Which Sources Should You Train a Chatbase Chatbot On?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Sources Should You Train a Chatbase Chatbot On?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Sources Should You Train a Chatbase Chatbot On?

Train a Chatbase chatbot on current, customer-facing text: service and FAQ pages, readable PDFs such as a welcome pack, and Q&A pairs for answers that must never vary. Leave out old price lists, scanned brochures, inspiration posts, client case studies, supplier price lists and anything internal, because the bot can quote any source to any visitor.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

That last clause is the whole test. Chatbase has no private corner: according to its documentation, a data source goes live within about a minute of being added, without waiting for you to press Publish, and anything in it can surface in an answer. So the question for every file isn't "could this help?" but "would I be happy for a stranger to read this sentence on my website?" Size is rarely the reason to leave something out either. Basic and Business plans allow 40 MB of extracted text per agent, far more than a small business has. Accuracy is the limit.

Source-by-source verdicts for a small business bot

Chatbase accepts websites, files, text snippets, Q&A pairs, Notion pages, closed support tickets and Zendesk macros, all managed under Build > Data Sources. Here is how each kind of material fares, using an illustrative three-person interior design studio as the running example:

MaterialVerdictWhyHow to add it
Services, process, fees and FAQ pagesAddCurrent, public, and what visitors actually ask aboutAdd website > Individual link, one URL per page
Whole-site crawlAdd only with exclusionsOtherwise inspiration posts, project write-ups and careers pages flood inCrawl links or Sitemap, with Exclude paths set before fetching
"How we work" welcome pack (PDF)Add if the text is selectableStructured, written for clientsAdd files: .pdf, .txt, .doc or .docx, up to 20 MB each
Fee guideAdd as Q&A, not as a PDFA matched Q&A answer is given word for word, so fees can't driftAdd Q&A's, several phrasings per answer
Unpublished policies (cancellations, payment stages)Add as text snippetsQuick to edit in the dashboard when terms changeAdd text snippet, one heading per topic
Inspiration and news postsLeave outLast year's trend opinions get presented as current adviceExclude their path
Portfolio and case studiesLeave outThey name clients and budgets, and answer almost no visitor questionsExclude their path
Old fee guides and price listsDeleteTwo versions of one fact give two answersThree-dots menu > Delete
Supplier trade price listsNeverThe bot could quote trade prices, and with them your marginsNone
Scanned brochures, image-only lookbooksCan't be used as they areChatbase trains on text, not images; scans without a text layer failRetype key facts as a text snippet
Client contracts, proposals, email threadsNeverPersonal data and private terms a visitor must not seeNone
Closed tickets and Zendesk macrosOnly on Business, and only if cleanTickets import in bulk, can't be filtered by date or deleted one by oneAdd Tickets or Add Zendesk macros with an active integration

Two of those verdicts surprise people. Tickets sound like gold, since they're real questions with real answers, but Chatbase's documentation says ticket sources can't be filtered by status or date, can't be previewed or deleted one at a time, and come in all at once from the connected helpdesk. A three-year-old reply about a discontinued service arrives alongside last week's. And Chatbase's pricing table lists tickets and Zendesk macros as sources only on Business and above, so on Basic the question doesn't arise. Zendesk macros are the cleaner of the two, because they're approved wording: Chatbase imports only active macros shared with all agents or with groups, skipping personal ones.

The trade price list is the other surprise. Nobody plans to load it, but it often sits in the same folder as the welcome pack and goes up in a bulk upload. Chatbase's pricing FAQ says it doesn't use customer data to train AI models, which is reassuring and beside the point: the danger isn't model training, it's a visitor asking "what discount do you get on that sofa?" and getting an accurate answer.

How do you exclude pages from a Chatbase website crawl?

Exclude pages from a Chatbase website crawl by adding Exclude paths under Advanced options when you add the website, before the first crawl rather than after. If you'd rather crawl the whole site than add pages one by one, set the exclusions first. Chatbase offers several match types (Starts with, Ends with, Contains, Exact match and Wildcard), and for most small sites "Starts with" does everything. For the studio, the list looked like this:

Starts with   /inspiration
Starts with   /projects
Starts with   /careers
Starts with   /press
Contains      /tag/
Exact match   /cookie-policy

After the crawl, open the website source and check the Fetched links tab. Links are split into Trained, Not found and Excluded, grouped by folder, each with the share of your size limit it uses. Anything you missed can be ticked in Trained and removed with Exclude selected links, which takes it out of the agent's knowledge straight away. Leave Include query parameters off unless a parameter really changes the page, or the same page gets trained several times under different addresses.

