an import is running — the trash icons stay available, and deleting a page that’s still waiting stops it from being imported at all.

If Your Own AI Key Stops Working Mid-Import #

This applies only if you’ve connected your own AI key (BYOK). Importing pages with your own key doesn’t use AI credits.

If your key stops working part-way through a large import — most often because the balance on your AI provider account has run out — the remaining pages are paused rather than finished. They’re marked Paused in the list, and nothing is charged while they wait. We do this because finishing them would mean switching to your AI credits, and that’s your decision to make, not something to discover afterwards on your balance.

You have two options:

1. Top up or fix your AI provider key, then press Resume on that domain. The remaining pages continue on your own key, at no credit cost.
2. Press Resume without fixing the key. The remaining pages are read using your AI credits instead. The confirmation says so before anything runs.

Either way, pages already imported are kept.

Before You Refresh: Which Copy Is the Master? #

A knowledge base that came from your website can be kept up to date in two ways, and it’s worth deciding which one you’re using before you click Refresh:

• Your website is the master. You keep your site up to date, and whenever it changes you refresh the source here. The platform rebuilds that page’s FAQs to match the site, and after a domain refresh it consolidates overlapping answers so your bot doesn’t carry near-duplicates. This is the right mode if your site is where the truth lives.
• The knowledge base here is the master. You imported or generated a starting set once, then invested time editing, adding and organizing FAQs by hand. In this mode, make your changes directly in the FAQ library — and treat Refresh as a “start over from the website” button, because a refresh rebuilds generated FAQs from the live page, and the cleanup pass that follows a domain refresh can merge or reword FAQs you wrote or imported yourself if they cover the same ground as the site.

Mixing the two — hand-curating FAQs and then re-crawling the same content — is what surprises people: the refresh does its job (making the knowledge match the site and consolidating duplicates), but that job includes reworking hand-written FAQs that overlap with page content. The refresh buttons ask you to confirm before any of this runs, so nothing starts by accident. If you’ve spent real time curating, export your FAQ library first (see Importing and Exporting FAQs) — it’s a one-click backup.

Refresh Domain (Multi-Page Re-Crawl) #

For URL sources with multiple pages, the Refresh action operates at the domain level. Click Refresh Domain next to the source and the platform will:

1. Re-read the sitemap and discover new pages that didn’t exist last time (you’ll see them appear in the list).
2. Update FAQs for pages whose content changed.
3. Drop or merge FAQs for content that no longer exists or has become irrelevant.
4. Generate new FAQs for new pages and new sections.

If new pages are found, a New pages found window opens so you can pick which ones to import — nothing is pre-selected, since each imported page costs credits. When there are many new pages (for example a site with lots of language versions), they’re grouped into collapsible sections by site area (like /es/ or /blog/): click a section to expand its pages, or use the section’s checkbox to select or deselect the whole group at once. Confirming with nothing selected just re-scrapes the pages you already imported.

Each page now tells you whether it’s worth adding. The list is sorted with the most useful pages on top, and pages that usually don’t belong in a knowledge base carry a Skip (suggested) label with the reason in plain words — a changelog is marked “History page — its entries go stale and your current pages already cover the features”, and legal text, login screens, search results and tag listings get similar notes. Pages like your homepage, pricing, products and FAQ show as Recommended, and a Select recommended button ticks exactly those in one click (it adds to whatever you’ve already ticked, never replaces it). The labels are advice, not rules: every page stays one click away, whatever its label says.

As you tick pages, the window shows the estimated credit cost of your selection and your current balance right above the import button. If the selection is large, or would use more than half your remaining credits, you’ll get a confirmation step before anything starts.

Watch out for language versions. If your site publishes a copy of every page in each language, selecting everything imports the same content many times over — a 200-page site in 20 languages is 4,000 pages, and you’re charged for each one. Import the language you actually need. The picker helps here: a page whose address is the same as one you already have apart from its language part is marked Optional — language copy of a page you already have, and left out of Select recommended. On a first import the original stays recommended and its translations don’t, so one click gets you one language rather than twenty.

New pages join the agents that already use the whole site. If an agent is linked to every page of a domain, pages found by a later refresh are added to that agent automatically — you don’t have to re-link the domain after each refresh. Agents linked to only some pages of the domain are left alone, since picking a few pages isn’t the same as asking for the site; link the domain to that agent once and it will keep up from then on.

While the refresh runs, the domain shows a Refreshing 40 of 249 counter next to its name that ticks up as pages are re-read, then changes to Tidying up duplicates… for the last step. A big site takes a while, and the Refresh button stays greyed out until it’s done — you can leave the page and come back, the counter picks up where it left off. You’ll also get an email when the whole refresh finishes.

