automation
How to build a knowledge base for a Telegram support bot
A customer asks your bot one question outside its script and the whole conversation falls apart. Here's how to build a knowledge base that actually holds up.
Your bot answers the first question perfectly. Then a customer asks something slightly off-script — "does this work with my old plan?" — and the bot either loops, hallucinates an answer, or dumps them into a generic "contact support" message. That's not a bot problem. That's a knowledge base problem.
Most teams build the prompt first and treat the knowledge base as an afterthought. It's the other way around. The prompt tells the bot how to talk. The knowledge base tells it what's actually true. Without one, you get a confident bot that's confidently wrong.
How big does a Telegram support bot's knowledge base need to be?
For most support use cases, 15-40 well-written Q&A entries covering your top recurring questions is enough to handle 80% of incoming volume. You don't need a 200-page manual — you need the specific questions your customers actually ask, answered in plain language the bot can reuse almost verbatim.
If your product is simple (a single plan, a single flow), you might only need a handful of FAQs plus your pricing and refund policy. If your product is complex — multiple tiers, integrations, edge cases — that's exactly when a knowledge base stops being optional. A Telegram AI agent can handle the conversation on its own, or wait for your intervention when the question is unclear — but it can only do that well if it has real source material to draw from instead of guessing.
What should actually go in the knowledge base?
Skip the temptation to paste your entire help center into a doc and call it done. Bots perform better with curated, structured content than with raw dumps. Build it around these categories:
Top 20 support tickets from the last 90 days. Pull your actual ticket history — these are the real questions, not the ones you assume people ask.
Pricing and plan differences. Write these as direct comparisons, not marketing copy. "Plan A includes X, Plan B includes X and Y."
Refund, cancellation, and billing policies. Spell out exact timeframes and conditions — vague answers here create escalations.
Product limitations. What it can't do is as important as what it can. A bot that admits a limitation looks more trustworthy than one that dodges.
Escalation triggers. A short list of question types the bot should never try to answer alone — account deletion requests, legal complaints, refund disputes above a certain amount.
Tone and formatting examples. Two or three sample answers written the way you want the bot to sound, so it mirrors your voice instead of a generic assistant tone.
How do you write entries the bot can actually use?
Format matters more than volume. Write each entry as a direct question paired with a short, complete answer — 2-4 sentences, no filler. Avoid burying the answer in a paragraph of context the bot has to parse. If a human support agent would need to scroll to find the answer, so will the bot's retrieval logic.
A good test: could a new hire answer a customer using only this entry, with zero outside knowledge? If not, rewrite it. Vague entries like "we support most integrations" produce vague bot answers. Specific entries like "we support Zapier, Make, and native API access via webhooks" produce specific, quotable answers.
How does the knowledge base connect to the bot's prompt?
The knowledge base and the prompt are two separate pieces that work together. The prompt defines behavior — how the bot should greet people, when to push toward booking a call or filling out a form, what tone to use. The knowledge base supplies facts. CRMChat's Telegram AI Sales Agent sits in your Telegram account and replies to incoming messages using both: the prompt shapes how it responds, and the knowledge base is what it references when a client asks something the prompt alone can't answer.
CRMChat's AI agent setup uses exactly this split — you write a prompt using CRMChat's prompt-building assistant, then attach a knowledge base if your product is complex enough to need FAQs beyond the core script. Simple products might run fine on prompt alone. Anything with pricing tiers, onboarding steps, or policy edge cases needs the knowledge base layered on top.
How do you keep the knowledge base from going stale?
A knowledge base that isn't maintained becomes a liability faster than having no knowledge base at all — customers get confidently wrong answers about pricing that changed three months ago. Set a maintenance rhythm:
Review new support tickets weekly for the first month, then monthly after that, and add any recurring question that isn't already covered.
Flag every bot conversation where a human had to step in, and check whether a missing or outdated KB entry caused it.
Update pricing, policy, and feature entries the same day they change in production — not on the next content cycle.
Remove or archive entries for deprecated features so the bot never surfaces dead information.
Spot-check a sample of bot conversations monthly to catch drift between what the KB says and what the bot is actually telling people.
Should you test the bot before customers see it?
Yes, and skipping this step is the most common mistake teams make. Run the bot through your worst-case questions before launch: ambiguous phrasing, questions with no clean answer in the KB, and questions designed to bait it into a policy it shouldn't discuss. If it fails those in testing, it will fail them live in front of a paying customer. Test with people who didn't write the knowledge base — they'll phrase things in ways you didn't anticipate.
Once it's live, treat escalations as free QA. Every time a conversation gets handed to a human, that's a signal about a gap in the knowledge base, not just a one-off exception. Check CRMChat's Help Center for setup specifics if you're configuring this inside CRMChat directly, and the CRMChat API docs if you want to sync your knowledge base from an existing help desk or docs system instead of maintaining it separately.
