Back to all posts
ChatNexus Blog
April 11, 2026 9 min read Knowledge Base

How to Build an Effective Knowledge Base for Your AI Agent

The bot does not read your whole knowledge base on every question. It searches for a few short chunks and answers from those. If the chunk is vague, old, or split in a weird place, the answer will be too. Here is how to write the source so the search can find it.

The bot only sees a few chunks

ChatNexus splits what you add into pieces of about 1,500 characters. A question pulls a few of those pieces, plus the bits just before and after a match, and the model answers from that. Two things decide the reply: did the right piece get picked, and did that piece actually contain the answer.

For a longer explanation of why this beats retraining a model, see knowledge bases vs fine-tuning.

Five writing habits that help

1. Use the words customers use

A brochure heading like "Delivery Operations Policy" is a bad match for "Do you deliver on weekends?" Put the question in the heading, then the answer right under it. 

Do you deliver on weekends?
Yes, in metro areas, for orders placed before 2pm Friday. Rural orders placed
after Thursday go out Monday.

2. Say the actual rule

  • Too vague: "We offer flexible return options for most products."
  • Useful: "You can return an unused item within 30 days for a full refund. Final Sale items cannot be returned. Email returns@yourcompany.com with your order number."

The second one tells the visitor what to do. The first one makes them ask again.

3. One topic per section

A 2,000-word page called "Our Services" gets chopped into pieces that may not make sense alone. If a section cannot answer one question by itself, split it. Each section needs its own heading.

4. Update it when the business changes

Fix the source when:

  • Prices or plans change
  • You add or remove a product
  • A policy changes
  • The bot keeps giving the same wrong answer

On ChatNexus you can paste text, upload a .txt, .md, or PDF, crawl a normal web page, or paste a public sitemap.xml to import multiple pages. Google Docs and Google Sheets go in the Google Doc tab. Share the file with the address shown there first. The normal URL crawl cannot open Google links.

Auto-sync on a Google file checks about once an hour, and only if you turn it on. It does not update the moment you hit save in Google. You can also sync that file yourself from the knowledge base page. The steps are in connect a Google Doc or Sheet.

Free accounts can store 100 knowledge base entries per agent, and each upload can be up to 5 MB. Paid plans allow more.

5. Write the "we don't do that" answers

People ask about sold-out items, areas you do not serve, and competitors. If you do not write those answers, the bot may invent one. A short line is enough: "We don't ship to PO boxes. If the item is out of stock, there is no waitlist."

Industry Examples

The right knowledge base structure varies by industry. Here are practical illustrations across five sectors.

E-Commerce & Retail

An online clothing store's AI agent handles hundreds of queries daily about sizing, returns, shipping, and stock availability. An effective knowledge base for this context includes:

  • Size guide: Exact measurements for each size (S, M, L, XL), per product category (tops, bottoms, footwear). Include a note about whether items run small or large.
  • Shipping policy: Delivery times per region, cutoff times for same-day dispatch, carrier names, and tracking instructions.
  • Returns & exchanges: Step-by-step process, time windows, eligible vs. non-eligible items, and refund timeline.
  • Out of stock: What happens when an item is unavailable — do they get notified? Is there a waitlist?
  • Discount codes: How to apply a code at checkout, whether codes stack, and expiry policies.

Tip: Import your product FAQ page and returns policy PDF directly. These are usually well-structured and already written for customers.

Healthcare & Allied Health Clinics

A physiotherapy or general practice clinic can use an AI agent to handle appointment and admin queries — freeing up front desk staff. Effective content includes:

  • Services offered: List each treatment type with a plain-language description. Avoid clinical jargon unless your audience is professional.
  • Booking process: How to book (online, phone, walk-in), available time slots, and what to bring to a first appointment.
  • Insurance & billing: Which health funds are accepted, how to claim on the spot, and gap amounts.
  • Cancellation policy: Exact timeframe required for cancellations and any fees for late cancellations.
  • Scope limits: Explicitly state what the agent cannot help with — e.g., "For medical advice or urgent concerns, please call the clinic directly or contact emergency services."

Keep it to admin. Do not put patient records or treatment advice in the knowledge base. If the bot will handle health data, tenancy screening, or card numbers, ChatNexus will ask you to complete a data processing agreement before that agent can go live. See when a chatbot needs a DPA. The healthcare intake template is a good starting point for a clinic agent.

Legal & Professional Services

Law firms, accounting practices, and consultancies can use an AI agent to handle initial enquiries and reduce time spent on qualifying leads. Effective content includes:

  • Practice areas: What types of matters the firm handles (e.g., "We specialise in commercial property, not residential conveyancing").
  • Engagement process: How to engage the firm — initial consultation, whether it's free, what documents to bring.
  • Fee structure: Fixed-fee services vs. hourly billing, retainer arrangements, and payment terms.
  • Jurisdiction: Which states or countries the firm is licensed to practise in.
  • Clear disclaimers: "This agent provides general information only. Nothing here constitutes legal advice. Contact us for advice specific to your situation."

Tip: Use your existing website FAQ and "About" pages as a starting point. They're usually already scoped appropriately for public consumption.

Hospitality & Restaurants

Restaurants, hotels, and venues get the same questions all day. Put these in the knowledge base:

  • Menu & dietary requirements: Full menu text with allergen information. Clearly flag gluten-free, vegan, halal, and nut-free options.
  • Reservations: How to book, how far in advance, group booking minimums, and deposit requirements.
  • Opening hours: Include public holidays and any seasonal variations.
  • Location & parking: Full address, nearest public transport, and parking options including cost.
  • Private events: Minimum spend, room capacity, AV equipment availability, and how to enquire.

Tip: Upload your menu as a PDF. Include the allergen matrix if you have one — it's one of the most common questions restaurant agents receive.

SaaS & Tech Support

Software products generate a consistent stream of "how do I..." questions. A support-focused agent dramatically reduces support ticket volume when the knowledge base is thorough:

  • Getting started: Step-by-step onboarding flow — account creation, first setup, key terminology.
  • Feature guides: One section per key feature, written as "How to [do X]" with numbered steps.
  • Common errors: List frequent error messages with their cause and resolution. Users often paste error text directly into chat.
  • Billing & plan limits: What's included in each plan, how usage is counted, and how to upgrade.
  • Integrations: Which platforms integrate natively, how to connect them, and known limitations.

Tip: Look at your last support tickets. Write a short section for the 20 questions that show up most often. People often paste the error text straight into the chat, so include the exact error message in the heading. The customer support FAQ template and the SaaS onboarding template are good starting points for this kind of agent. Browse all industry templates for other starting points.

A few things that cause wrong answers

  • Two versions of the same rule. If returns are "14 days" in one file and "30 days" in another, the bot may quote either.
  • Internal docs. Staff handbooks and private price sheets do not belong in a public bot. It can repeat them.
  • A tone that does not match you. The reply often sounds like the source. If you write like a policy manual, the bot will too.

A practical way to start

Start with the page you already send people to: the FAQ, the returns policy, or the service list. Paste the text, upload the PDF, or crawl that URL. Then ask the bot the 10 questions you hear most. When an answer is wrong or vague, add one short section for that question and ask again.

Do not paste staff handbooks, internal price sheets, or anything you would not put on the public website. The bot can repeat whatever is in the knowledge base.

Start for free on ChatNexus. No credit card. The Free plan includes two agents and 30 days of access from signup.