Stopping a refresh you didn’t mean to start. While a site is being refreshed, its Refresh button becomes Cancel. Click it and the refresh stops where it is: pages already re-read keep their fresh content, pages still waiting are left as they were, and no completion email is sent. Once a refresh reaches its last step (tidying up duplicates) it can no longer be stopped — Cancel then tells you it’s already finishing up, because that final pass has to complete to leave your FAQs in one piece. A site that’s being refreshed can’t be selected or deleted in the meantime; finish or cancel the refresh first.

Contradictions are checked automatically after a refresh. When a refresh finishes, the platform re-reads that site’s pages against the rest of your knowledge base and opens the Find Contradictions review for anything that now disagrees. That matters because a page you refreshed may have changed its mind about a price or a limit while another page still states the old one. You no longer have to remember to run the check yourself after a refresh.

You’ll often see the FAQ count climb during the fetch and then settle back down — that’s the platform automatically merging FAQs that cover the same ground at the end of the run. This cleanup looks across your whole FAQ library, not just the refreshed site, so it can also fold in hand-written or imported FAQs that say the same thing as a page (see Before You Refresh above). Fewer, consolidated FAQs is the goal — near-duplicate answers make the bot’s search worse, not better.

Refreshing a Source (One-Click Update) #

When your website or document changes, click the Refresh icon next to that source in the Knowledge Base Sources list. The platform:

1. Re-scrapes the URL (or re-reads the file).
2. Diffs the new content against the last snapshot, section by section.
3. Regenerates the FAQs for the parts of the page that changed, together with anything that was generated alongside them (plus the cross-section synthesis pass on the whole source).

A single-page refresh rebuilds that page’s generated FAQs but does not run the library-wide duplicate cleanup — that only follows a domain refresh (or you can run Deduplicate yourself, below). You’ll be asked to confirm before the refresh starts.

You do not need to re-link the source to any Agent after a refresh. When you link a knowledge base to an Agent (with the Link… button), it stores a reference to the source itself — not to each individual FAQ. So after a refresh, every Agent that already had this source linked automatically uses the new FAQs the next time the bot answers a message. One Refresh click is enough.

You’ll see a summary toast after the refresh: how many sections were added / changed / removed and how many FAQs were updated. A refresh that finds zero content changes is free (0 credits); otherwise you’re charged for the regenerated sections only.

Deduplicate Across Sources #

A single-page refresh doesn’t merge duplicates at all, and if you have multiple sources (e.g. a help-docs domain and a changelog domain and an uploaded PDF), the same FAQ may end up generated by more than one of them.

Click the Deduplicate button at the top of the Knowledge Base Sources card whenever you want to clean this up. The platform analyses every active FAQ across every source, merges or removes duplicates, and shows you the result (“No duplicates found” or “X duplicates merged”). Safe to run any time — it only touches FAQs that overlap.

Find Contradictions #

Duplicates aren’t the only problem a large knowledge base develops — sometimes two FAQs give conflicting answers to the same question (an old blog page says one price, your help docs say another). The bot can only be as consistent as its knowledge, so conflicting answers mean customers can get told the wrong thing.

Click Find contradictions on the Knowledge Base page to scan for this. The audit runs in the background (a progress chip shows clusters checked and conflicts found) and sorts what it finds into three groups:

• Auto-resolved — when one answer comes from a clearly more trustworthy place (your help docs or a hand-written FAQ) and the other from a marketing, changelog or legal page, the trusted answer wins automatically. A changelog counts as history rather than current truth here: an old announcement loses to the page that documents the subject today. Two versions of the same help article count too — the newer version simply wins. And when two equally trustworthy pages disagree, the page whose address is dedicated to the disputed subject wins, automatically when the audit’s reviewer independently reached the same answer. The losing FAQ is held out of the bot’s knowledge.
• Recommended — the audit proposes which answer is correct; accept them one by one or hit Accept all.
• Needs review — genuinely ambiguous ones where you pick which answer to keep.

It only flags disagreements about your own business. Two pages quoting a different price for someone else’s product, an article saying how often it gets updated, or two guides that each unlock with a different keyword aren’t contradictions — they’re separate facts that happen to read alike, and the audit leaves them alone. What it looks for is your own pricing, plans, limits, features and policies being described two different ways.