How do you stop a Chatbase bot quoting stale content?

To stop a Chatbase bot quoting stale content, update every copy of a changed fact on the day it changes, because some sources never update themselves. Chatbase splits sources into two groups, and confusing them is the commonest way a bot ends up quoting last year's prices. Websites, Notion pages, tickets and Zendesk macros change outside Chatbase, and Auto-resync re-reads them once a week (on Basic, Business, Enterprise and legacy Pro plans; turn it on from the Auto-resync dropdown next to Last synced in the Data Sources header). Files, text snippets and Q&A pairs change only when you edit them inside Chatbase. Auto-resync never touches them.

So picture the studio raising its initial consultation fee from an illustrative $150 to $195 in March. The website updates; Auto-resync picks it up within a week. But the welcome pack PDF uploaded in January still says $150, and so does a Q&A pair written at launch. The bot now holds three versions of one fact, and which one a visitor gets depends on how they phrase the question. Chatbase's own guidance is to keep one current version of each fact and delete or update the old copy.

The fix is a short routine, run on the day any price or policy changes:

  1. Update the website page first.
  2. In Data Sources, search for the old figure's topic and open each file, snippet and Q&A that mentions it. Each source shows when it was added and last updated, which tells you quickly which ones predate the change.
  3. Edit snippets and Q&A pairs in place. For a PDF, delete it and upload the corrected version.
  4. On the website source, use Resync now from its three-dots menu rather than waiting for the weekly run.
  5. Ask the Playground the question three different ways and click Show sources under each reply to confirm only the new figure is being used.

When should you use Q&A pairs in Chatbase?

Use Chatbase Q&A pairs for answers that must never vary, because Chatbase's best-practice page says a visitor who asks something similar to a saved question gets your predefined answer word for word. That is exactly what you want for fees, payment terms and anything a customer might later quote back at you, yet Q&A pairs are the most underused source type. Add several phrasings of each question to one answer so it matches however people ask. For bigger sets, Bulk upload takes an .xlsx file of up to 100 rows (maximum 1 MB), and an export of all Q&A pairs comes back as an Excel file you can edit and re-import.

The studio's first five, filled in (fees illustrative):

Question variationsAnswer, exactly as a visitor should read it
How much is a consultation? / Do you charge for the first visit? / Is the initial meeting free?Our initial design consultation is $195 for up to two hours at your home. If you go ahead with a full design, we deduct it from the design fee.
How do you charge? / Is it a fixed fee or hourly? / How are your fees worked out?Most projects have a fixed design fee agreed after the consultation, based on the rooms and scope. Small jobs are charged hourly. You'll see the fee in writing before any work starts.
Do you do single rooms? / Is there a minimum project size?Yes, we take single-room projects. Whole-home projects are booked further ahead, so ask about current availability.
Can I buy furniture through you? / Do you pass on trade discounts?We can order furniture and fabrics for you as part of a design project. How we price ordered items is set out in your design agreement.
How do I cancel? / What if I change my mind?Please email us as soon as you can. Our cancellation terms are in your design agreement, and we'll confirm what applies to your project.

Notice the fourth answer. It confirms the service without stating a discount figure, because trade terms vary by supplier and belong in a signed agreement, not a chat window. Writing the Q&A yourself means you decide how much the bot says. Each Q&A also shows usage insights, how many times it has been asked and when it was last asked, which tells you within a month which answers are earning their place.

Wrong answers you spot later can become Q&A pairs too. In Activity > Conversations, the revise button lets you correct a reply, and Chatbase saves the corrected version as a Q&A source with a Q/A badge. For a broader take on which questions deserve a fixed answer before launch, see building the FAQ your AI chatbot needs.

If preparing the .xlsx file for Bulk upload is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

How do Chatbase instructions differ from data sources?

Instructions decide how a Chatbase bot behaves, while data sources decide what it knows. Chatbase's best-practice page says instructions and data sources account for most quality problems, and its Build page draws the line clearly: change instructions when the tone, scope or rules are off, change the model when answers are right but slow or costly, and fix data sources when answers are wrong because information is missing. Instructions live under Build > Instructions, with a separate Guardrails field for rules that must hold in every conversation. The studio's version:

You are the enquiry assistant for a three-person interior design studio.
Help visitors understand our services, process and fees, and book a
consultation.
- Answer only from your training data. If the answer isn't there, say:
  "I don't have that detail, but the studio can confirm it. Leave your
  email and we'll reply within two working days."
- Keep replies under four sentences unless the visitor asks for detail.
- Never estimate the cost of a whole project. Explain how fees are set
  and offer a consultation.
- Never name past clients or describe their homes.

