AI voice bot knowledge base: price list, catalog and webhooks

A client asks about a price or a free slot, the bot promises that “a manager will call you back” — and the client cools off and goes to a competitor. An AI voice bot knowledge base solves this: through webhooks, the bot answers from live data — your price list, stock, catalog and the client card from your system. And it sends the results of the conversation by itself to where they are used — with no manual entry.

The bot's main strength is large databases. It keeps the whole conversation and large volumes of data in memory and doesn't forget what the client said at the start. And database search with parameters keeps it from making things up: the bot books the exact time and doesn't mix up addresses, staff and dates.

A webhook is a link through which the dashboard fetches or sends data by itself. Webhooks and triggers aren't billed separately: you pay only for conversation minutes, $0.12 per minute on average. Watch the two-and-a-half-minute video, and below you'll find the step-by-step setup.

Why businesses need it

• An online school keeps its program and prices in a Google document. Change a price in the document — and the bot quotes the new one on the very next call.
• A real estate agency connects its own database: the client names the district, budget and number of rooms, and the bot finds 3–5 matching properties right during the conversation.
• An online store answers questions about stock and order status by item number or order number. The call result, together with a link to the recording, goes to its system.
• A store with a large catalog pays less. The dynamic database gives AI only the rows it needs, not the whole catalog, and the bot doesn't get confused.
• A taxi or delivery service: the bot asks for the pickup and drop-off addresses, the server calculates the price, and the client hears the fare right in the conversation — with no wait for an operator.
• A salon, clinic or car rental: the bot checks free slots or available cars and books, reschedules and cancels appointments by itself. The administrator doesn't have to call clients back.
• Every conversation is transparent: the call review shows what data the bot received, what it understood from the conversation and what it answered. You spot mistakes right away, not from client complaints.

How the bot gets data: three types of webhooks

Webhooks come in three types:

1. A loading webhook fires before the call. The data arrives together with the instruction, and the bot relies on it from the very first phrase.
2. A dynamic webhook fires during the call, when the trigger keywords come up in the conversation — from the client or from the bot itself. AI takes the needed values from the conversation, the dashboard sends a request, and your server returns a short answer. While the request is running, the bot says “one second, let me check”.
3. A writing webhook fires after the call. AI fills in the link parameters from the dialog and sends the result.

How to connect a Google document to an AI voice bot

This is the simplest option: you need neither hosting nor a programmer. The document is pulled in while the phone is still ringing, so the bot knows your price list from the first phrase.

1. 1
   Share the document

   In Google, click “Share” → “Anyone with the link”, role “Viewer”. Copy the link.
2. 2
   Add a webhook

   Bot card → “Webhooks” → “Load webhook, google document, txt” → “Add webhook”.
3. 3
   Fill in the fields

   “Webhook name”, for example “Price list”, and “Request URL”. A document link turns into a plain-text export by itself, and a spreadsheet link into CSV. A txt file or your API (the address where your system serves its data) works too.
4. 4
   Pass client data

   If the link points to your API, tick the “Quick parameters”: “Pass phone”, “Pass lead ID”. That way, the server knows whose card to return.
5. 5
   Save

   Click “Save” on the bottom bar. The dashboard checks the data size and loading speed.
6. 6
   Check

   In “Training AI”, the “Loading the database from the webhook” window shows what the bot received. How to work with this tab is described in “AI voice bot training”.

How to send call results to your own system

Open “Write webhook” → “Add webhook”, tick “Send webhook” and enter the receiver's address: every write webhook works as a “Custom webhook” with “Quick parameters”. The request method is GET only, meaning the data is passed right in the link.

• “Quick parameters” insert the phone, name, lead number and a link to the call recording.
• “Add parameter” — your own value from the conversation. Enter the “Parameter name” in Latin characters and the “Parameter search prompt”, for example summary and “Write a summary of the conversation in two sentences”. The parameter is added to the link by itself as summary=${summary}.
• Webhooks for amoCRM and Yandex Forms are created by the wizards on the “Integrations” tab: they keep their own setup buttons, and the wizard fills in the link and parameters itself.

How the bot fills in fields and steps in the dashboard's own CRM is covered in “Writing call results to CRM”.

Large databases: “Database search”

AI receives the loading data in full at the start of the call. The dashboard recommends keeping all loading webhooks together within 50 KB and 7 seconds of loading. Above that, a warning on save says the call will cost more or start with silence, and saving needs “I accept the risk, save”. Above 100 KB or 10 seconds, saving is blocked. A catalog of thousands of items doesn't fit and isn't needed in full.

The dynamic database solves this — a search in your data during the call: even from a million rows, the bot gets only the records it needs. How to connect it:

1. Open “Triggers” and click “Add trigger”.
2. Enter the “Keywords to launch the trigger” — the client's phrases (“in stock”, “for tomorrow”, “two-room flat”) or the bot's own phrase, such as “checking the schedule”.
3. Set the “Pause filler phrase” — what the bot says while it waits for the answer.
4. Specify the “Dynamic database request URL” and parameters with a “Parameter prompt”, for example “Which district did the client name? One word”.
5. Write the “Dynamic webhook instruction” — how to answer based on what was found. If left empty, a ready-made instruction for free booking slots is used. For a catalog or stock, rewrite it for your task.