Re-running the audit also tidies up after itself: anything on the list that no longer disagrees — because you edited a page, deleted a source, or re-imported it — is closed automatically, and any FAQ that was being held out because of it goes back into your bot’s knowledge.

Click Review on the result chip to open the review panel. Keeping an answer holds the conflicting FAQs out of the bot’s knowledge (they’re not deleted — resolving a conflict the other way restores them); Dismiss tells the audit the two answers don’t actually conflict and both stay live.

This also runs automatically when you import. Whenever you add a new website page, document, or run a bulk import, the platform checks the new FAQs against your existing knowledge. New FAQs that contradict what your bot already knows are held back until you review them — a freshly imported page can never silently overwrite the answers your bot gives.

Linking a whole website or section at once #

When you click Link sources on an Agent’s FAQs & Knowledge tab, the picker shows your sources as a tree: each website at the top level, the sections of that site as folders inside it, and the individual pages inside those. Every level has its own checkbox — tick the website to select every imported page from that site in one click, tick a folder to select everything in that section, or expand a folder and tick single pages. A dash in a checkbox means only some of the pages inside are selected. Uploaded documents are grouped under Files with their own select-all checkbox. Nothing changes until you click Save selection.

Removing many linked sources from an Agent #

On the FAQs & Knowledge tab of an Agent, the Linked Knowledge Bases card groups linked sources by domain (or “Files” for uploaded documents). Each domain row shows the page count and a Remove all from Agent button — one click unlinks every page from that domain in one go. Click a domain header to expand and see the individual pages if you only want to remove some of them; each row has its own Remove button.

You can also pick and remove pages from anywhere in the list at once: every page row has a checkbox, and each website or folder header has a select-all that ticks everything inside it (a dash means only some of it is ticked). Tick the pages that should go — from one site or several — and a bar appears at the bottom of the screen with an Unlink button that detaches the whole selection after a single confirmation.

If you removed sources from your knowledge base previously and an Agent is still showing a “Deleted sources” group, that’s a leftover reference to a source that no longer exists. Click Remove all on that group to clean it up — the platform now also clears those references automatically the next time you delete a source.

Deleting a source keeps its FAQs by default. Deleting a knowledge base source removes the source itself but, by default, keeps the FAQs it generated (so anything you hand-edited or reused isn’t lost) — they stay in your library and in any linked Agents. Choose the “also delete its FAQs” option if you want the generated FAQs removed and unlinked from everywhere.

What happens behind the scenes:

• The AI analyzes the full content of each document
• It identifies key topics, facts, and details that customers are likely to ask about
• It generates natural-sounding questions and accurate answers
• Duplicate FAQs (ones that overlap with your existing FAQs) are automatically filtered out
• The process uses AI credits, and the cost depends on how much content there is. Knowledge-base generation costs 1 credit per content chunk, or 0.25 on accounts with the Max/Mini rate. This one follows your account, not the AI Quality setting on the Agent. The platform splits each page or document into chunks, so longer sources cost more, and the estimate shown before you confirm already reflects this.
• If you’ve connected your own Anthropic key (BYOK), knowledge-base generation runs on your key instead — the work is billed directly to your Anthropic account and no credits are charged. The exception is accounts on the Max/Mini rate, which pay the 0.25-credit chunk rate regardless of BYOK. See AI Model & BYOK for the full list of what your key covers.

Tips for best results:

• Upload documents that contain factual, specific information (pricing sheets, product specs, policies). These produce the best FAQs.
• Review every generated FAQ before enabling it. Remove any that are irrelevant or inaccurate.
• Documents with clear structure (headings, bullet points, tables) produce better results than unformatted text.
• Each upload adds one document as its own source. To cover several topics, add the documents one at a time — each becomes a separate source you can refresh or remove on its own.

Importing and Exporting FAQs #

Both live behind one button: click Import / Export at the top of the Frequently Asked Questions list on the Knowledge Base page. It opens a single pop-up with an Export section and an Import section stacked in it — there’s no separate Import button and Export button.

Importing #

If you already have FAQs (from a spreadsheet, help center, or another tool):

1. Prepare a CSV with question, answer, and optionally category columns (a JSON file with the same fields also works)
2. Go to AI Studio → Knowledge Base and click Import / Export
3. In the Import section, click Choose a file… and pick your CSV or JSON file
4. Optional: if you have Knowledge Groups, an Import into Knowledge Group picker appears — choose a group and the imported FAQs are added to that group as well as your library. Leave it on Main library (no group) to import ungrouped. (Importing one file per knowledge set and picking its group each time is the fastest way to build out grouped knowledge.)

