---
title: Avra RFM
description: The foundation model for relational data. Pre-trained on diverse relational data, it reads your database as it is and predicts over it without a training project.
sidebar:
  label: Overview
---

Avra RFM, the Relational Foundation Model, is pre-trained on diverse relational data. It reads a company's database as it is — tables linked by keys, any schema — and predicts over it without a training project.

It's the center of Avra. [Avra Core](/pt-BR/core) delivers it into your systems, and [data ingestion](/pt-BR/data-ingestion) gets your data to it.

<svg class="fig" viewBox="0 0 640 210" role="img" aria-label="On the left, your database: customers, orders, and payments tables linked by keys. In the middle, Avra RFM. On the right, scores delivered through live predictions and batch inference.">
  <defs>
    <marker id="fig-rfm-arrow" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
      <path d="M0,0 L10,5 L0,10 z" class="arrow" />
    </marker>
  </defs>
  <text class="small title label" x="24" y="30">Your database</text>
  <rect class="card" x="24" y="44" width="150" height="34" rx="6" />
  <text class="mono" x="40" y="66">customers</text>
  <rect class="card" x="24" y="94" width="150" height="34" rx="6" />
  <text class="mono" x="40" y="116">orders</text>
  <rect class="card" x="24" y="144" width="150" height="34" rx="6" />
  <text class="mono" x="40" y="166">payments</text>
  <path class="edge-accent" d="M158,78 v16" />
  <path class="edge-accent" d="M158,128 v16" />
  <circle class="solid" cx="158" cy="78" r="3" />
  <circle class="solid" cx="158" cy="94" r="3" />
  <circle class="solid" cx="158" cy="128" r="3" />
  <circle class="solid" cx="158" cy="144" r="3" />
  <line class="edge" x1="184" y1="111" x2="242" y2="111" marker-end="url(#fig-rfm-arrow)" />
  <rect class="solid" x="248" y="80" width="144" height="62" rx="10" />
  <text class="mid on-solid title" x="320" y="108">Avra RFM</text>
  <text class="small mid on-solid" x="320" y="127">in context</text>
  <text class="small mid label" x="320" y="164">pre-trained on diverse</text>
  <text class="small mid label" x="320" y="180">relational data</text>
  <path class="edge" d="M392,104 C420,104 424,76 450,76" marker-end="url(#fig-rfm-arrow)" />
  <path class="edge" d="M392,118 C420,118 424,146 450,146" marker-end="url(#fig-rfm-arrow)" />
  <text class="small title label" x="456" y="30">Scores in your systems</text>
  <rect class="soft" x="456" y="56" width="160" height="40" rx="8" />
  <text class="small title" x="472" y="74">Live predictions</text>
  <text class="small label" x="472" y="89">one record, by API</text>
  <rect class="soft" x="456" y="126" width="160" height="40" rx="8" />
  <text class="small title" x="472" y="144">Batch inference</text>
  <text class="small label" x="472" y="159">a whole portfolio</text>
</svg>

Your tables go in as they are, and scores for your records come out, live or in batch. No features to build and no model to train for each question.

## Learn the concepts

**[Relational data](/pt-BR/rfm/relational-data)**

How the model reads your tables, keys, names, and time

**[What you can predict](/pt-BR/rfm/what-you-can-predict)**

Frame a prediction: the records, who's in, the outcome, and the window

**[How predictions work](/pt-BR/rfm/predictions)**

What the model reads for each prediction, and what it never sees

**[Adaptation and fine-tuning](/pt-BR/rfm/adaptation)**

When a copy of the model learns from your data

**[Evaluation](/pt-BR/rfm/evaluation)**

Backtest, shadow, and promote: how every model earns its place

**[A worked example](/pt-BR/rfm/worked-example)**

One prediction from start to finish, from the data to a weekly list

For how the RFM compares with language models and what it needs from your data, see the [FAQ](/pt-BR/resources/faq).

## What it reads

The RFM reads your tables as they sit in your systems, and the relationships come from their keys, inferred from the data or declared. Table names, column names, and text values are read as language, Portuguese included.

Each table also says what its time means. Some rows have no time, like a product catalog. Some are events, which happened at a moment, and some are records that became known at a moment. See [Relational data](/pt-BR/rfm/relational-data).

## What it returns

A score for each record you ask about, for a yes-or-no outcome, or an estimate, for a number: each customer, account, or company in the question. Scores rank and compare records; they aren't calibrated probabilities. See [What you can predict](/pt-BR/rfm/what-you-can-predict).

## Three ways to use your data

| Mode | What happens | Effect on the model |
|------|--------------|---------------------|
| **In-context prediction** (default) | Your history is read as context when a prediction is made | No weights change |
| **Adaptation** | Self-supervised learning on your database | A copy exclusive to your workspace |
| **Fine-tuning** | Learning one particular task | A copy exclusive to your workspace |

See [Adaptation and fine-tuning](/pt-BR/rfm/adaptation).

## Where it runs

In Avra's cloud, Avra serves the model in your isolated workspace, and you reach it through Avra Core: the API and batch.

In your AWS account, the model runs on Amazon SageMaker as an AWS SageMaker algorithm, available on AWS Marketplace. Its weights are never delivered as usable files. See [Avra RFM on AWS SageMaker](/pt-BR/deployment/aws-sagemaker).

## What stays private

Avra doesn't publish the model's architecture, pretraining data, or benchmark claims. Every evaluation runs on your data, against the model you run today, and the results are shared with you. See [Evaluation](/pt-BR/rfm/evaluation).
