TeklTalk documentation
TeklTalk reads the pages you already have and answers your visitors' questions from them, with links back to the page each answer came from. These pages describe exactly how that works — and what it does not do.
The shape of it
There is one unit of everything here, and it is a project. A project owns a set of content, a widget appearance, and a chat history. One project is one bot on one website. Accounts can have as many as they like on every plan.
A project goes through the same three stages every time:
- You add content. A sitemap URL, a list of page URLs, a block of text you paste in, or a document you upload. See Knowledge sources.
- We read and index it. Each page is fetched, stripped down to its readable article text, split into overlapping chunks, and turned into embeddings stored alongside the text.
- Visitors ask questions. Their question is embedded the same way, the closest chunks are retrieved, and the model answers using only those chunks as its source material.
What happens on a question
The retrieval step is the part worth understanding, because it explains most of the bot’s behaviour.
Every question is compared against your indexed content, and the best match gets a similarity score between 0 and 1. If that score does not beat your project’s similarityThreshold, the bot replies with your fallback message and no model call is made at all. That is deliberate: it is the mechanism that stops the bot inventing an answer to a question your site does not cover.
When the score does clear the threshold, the top matches are handed to the model as reference material, with instructions to base every factual claim on them and not to invent specifics. The pages those chunks came from are attached to the reply as citations, so a visitor can check the answer and you can see what the bot is reading.
What it is not
- It is not a general-purpose chatbot. It answers from your content, and says so when it cannot.
- It does not browse the live web at question time. It answers from what was indexed the last time you crawled.
- It does not take actions on your behalf — no bookings, orders, or account changes. It reads and it answers.
Where to go next
- Install — The script tag, where it goes, and how to check it worked.
- Knowledge sources — Sitemaps, pasted text and uploaded documents — and what each costs.
- Configuration — Every widget setting, what it does, and what it is bounded by.
- Plans and limits — Credits, conversations, rate limits and message size.
- Troubleshooting — The widget won't show, the bot won't answer, a page won't index.
If something here does not answer your question, email support@tekltalk.com.