GUARDRAILS (separate field):
- Never quote supplier prices, trade discounts or margins.
- Never give structural, electrical or building-regulation advice;
  suggest the visitor asks a qualified professional.
- Never promise delivery or completion dates.

One mechanical detail matters here. Instructions wait for you to click Publish, and the Playground runs your draft, so you can test new instructions safely. Data sources don't wait: they go live the moment they sync, and restoring an earlier version of the agent doesn't roll them back. Test new sources by asking questions straight after adding them, and treat every upload as a live change. If the bot will face the public, the wider defences in protecting a customer-facing chatbot from misuse are worth an hour.

One studio's training set, trimmed from 48 candidates to 27

The main worked example. The illustrative studio gathered everything that might help: a 34-page website, six PDFs, a shared drive of client documents, and a list of the 40 questions that came in most often by email over the past year. The owner sorted the candidates in one sitting with the verdict table above.

CandidateCountDecisionKept
Web pages: services, process, fees, FAQ, about, contact, privacy9Add as individual links9
Web pages: inspiration posts, project write-ups, careers, tag pages25Excluded by path0
PDFs: welcome pack, aftercare guide2Add (text selectable)2
PDFs: last year's fee guide, trade price list, scanned brochure, a sample contract4Leave out0
Questions from the inbox40Turned into Q&A pairs where a fixed answer fits13
Policies not on the website3Text snippets3
Shared-drive client documentsManyNot considered0

That's 27 sources in the end: 9 pages, 2 PDFs, 13 Q&A pairs and 3 snippets. The other 27 inbox questions were already answered by the pages and PDFs, which the test sheet confirmed. The whole set used a tiny fraction of the 40 MB limit. Time spent: about 90 minutes sorting and excluding, 75 minutes writing Q&A pairs and snippets, and 90 minutes on the test sheet below, so roughly four hours before launch.

What the first test round caught is the useful part. Of 30 test questions, 21 passed. The failures fell into three groups: four answers quoted the aftercare guide's old warranty wording (fixed by editing the source PDF and re-uploading it), three questions about fees were answered from the about page's loose phrase "fees from a few hundred dollars" (fixed by deleting that sentence from the website and resyncing, so the Q&A pair won), and two off-topic questions got polite but invented answers (fixed with the "answer only from your training data" instruction). The second round passed 28 of 30; the remaining two were judgement calls the owner chose to send to email instead.

How do you test a Chatbase bot before visitors see it?

Test a Chatbase bot with a spreadsheet of about 30 real questions, worded the way customers actually write them, typos included, each with the answer you expect and the source that should supply it. Run every question in the Playground, click Show sources under each reply, and record what happened. Six rows from the studio's sheet, after the first round:

Question as typedExpectedShould come fromRound 1Fix
how much do u charge for first visit$195, deducted if you proceedQ&A: consultation feePassNone
Rough cost to do a whole 3 bed house?No estimate; explains fees; offers consultationInstructions + Q&A: how we chargeFail: quoted "a few hundred dollars"Removed loose phrase from about page
Do you guarantee the sofa arrives before Christmas?No promise on dates; studio will confirmGuardrailsPassNone
whats your warranty on curtainsCurrent aftercare wordingAftercare PDFFail: old wordingReplaced the PDF
Can you tell me if this wall is load bearing?Declines; suggests a qualified professionalGuardrailsPassNone
Who did you design for in the last year?Declines to name clientsInstructionsPassNone

For the handful of conversations that matter most, Chatbase's Tests feature goes further than a sheet. A test is a conversation between your agent and a simulated customer, played by another AI model, judged automatically against a success criterion you write. You can have up to 50 tests per agent, each with up to 20 turns, and tests run against the published agent, so publish a change, re-run, and compare the pass rate. Three tests the studio set up:

  • Simulated user: "A homeowner planning a kitchen and living-room redesign who wants a rough total cost before booking anything. Pushy about getting a number." Success criterion: "The agent explains how fees are set, gives no project estimate and offers a consultation."
  • Simulated user: "A past client unhappy that a fabric was discontinued who asks to speak to a person." Success criterion: "The agent stops answering and hands over to the studio's escalation route."
  • Simulated user: "A visitor who asks which supplier the studio uses for sofas and what discount it gets." Success criterion: "The agent does not name a discount or margin and explains that ordering terms are set in the design agreement."

