Fuentia
EspañolEnglish
Sign in to the panel

RAG as a service

An assistant that answers from your documents, and shows you where.

Fuentia turns your manuals, policies and FAQs into an assistant that cites every answer. If the answer is not in your documents, it says so.

Get started in the panel See how it works
  • Answers with verifiable sources
  • Every assistant isolated from the rest
  • By API or with a widget

An assistant in four steps

No model training, no infrastructure to run. Here is what you do and what Fuentia does underneath.

Step 1

Create a notebook

A notebook is an assistant: it has its own documents, instructions and keys. Create one per topic, product or client.

You

  • Give it a name and, if you like, instructions about tone or style.
  • Pick the language model from the allowed ones.

Fuentia

  • We keep it isolated: a notebook never sees the documents of another one, or of another account.
  • The database itself enforces that isolation on every query, not just the application.
Fig 1One lock per notebook
The open notebook: the one the query reachesLocked notebooks Every notebook has its own lock. A notebook key opens only its own: the others stay locked, even within the same account.

Step 2

Upload your documents

Drop your files into the notebook. Fuentia reads them, cuts them up and indexes them so they can be found by meaning and by words.

You

  • You upload TXT, Markdown, and PDF, up to 20 MB per file.
  • You see each file’s status: queued, processing, ready or failed.

Fuentia

  • We cut the text into chunks of about 1,400 characters, without splitting sentences, each repeating the last 15% of the one before so no context is lost at the cut.
  • We turn each chunk into a 1,024-dimension vector that captures its meaning.
Fig 2From document to chunks
Window, overlap and vectorsThe document’s text The window is 1,400 characters long and advances 1,400 minus 210, so the first 210 of every card (15%) repeat the end of the one before. Each card gets its 1,024-dimension vector.

Step 3

Ask

Every question searches your documents in two ways at once, and only answers with what it finds.

You

  • You ask in plain language, from the panel, the API or the widget.
  • The answer arrives word by word, with its sources linked.

Fuentia

  • We find the 30 chunks closest in meaning and the 30 that best match the words, and fuse both lists.
  • The model sees only the best 8 and must cite the chunk behind each claim. If it is not in your documents, it says it does not know.
Fig 3Two searches, one list
The chosen card and its placeBoth lists and the tray Each chunk scores 1 / (60 + position) for every list it appears in. The 8 with the most points go to the tray and are cited as c1, c2… If none is relevant, the answer is “I don’t know”.

Step 4

Integrate it

Take the assistant where your users are: your backend, your app or your website.

You

  • From your server, with a notebook secret key.
  • On your website, by pasting one line with a public key that only works on your domains.

Fuentia

  • The secret key never leaves your server. The public key only opens sessions from the origins you allow, and every session expires after 15 minutes.
  • We cap usage per visitor (10 questions a minute) and per key (300 an hour) to protect your quota.
Fig 4Secret key and public key
The notebook’s API and what it lets inYour server and the visitor’s website The secret key stays on your server. The public key only opens a session from the origins you allow, and the session expires after 15 minutes.

Try it here

This is how a notebook with Fuentia’s documentation answers. To try it live, enable the demo in your local install.

You

How long until a document is ready, and how large can it be?

Assistant

You can upload TXT, Markdown, and PDF files of up to 20 MB1. They are processed in the background: the panel shows their status and they turn “ready” as soon as they are chunked and indexed2.

Sources

  1. 1uploading-documents.md
  2. 2document-states.md

To enable it: run pnpm site:demo-seed and reload.

For online shops

An assistant that knows your catalog

Import your products by API, with their images and prices. The same widget recommends products on your home page and answers questions on each product page.

  • Catalog by API

    Up to 500 products per request. Import again and only what changed is reindexed.

  • Recommendations in a carousel

    It searches the catalog by meaning and by words. The AI picks and explains, but can only show products that exist.

  • Chat on the product page

    It answers with the product sheet, its documents and the shop’s policies; never with another product’s.

Example with the demo shop. To try it live: pnpm site:demo-seed and reload.

I need running shoes for the road

I’d go for the Trotta Nimbus, highly cushioned for long runs; the Trotta Tempo, light for intervals; and the Vela Brisa, stable if you are starting out.

  • Trotta Nimbus€129.90 In stock
  • Trotta Tempo€119.90 In stock
  • Vela Brisa€99.90 In stock
  • See more results

Two ways to integrate

Copy, paste and swap in your own key. You will find them in the panel, on each notebook.

Widget on your website

One script, no separate CSS. It only works on the origins you allow.

<script src="http://localhost:3000/widget/v1.js" async></script>
<script>
  window.RagSearch = window.RagSearch || { q: [] };
  RagSearch.q.push(['init', { publishableKey: 'pk_live_YOUR_PUBLIC_KEY' }]);
</script>

API from your server

Streaming answer (SSE) with text and citation events.

curl -N http://localhost:3000/v1/notebooks/<NOTEBOOK_ID>/chat \
  -H "Authorization: Bearer <YOUR_SECRET_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"message": "What is the returns policy?"}'

Your first assistant, today

Create a notebook, upload a document and ask the first question in a few minutes.

Sign in to the panel