On your side, you need an address that accepts a GET request and quickly returns 3–10 matching records. A programmer can build this “middleware” — a small piece of code between the bot and your database — or order a turnkey connection with the button at the end of the article. A new trigger comes with a demo database: replace the link, parameters and instruction with your own. The same tab also handles other bot actions, such as transferring a call from the bot to a manager. For details, see “AI voice bot database lookup”.

Dynamic database checklist: a brief for your developer

First check whether you need a dynamic database at all: a loading webhook is usually enough, and you save on a developer. Load your data as a Google document and ask the bot about dates, services and prices in “Training AI”. If the bot doesn't mix things up and the data fits within 50 KB, you don't need a dynamic database. If it does mix them up, you need database search. A schedule in Yclients or Google Calendar is connected by the wizards on the “Integrations” tab, with no developer: see “Yclients integration”. For your own database, give your developer this brief:

• List the search parameters, for example date, service and staff member. Booking usually needs three addresses: a directory of services and staff with their IDs for the loading webhook, a free-slot search for “Database search”, and the booking itself for the write webhook.
• Keep the answer short: for booking, three free slots closest to the time the client named are enough. Plain text, JSON or XML with clear field names all work. In our setup experience, the bot reads XML most reliably.
• A clear phrase instead of an error. An empty date or an unknown service must not return “error” or code 500. Let the server answer: “Please choose a day that suits you for the appointment”. The bot reads it and asks the client again.
• Before setup, open every link in a browser: remove the date, change the service, pick a day off. Every answer should suggest the next step rather than show an error.

How to write a “Parameter prompt” and start the search at the right moment

AI fills in parameters from the conversation, the loading data and the database's previous answers. It doesn't see the main instruction, so put everything it needs into the prompt itself:

• Ask for one value in a strict format: “Write the date the client wants to book. Strictly DD.MM.YYYY, and if they named a time — DD.MM.YYYY HH:MM. No comments and no period at the end”.
• IDs — digits only: “Write the ID of the service the client wants to book. Digits only, for example 1”. AI finds the ID in the directory from the loading webhook. A short list can be spelled out right in the prompt: “If the client named the branch on Lenin Street — write 101”.
• Say what to write when there's no value: “If the client didn't name a date — write -”. Put constant values, such as the city, straight into the link: city=msk.
• Start the search with the bot's own phrase. Add a step to the “Conversation plan”: “When the client has named the service and the day, write ‘Checking the schedule...’, add nothing else and wait for an answer”. Enter “checking the schedule” in the keywords. A second step of this kind covers another date: “If the client wants a different day, write ‘Checking the schedule...’ again”.
• Ask the bot to name the service exactly as it appears in the directory: the client says “hair dye”, the bot confirms “hair coloring”. That way AI picks the right ID. Put the steps “service — name from the directory — day — search phrase” in a row, and other questions before or after them.

How to test the dynamic database and the write webhook in “Training AI”

Check the dynamic database and the write webhook in the test chat before the first calls:

• Start the conversation with the trigger’s keywords, for example “do you have anything free tomorrow?”. A “Webhook: <trigger name>” block appears in the chat: which word fired the trigger, what AI put into each parameter, the final link and the database's answer. How to review a real call is covered below.
• The bot writes the reply right after this block from the trigger’s “Dynamic webhook instruction”, not from the main instruction. That is why “Fix” on such a reply edits exactly that instruction: the window shows the “Fixing the dynamic webhook instruction” banner.
• “Check answer stability” under such a reply repeats the request 10 times and shows whether dates, slots and prices drift.
• The write webhook is checked in the “Data write test”: select “Webhook: <address>” and click “Test the data write”. Each parameter shows its “AI answer”, and you fix an inaccurate prompt with “✎ Edit the prompt”. The request is sent for real, so the data will appear in your system.

How to review a call with a dynamic database

If the bot named the wrong time on a real call or didn't book the client, open “Training AI” → “Find a call”. Paste the ID from “Call details” or pick the call from the list for the last 7 days. The conversation loads into the chat with every webhook in its place. Go through them in order:

• Loading webhooks — what the bot received before the conversation. No service directory means AI can't find the right ID.
• The dynamic database block — the request link, the AI answer for each parameter and your server's response. A wrong date or ID means you should refine the trigger's “Parameter prompt”.
• The bot's reply built from the database data — what the bot said based on it. Wrong slots or format — “Fix” under the bot's reply corrects the “Dynamic webhook instruction”.
• The write webhook block — the parameters and your system's response. For example, AI wrote update instead of create, and the booking wasn't created. Refine the prompt: “Confirmed a new booking — write create, rescheduled — update, cancelled — delete”.

Compare the bot's reply with the server response saved in the call. Don't open the request link again: the free slot may already be taken and the date may have passed.

Read also: AI voice bot call transfer to a manager and calendar booking, How to train an AI voice bot to answer like your best sales rep, AI voice bot CRM integration: call results and funnel step, Ai voice bot personalization