An AI chatbot for your site, without building RAG

Someone opens your site with a specific question. Not “tell me about your product” but the real one: does this handle my case, what does it cost for my situation, is the thing they need buried somewhere in your docs. They do not want to read your whole site. They want one answer.

A search box gives them links. A generic AI chatbot gives them a confident guess that may be wrong. What they actually want is an assistant that answers from your material, and only your material, and shows where each answer came from.

There are two ways to build that. You can assemble the machinery yourself, or you can point at something that already has it. It is worth understanding what the first path involves before you decide.

What “answering from your content” actually takes

The technique behind a grounded chatbot is retrieval-augmented generation, usually shortened to RAG. The idea is simple to say and a lot of work to ship: before the model answers, you retrieve the relevant pieces of your own content and make it answer from those, not from whatever it happened to be trained on.

Doing that well means building, and then maintaining, a stack:

• Ingestion for every format you care about: PDFs, web pages, docs, transcripts. Each one parses differently, and each one breaks differently.
• Chunking that respects the structure of a document, so a retrieved passage is actually coherent instead of cut off mid-thought.
• Embeddings and a vector store, plus the retrieval tuning that decides which chunks come back for a given question. This is the part that quietly makes or breaks answer quality.
• Grounding logic so the model stays inside the retrieved material instead of drifting into plausible invention.
• A citation interface, because an answer nobody can trace is an answer nobody trusts.
• Refusal behavior for questions your content does not cover, so the bot says “I do not know” instead of guessing.

Then the unglamorous half: hosting, monitoring, and some way to tell whether a change made answers better or worse. Add real-time voice and you roughly double the surface area again.

None of this is impossible. Plenty of teams build it. But it is a project with a long tail of maintenance, and for most people the goal was never to own a retrieval pipeline. The goal was to let visitors ask a question and get a grounded answer.

The shortcut: a hosted, grounded chatbot

Mindola is that second path. You give it your content, and it turns that content into a lens: one public link where anyone can ask questions and get answers drawn only from your material, with the sources shown. When a question falls outside what you have given it, the lens says so honestly instead of filling the silence with a guess.

You do not touch chunking, embeddings, retrieval, or a citation UI. That work is done and hosted. You provide the knowledge and the voice; the grounded answers, the citations, the honest refusals, and the page all come with it.

The point of RAG was never the pipeline. It was a visitor getting a straight answer from your own words. You can have the second without owning the first.

Two ways to wire it up

From your AI assistant. Mindola runs a hosted MCP server, so an assistant like Claude, Cursor, or ChatGPT can create and fill a lens for you inside one conversation. You connect the server once, point it at your docs or paste your material, and ask it to build the lens. No account is needed to start. The connect guides walk through each client.

From code. If you would rather script it, there is a small REST API. You create a key, create a lens, and add knowledge as plain text or URLs. Here is the whole first call:

curl -s https://app.mindola.ai/api/v1/lenses \
  -H "Authorization: Bearer mnd_your_api_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Product docs assistant",
    "knowledge": [
      { "type": "url",  "content": "https://example.com/docs/getting-started" },
      { "type": "text", "content": "Refunds are available within 30 days." }
    ]
  }'

Ingestion runs in the background. You poll the lens until it is ready, then open the link and ask it something. Runnable examples in curl, Node, and Python live in the open-source repo.

When you should still build your own

Hosted is not always the right call, and it is worth being honest about the edges. Build your own RAG when you need full control over which models run, where your data is stored, or exactly how retrieval is ranked. The same goes for embedding the pipeline deep inside an existing product, where a hosted page is the wrong shape. If that is you, roll your own and keep the control.

For everyone else, whose actual goal is a grounded assistant on a link they can share, the hosted path skips the months and the maintenance.

Try it before you build anything

You can talk to a live lens right now, no signup: the Charles Darwin demo answers only from public-domain writing and admits what it does not know. The gallery has a dozen more. When you are ready to make your own, the docs cover every path end to end.

Frequently asked questions

What is grounded AI, and how is it different from a normal chatbot?

A grounded AI answers only from a specific set of source material and cites where each answer came from. A normal chatbot answers from its general training, so it will confidently make things up about you or your product. Grounding is what lets you put your name on the answers.

Does the chatbot hallucinate?

Answers come from your material only. When a question falls outside it, the lens refuses honestly and can pass the question to you, rather than inventing a plausible wrong answer. Zero guesses is the point, not a limitation.

What content can I add?

Through the API and the MCP tools, you add two kinds of item: plain text and URLs. In the Mindola app you can also add richer sources such as PDF uploads. Everything runs through the same pipeline so the lens can answer from all of it, with citations.

Do I need an account to try it?

No. An AI assistant connected to the MCP server can create a lens for you with no signup. You get a live link and a claim link, and you sign up for free only when you want to keep the lens.

Is there an API and is it open?

Yes. The REST API and the MCP server are both documented, and the examples repo is open source under an MIT license. The hosted service that runs the retrieval is closed source.