Training is the single most important step in building a useful chatbot. A well-trained chatbot gives accurate, helpful answers that reflect your business. A poorly trained one frustrates visitors and damages trust. This guide covers every training method available in Agentkit, when to use each one, and how to build a knowledge base that actually works.
The four source types
Agentkit supports four ways to feed your chatbot with knowledge. Each serves a different purpose, and the strongest chatbots use a combination of all four.
| Source type | What it is | Best for | Setup effort |
|---|---|---|---|
| Website | Crawl your site pages and index the content | Broad knowledge base from existing content | Low |
| Documents | Upload PDF, DOCX, or TXT files | Product specs, guides, internal docs | Low |
| Q&A pairs | Write specific question-answer pairs | Controlling exact responses to key questions | Medium |
| Text snippets | Paste raw text directly | Policy text, product details, announcements | Low |
Think of website crawling as the foundation, documents as supplementary depth, Q&A pairs as precision tools for critical answers, and text snippets as quick additions for specific content.
Website training: crawling your site
Website crawling is where most chatbots start. You give Agentkit a URL, and the crawler visits that page and follows links to index your site content.
How it works
- You enter a starting URL (usually your homepage or a specific section).
- The crawler visits that page and extracts the text content.
- It follows internal links to discover additional pages.
- Each page's content is processed, chunked, and converted into vector embeddings.
- When a visitor asks a question, the chatbot searches these embeddings to find the most relevant content.
The crawler respects your site's robots.txt rules and processes pages at a reasonable rate to avoid overloading your server.
Adding a website source
- In the Agentkit dashboard, go to your chatbot's Sources.
- Click Add Source and select Website.
- Enter your starting URL.
- Choose the crawl scope:
- Single page: Only the URL you enter.
- Subdirectory: The URL and all pages under that path (e.g.,
yoursite.com/docs/crawls everything in/docs/). - Full site: All pages the crawler can reach from the starting URL.
- Start the crawl.
The crawl runs in the background. Depending on the size of your site, it takes anywhere from a few seconds (single page) to several minutes (hundreds of pages).
For detailed options and settings, see the website sources documentation.
Handling JavaScript-rendered sites
Many modern websites use JavaScript frameworks (React, Vue, Next.js) that render content client-side. A basic crawler that only reads the initial HTML will miss this content.
Agentkit handles this with Firecrawl-powered rendering. When you add a website source, JavaScript-heavy pages are rendered in a headless browser before the content is extracted. You do not need to do anything special to enable this. If the crawler detects that a page needs JavaScript rendering, it handles it automatically.
What makes a good website source
Not all pages on your site are equally useful for training. Focus on pages that:
- Answer questions visitors actually ask (FAQ, help center, knowledge base)
- Describe your products or services in detail
- Explain policies (shipping, returns, pricing, terms)
- Provide how-to information relevant to your customers
What to exclude
Some content actively hurts chatbot quality. Avoid training on:
| Content to exclude | Why |
|---|---|
| Login/authentication pages | No useful information, may confuse the chatbot |
| Admin or dashboard pages | Internal content not relevant to visitors |
| Checkout pages | Transaction-specific, not informational |
| Duplicate content (print versions, AMP pages) | Creates noise and redundancy |
| Very thin pages (tag archives, empty categories) | Low information density |
| User-generated content (comments, forums) | Quality varies wildly, may contain incorrect information |
| Outdated blog posts with stale information | Can lead to wrong answers |
You can manage which URLs are included by using the subdirectory scope or by removing specific pages from your source list after crawling.
Document training: uploading files
Some of your best content may live in files rather than on your website. Product specification sheets, employee handbooks, training manuals, research reports, detailed guides. Document upload lets you add these directly.
Supported file types
| File type | Extensions | Notes |
|---|---|---|
| Most common format for documentation | ||
| Word | .docx | Good for internal documents |
| Text | .txt | Plain text files |
How to upload documents
- In your chatbot's Sources, click Add Source and select Documents.
- Drag and drop files or click to browse.
- Upload your files.
- The system processes each document, extracts text, and creates embeddings.
Processing time depends on document length. Most files complete within a minute.