Duplicate questions and blank rows are skipped automatically, and the pop-up reports how many FAQs were imported vs. skipped. Skipped duplicates are not added to the selected group — they stay exactly where they already are.

Exporting #

To back up or share your FAQs:

1. Go to AI Studio → Knowledge Base and click Import / Export
2. In the Export section, click Export CSV or Export JSON to download your whole FAQ library (question, answer, category)

Bulk editing FAQs via Import/Export #

When you have many FAQs to fix in one pass — rewording, adjusting tone, or adding a consistent disclaimer across every answer — the fastest workflow is to export the FAQ list to a text/CSV file, edit it externally, and re-import.

• Open the exported file in your favourite editor (spreadsheet app, text editor) or paste the contents into an AI assistant and ask it to do the rewrite for you.
• This is dramatically faster than clicking through dozens of individual edit forms one by one.
• Re-importing never changes existing FAQs. A row whose question matches an existing FAQ is skipped (it does not update the answer); only rows with new questions are added. Importing also never deletes anything.
• So the bulk-edit workflow is: export → edit the file → select and delete the old FAQs in the list (tick them and use the bulk delete) → import your edited file. Keep the original export file untouched as a backup before you start, just in case you need to roll back.
• If you only reword questions, a plain re-import works — the reworded rows count as new questions — but the old versions stay behind until you delete them.

Tracking FAQ Performance #

The app shows you how your FAQs are performing:

• Usage Count - How many times each FAQ has been used by the bot, shown on each FAQ row once it’s been used at least once
• Sort the FAQ list by Most used (from the sort dropdown) to see your busiest FAQs first

Use these numbers to:

• Identify your most valuable FAQs
• Find FAQs that are never used (consider rewording or removing them)
• Spot gaps where customers ask questions you do not have FAQs for (check your live conversations for patterns and add new FAQs)

Best Practices #

1. Start with 10-20 FAQs. Cover the most common questions first. You can always add more as you learn what customers are asking.

2. Review live conversations weekly. Look for questions the bot struggled with and create new FAQs to fill the gaps.

3. Keep FAQs current. If your pricing or policies change, update the FAQ immediately. An outdated FAQ is worse than no FAQ at all.

4. Do not duplicate your bot instructions. FAQs are for specific facts and details. General behavior and personality belong in the AI Instructions.

5. Write naturally. Write FAQs the way a customer would ask and the way you would answer in a real conversation.

6. Test after adding new FAQs. Use the Agent’s Try Out tab to verify that the bot correctly uses your new FAQs when relevant questions come up.

Common questions about the Knowledge Base #

I linked a knowledge group to my Agent — do I still see the individual FAQs on the Knowledge Base page? Yes. Groups are just a shortcut for linking many FAQs at once; every FAQ still lives in, and is editable from, the main FAQ library.

A FAQ I edited doesn’t seem to be used by the bot — why? Check that it’s actually linked to the Agent answering that conversation (or that the Agent is set to use your whole knowledge base). An unlinked FAQ, however well-written, is invisible to that Agent’s bot. (The one exception: a FAQ held out by Find Contradictions stays linked but is excluded from the bot’s knowledge until you resolve the conflict.)

Do I need 3+ FAQs before the smart search is worth setting up? No — even 1-2 FAQs help immediately via keyword matching. The meaning-based search is a bonus that kicks in automatically once you cross 3, no extra setup required.

Is there a limit on how many FAQs my knowledge base can hold? No. There is no cap on the number of entries, and no storage quota — the largest accounts on the platform run well over a hundred thousand entries with no trouble. The limits that do exist are per-import, not per-library: 50 MB per uploaded file, 2,000 rows per spreadsheet, and 100 pages per website import or refresh.

So a bigger knowledge base always makes the bot smarter? No, and this is the part worth internalising. Your Agent doesn’t read the whole library on every message — it searches it and hands the AI only the entries that match the question. Adding near-duplicate or contradictory entries makes that search worse, which is why Deduplicate Across Sources and Find Contradictions exist. Fewer, cleaner entries beat more entries every time.

Is there a limit on how much I can write in the Agent’s instructions? Yes, and that one is a real cap: it’s the Context budget on the Agent’s Bot Settings tab, and it varies by plan. See Setting Up Your AI Bot.

Next Steps #

• AI Agents — where FAQs plug into the rest of an Agent’s setup.
• Setting Up Your AI Bot — the full guide to writing instructions and playbooks.
• AI Agents — the main page of the AI Studio group Knowledge Base lives in, alongside Templates, Media Library, Custom Functions, and MCP Servers.