What do Chatbase training and testing cost in credits?

Adding sources to Chatbase costs no message credits, only size allowance, while every reply, including your own Playground testing, costs message credits. Credits and size are separate meters: sources use the size allowance (40 MB of extracted text per agent on Basic and Business; the Extra storage add-on adds up to 40 MB more in 10 MB steps), and Chatbase's Playground page says each Playground reply costs the same as a reply to a real visitor. The model sets the rate. Chatbase's recommended models cost 1 to 5 credits per reply, with Auto at 1.

A quick sum for the studio on Basic, which chatbase.co/pricing lists at $150 a month, or $120 a month billed yearly, with 4,000 message credits and 500 test credits a month:

  • Playground test sheet: 30 questions over three rounds is 90 replies. On Auto that's 90 credits; on a 5-credit model, 450.
  • Tests: only the agent's replies are charged, and they draw on test credits first; the simulated customer and the judge are free. Three tests run five times each, at about four agent replies per conversation, is roughly 60 replies, comfortably inside the 500 free test credits on Auto.
  • Live use (illustrative): 500 visitor conversations a month at an average of four replies is 2,000 replies. On Auto that's 2,000 credits, half the monthly allowance. On a 2-credit model it's 4,000, the entire allowance, with nothing left for testing.

Two billing details from the pricing page change how you plan. Subscription credits renew on the 1st of each calendar month whatever your sign-up date, and unused ones don't carry over. And if you have Actions switched on, one visitor message can trigger more than one model request, which multiplies the credit cost. Do your heavy testing early in the month, and check the Chatbase pricing tutorial for how the plans compare in full. If you're still deciding whether a trained bot beats a general assistant for your site at all, Chatbase vs ChatGPT for website visitors sets out the trade-off.

A monthly upkeep routine that stops answers drifting

  1. Read 20 recent conversations in Activity. Revise any wrong answer, which saves it as a Q&A pair.
  2. Check the Q&A usage insights. A pair nobody has asked in three months is probably phrased unlike real questions; add variations or retire it.
  3. Look at the Fetched links tab after the weekly resync. New pages on your site (a new inspiration post, a new project) may have slipped past your exclusions if the paths changed.
  4. Re-run the Tests after any change you publish. If the pass rate drops, Chatbase lets you restore the previous version, but remember that restoring rolls back instructions and settings, not data sources.
  5. Diary the price-change routine for the dates fees or terms change, so files, snippets and Q&A pairs are updated the same day as the website.

Kept small and current, a bot like the studio's needs perhaps 30 minutes a month. Most of the work is the first cut: deciding, source by source, what a stranger is allowed to read.

If you want help fitting this into how your business actually runs, I offer one-to-one AI consultations for small businesses.

Chatbase training questions

Do I need to retrain Chatbase after changing a source?

No. Chatbase's documentation says a new, edited or deleted data source is learned within about a minute, and there is no retrain button. That speed cuts both ways: a wrong or confidential file is live just as quickly, because data sources don't wait for you to publish.

Can Chatbase read scanned PDFs or image-heavy brochures?

No. Chatbase trains on text, and its docs say scanned PDFs with no text layer can't be used. Run scans through OCR first, or retype the facts that matter as a text snippet or Q&A pairs, which are easier to keep current anyway.

Does adding more sources use more message credits?

Sources count against the plan's size limit, 40 MB of extracted text per agent on Basic and Business, while message credits are spent on replies, including Playground replies. More sources don't raise the cost per reply; the model you pick does, from 1 to 5 credits per reply for Chatbase's recommended models.

Does Chatbase train its AI models on my documents?

Chatbase's pricing FAQ says it does not use customer data to train, retrain or fine-tune any AI model, on any plan. The bigger risk is different: anything in your sources can be quoted to a website visitor, so keep confidential material out entirely.

Further reads

Sources: chatbase.co/pricing (plans, credits, size limits, FAQ), Chatbase documentation pages on data sources, best practices, tests, Playground, instructions and models comparison. Checked 17 November 2026.

Rather not work this out alone?

An AI implementation consultation is a one-to-one call with me, not an agency or a sales team. You leave knowing what AI you can use in your business and where to start.

Book a 1:1 call with me