For complete details on supported formats and limits, see the document sources documentation.
Tips for document sources
Quality over quantity. One well-written product guide provides more value than ten poorly formatted documents. The chatbot's answers are only as good as the source material.
Keep documents current. If you upload a product catalog from last year, the chatbot may reference discontinued products or old pricing. Replace outdated documents when new versions are available.
Structure matters. Documents with clear headings, sections, and organized content produce better results than walls of unformatted text. The chunking algorithm uses structural cues to create meaningful segments.
Check for scanned PDFs. If your PDF is a scanned image (no selectable text), the text extraction will not work. Convert scanned documents to text-based PDFs using OCR software first.
Q&A pairs: precision answers for key questions
Q&A pairs are your most precise training tool. When you create a Q&A pair, you define both the question and the exact answer. The chatbot uses these as high-priority knowledge, meaning a well-matched Q&A pair takes precedence over crawled or document content.
When to use Q&A pairs
Q&A pairs are ideal for:
- High-stakes answers where accuracy is critical (pricing, legal, medical disclaimers)
- Frequently asked questions where you want a consistent, carefully worded response
- Topics not covered on your website (internal knowledge, context that lives in your team's heads)
- Correcting chatbot mistakes (if the chatbot gets something wrong from crawled content, a Q&A pair overrides it)
- Time-sensitive information (current promotions, seasonal hours, temporary policy changes)
Writing effective Q&A pairs
A good Q&A pair has a question that matches how real people ask, and an answer that is complete, specific, and actionable.
Weak Q&A pair:
- Question: "Pricing?"
- Answer: "Check our pricing page."
Strong Q&A pair:
- Question: "How much does your service cost?"
- Answer: "We offer four plans. The Free plan is $0 with 50 messages per month and 1 chatbot. The Hobby plan is $29.99/month with 2,000 messages and API access. The Standard plan is $119.99/month with 12,000 messages, 2 chatbots, and 3 team seats. The Pro plan is $399.99/month with 40,000 messages, 3 chatbots, and 5 team seats. Annual billing saves about 20%. See full details at /pricing."
The strong version anticipates what the visitor actually wants to know and provides a complete answer.
Q&A pair best practices
| Practice | Explanation |
|---|---|
| Write questions in natural language | Use "How do I return an item?" not "Return procedure" |
| Include question variations | People phrase things differently. Add multiple Q&A pairs for important topics with different phrasing. |
| Keep answers concise but complete | Answer the question fully without unnecessary padding |
| Include next steps | If relevant, tell the visitor what to do next |
| Update regularly | Review and update Q&A pairs when information changes |
How many Q&A pairs do you need
There is no magic number. Start with 10-20 pairs covering your most common questions. Monitor conversations after launch to see what visitors ask, then add more pairs for topics where the chatbot's crawled-content answers are not sufficient.
Text snippets: quick additions
Text snippets let you paste raw text directly into your chatbot's knowledge base. No URLs, no files, just text.
When to use text snippets
- Product specifications that are not on your website yet
- Policy text you want the chatbot to reference verbatim
- Announcements (holiday hours, temporary closures, product launches)
- Brand information (mission statement, company background)
- Content from other sources (email templates, support scripts)
Adding a text snippet
- In your chatbot's Sources, click Add Source and select Text.
- Give the snippet a descriptive title (e.g., "2026 Holiday Shipping Deadlines").
- Paste your text content.
- Save.
Text snippets are processed immediately. They are useful for quick updates that do not warrant a full document upload or website recrawl.
The training workflow: playground testing
Training is not a one-time event. It is an iterative process. The Agentkit playground is your testing environment for refining the chatbot before and after it goes live.
Step-by-step training workflow
- Add initial sources. Start with your website URL and the most important Q&A pairs.
- Test in the playground. Open the playground and ask questions your visitors would ask.
- Evaluate responses. Are the answers accurate? Complete? In the right tone?
- Identify gaps. Note questions where the chatbot gives weak or incorrect answers.
- Add more training data. Upload documents, add Q&A pairs, or add text snippets to fill gaps.
- Test again. Verify the improvements.
- Repeat. Continue until the chatbot handles the majority of expected questions well.
