Guide

How Neolo Chat works, step by step

1. Key concepts

Agent

A chat assistant with its own knowledge base, instructions, model, look and limits. An account can have several agents, depending on the plan.

Data source

Where the agent learns from: a URL, a whole domain, a sitemap.xml or pasted text.

Chunk

A piece of text of about 1,800 characters that the content is split into so it can be searched precisely.

Embedding

A numeric representation of the meaning of a text. Two texts with similar meaning have nearby embeddings, even if they share no words.

Semantic search (RAG)

Looking up the chunks in the knowledge base that are closest in meaning to the question, and giving them to the model as context before it answers.

Instructions (system prompt)

The text that defines how the agent should behave: tone, scope, what to do and what not to do.

Widget

The chat the visitor sees on your site: a floating bubble that opens a panel, or an embedded iframe.

Visitor and conversation

Each browser gets an anonymous code and has one conversation per agent, kept if they come back to the site.

2. Step by step: from zero to a working agent

Everything is done from the dashboard at neolochat.com/dashboard. It takes about 10 minutes and needs no coding.

1. Create your account. Enter your email and open the link we send you (magic link). There are no passwords. The free account includes 1 agent and 50 messages per month, no credit card needed.
2. Create an agent. Click New agent and name it. Each agent has the tabs General, Model, Instructions, Data sources, Playground, Guardrails, Embed, Chat log and Analytics.
3. Train it on your site (Data sources tab). Choose one or more sources:
   • URL: reads a single page.
   • Domain crawl: starts from a URL and follows internal links on the same domain, up to 40 pages.
   • Sitemap: reads the URLs in your sitemap.xml, up to 150 pages.
   • Text: you paste the information yourself (prices, policies, FAQs).
   Each source goes through the states processing → indexed (or failed, with the error reason).
4. Write the instructions (Instructions tab). Define the tone, which topics it may answer, which links to use, which language to reply in and when to hand over to a person. This is the part that changes answer quality the most.
5. Choose the model (Model tab). It runs on OpenAI models. You don't need an account or your own key: Neolo's infrastructure is used.
6. Customize the look (General tab): color (picker or HEX code), font, font size, agent picture, bubble icon, welcome message, message field text and footer. There is a live preview.
   • Automatic translation: if the visitor's browser is in another language, the welcome message, message field and footer are translated automatically.
   • Ask for the email: optional, with three modes: off, optional (the visitor can skip it) or required before chatting. The email is saved with the conversation.
7. Test it (Playground tab). Chat with the agent using its real instructions, model and knowledge. These tests are not mixed with visitor conversations.
8. Set the limits (Guardrails tab):
   • Maximum messages per visitor in a period (default: 20 messages every 60 seconds) and the text they see when they exceed it.
   • Monthly and daily budget in dollars: when reached, the agent stops answering.
   • Blocked countries: those visitors are not shown the widget.
   • Allowed domains: only your sites can load the chat.
9. Publish it (Embed tab). Copy this line and paste it before the </body> tag of your site:

   <script src="https://www.neolochat.com/widget.js"
           data-agent-id="YOUR-AGENT-ID" defer></script>

   If you prefer a chat embedded in the page instead of a floating bubble, use the iframe:

   <iframe src="https://www.neolochat.com/iframe/YOUR-AGENT-ID"
           width="100%" style="height:100%;min-height:500px"
           frameborder="0"></iframe>

10. Review and improve. In Chat log you read the conversations (with country and email if you asked for it) and can export them to CSV. In Analytics you see views, conversations, messages, active users and thumbs up/down. If your site changes, click Retrain AI Agent in Data sources.

3. What happens when a visitor sends a message

From the moment the visitor hits send until they see the answer, the system does this:

1. The widget sends the message to the Neolo Chat server together with the visitor's anonymous code.
2. It checks the origin: if the agent has allowed domains, the request must come from one of them. If the visitor's country is blocked, it is rejected.
3. It applies the per-visitor limit. If exceeded, it replies with the configured limit text, without calling the model.
4. It checks budget and quota: the agent's daily and monthly budget and the plan's monthly messages. If exhausted, it replies that the assistant is unavailable for now.
5. It loads the visitor's conversation (or creates it) and takes the last 10 messages as conversation memory.
6. It turns the question into an embedding and compares it with those of all the agent's chunks.
7. It selects the 5 most similar chunks by cosine similarity. It also always adds the site's navigation-links chunk, so the agent knows the real URLs.
8. It builds the model request: instructions + retrieved chunks as context + recent history + the question.
9. The model writes the answer and the server stores it together with the tokens used and the cost.
10. The widget shows the answer, with clickable links and 👍/👎 buttons so the visitor can rate it.

If the context does not contain the answer, the behavior depends on your instructions: it can say it is not sure, offer human contact, or answer from general knowledge.

4. How the agent is trained, in detail

1. Download: the server fetches the content of each source. It blocks internal or private addresses for security.
2. Text extraction: code and navigation are removed, leaving the readable text of each page with its title.
3. Splitting into chunks: the text is cut into pieces of about 1,800 characters with 200 overlapping, so ideas are not cut in half.
4. Embeddings: each chunk is converted into a numeric vector with an OpenAI embeddings model.
5. Storage: chunk and vector are saved in the database linked to the agent. When retraining, the chunks of that source are replaced.
6. Navigation links: when crawling a domain, the list of the site's links is also saved as a special chunk.

5. Limits and rules

6. Privacy and isolation

• The widget is drawn inside a Shadow DOM: its styles don't affect your site and your site's styles don't affect the chat.
• It uses no third-party cookies. The visitor's anonymous code is stored in the local storage of your own domain.
• The platform's API keys never reach the visitor's browser.
• The visitor's email is only saved if you enable that option and they type it in.

7. Frequently asked questions

Where does the agent get its answers?

From the chunks of your content closest to the question, plus the instructions you wrote. If it's not in your content, it depends on your instructions.

Do I need to retrain every time my website changes?

Yes, by clicking Retrain AI Agent. Meanwhile, the agent keeps answering with the previous content.

Do visitors need to sign up?

No. They are identified by an anonymous code and, if you enable it, they can leave their email.

Does it answer in the visitor's language?

Yes, if your instructions say so. The welcome message and footer are also translated according to the browser language.

Can I see what visitors ask?

Yes. The Chat log shows each conversation and it can be exported to CSV.

8. Key facts as a list

• Product: Neolo Chat, by Neolo (neolo.com). Site: neolochat.com.
• What it is: an AI chat widget trained on the customer's own website content.
• Technique: retrieval-augmented generation (RAG) with embeddings and cosine similarity.
• Training sources: URL, domain crawl, sitemap.xml, text.
• Installation: one <script> line before </body>, or an <iframe>.
• Account: sign-in with an email magic link, no password.
• Control: instructions, per-visitor limit, daily and monthly budget, country blocking, allowed domains.
• Follow-up: Chat log, CSV export, Analytics, option to ask for the visitor's email.
• Dashboard languages: Spanish and English. The agent can answer in any language.