What to test
Use this checklist to systematically evaluate your chatbot's training:
| Test category | Example questions | What to check |
|---|---|---|
| Core product/service | "What do you offer?" "How does X work?" | Accuracy, completeness |
| Pricing | "How much does it cost?" "Do you have a free plan?" | Correct numbers, current plans |
| Policies | "What's your return policy?" "How long does shipping take?" | Specific details, not vague answers |
| Edge cases | "Can I get a refund after 60 days?" "Do you ship to Antarctica?" | Graceful handling of unusual questions |
| Off-topic | "What's the weather?" "Tell me a joke." | Polite redirection to relevant topics |
| Competitor comparison | "How are you different from [competitor]?" | Factual, neutral response |
| Multi-turn conversations | Ask a follow-up question that depends on the previous answer | Context retention |
Keeping your chatbot up to date
Your chatbot's knowledge base needs maintenance. Products change, policies update, and new content gets published.
Update methods by plan
| Plan | Auto-retrain | Manual retrain | Recommended approach |
|---|---|---|---|
| Free ($0) | No | Yes | Retrain monthly or when content changes |
| Hobby ($29.99) | No | Yes | Retrain biweekly or when content changes |
| Standard ($119.99) | Yes | Yes | Auto-retrain handles routine updates; manual for urgent changes |
| Pro ($399.99) | Yes | Yes | Auto-retrain handles routine updates; manual for urgent changes |
Auto-retrain (available on Standard and Pro plans) periodically re-crawls your website sources and updates the knowledge base automatically. This is especially valuable for sites that update frequently (blogs, product catalogs, news sections).
Manual retrain process
- Go to your chatbot's Sources.
- Click the refresh icon next to a website source to re-crawl it.
- Or upload new document versions to replace outdated files.
- Add or update Q&A pairs as needed.
Source management
Over time, you may need to reorganize your sources:
- Remove outdated sources. Delete website URLs that are no longer relevant or documents that have been replaced.
- Consolidate Q&A pairs. If you have many similar Q&A pairs, combine them into a single, comprehensive pair.
- Audit annually. Once a year, review all sources. Remove anything outdated and add new content areas.
Quality tips
These practices make the difference between a chatbot that frustrates visitors and one that genuinely helps.
Do
- Train on customer-facing content. The chatbot should know what your customers need to know.
- Write Q&A pairs in your brand voice. If your brand is casual, the Q&A answers should be casual. If professional, keep them professional.
- Test with real questions. Ask your support team what customers ask most often. Use those as test cases.
- Include context in answers. Instead of "Yes," answer "Yes, we ship internationally. Standard international shipping takes 7-14 business days and costs $15. Express international is available for $35."
- Start small and expand. A chatbot trained on 10 well-chosen pages outperforms one trained on 500 mediocre pages.
Avoid
- Do not train on outdated content. Old pricing, discontinued products, or expired promotions cause confusion.
- Do not train on internal documentation that references processes invisible to customers.
- Do not skip the playground. Every chatbot should be tested before going live.
- Do not rely on one source type alone. Website crawling misses nuance. Q&A pairs alone miss breadth. Use them together.
- Do not train on competitor content. Train on your own material. If visitors ask about competitors, add a Q&A pair with a factual, neutral comparison.
What to train on first
If you are starting from scratch, prioritize in this order:
- Your website (full site crawl or key sections)
- Top 10 FAQ answers as Q&A pairs
- Pricing details as a Q&A pair (ensures accuracy)
- Policy pages (if not covered by the website crawl)
- Product documentation (uploaded as documents)
- Brand and company background (text snippet)
This gives you broad coverage from the website, precision on critical topics from Q&A pairs, and depth from documents. You can always add more sources later as you identify gaps.
Next steps
Once your chatbot is trained and tested, you are ready to deploy it. Add it to your WordPress site, Shopify store, or any website with the JS widget. Then explore integrations to connect your chatbot to the rest of your tools.
Training is ongoing. The best chatbots improve over time as you review conversations, identify weak spots, and add training data to address them. Treat it like any other customer-facing content: review it regularly and keep it current.
No credit card required.
