---
title: FAQ
description: Common questions about Avra RFM, the data it needs, and running Avra, with short answers and links to the full explanation.
---

## The model

### How is Avra RFM different from a language model?

Language models are built for text. A prediction about a customer depends on the tables, keys, and timestamps that describe it, and a language model doesn't predict over a relational database. Avra RFM reads those tables directly, follows their keys, respects their time, and returns a score for each record. It doesn't generate text or write queries. See [Why Avra](/pt-BR/why-avra).

### How is it different from training a model on my data?

A traditional model needs a project for every question: build a dataset, engineer features into one wide row per record, train, and evaluate. Avra RFM reads your tables as they are and predicts in context, so a new question means framing it and evaluating it on your history, with no training run in between. Where an evaluation shows it pays, a copy can still be adapted to your database or fine-tuned on a task. See [How predictions work](/pt-BR/rfm/predictions).

### How is Avra RFM different from Avra GFM?

| | Avra GFM | Avra RFM |
|---|---|---|
| **Reads** | Avra's knowledge graph of the Brazilian economy | Your relational database, any schema |
| **You receive** | Embeddings and probabilities of default | Scores and estimates |
| **Runs in** | Avra's cloud only | Avra's cloud, or your AWS account |
| **Specializing it** | Separate downstream models on its outputs | Optional adaptation or fine-tuning of a copy |

See [Models](/pt-BR/models).

### Are RFM scores probabilities?

No. RFM scores rank and compare records; they aren't calibrated probabilities. Set thresholds and bands from your own outcomes. Avra GFM's PDs are probabilities of default. See [Risk bands](/pt-BR/use-cases/credit-intelligence/homogeneous-groups).

### Does Avra train on my data?

In-context prediction changes no weights. Adaptation and fine-tuning produce copies that are exclusive to your workspace and never merged back into a shared model. The shared RFM is pre-trained on diverse relational data, and Avra GFM on public, private, and licensed sources; no customer data goes into either. See [Data privacy & compliance](/pt-BR/security/data-privacy-and-compliance).

### Can I get the model's weights?

No. An adapted or fine-tuned copy is derived from Avra's model and licensed for use in your workspace. Its weights are never delivered as usable files, and in your AWS account they stay encrypted. See [Adaptation and fine-tuning](/pt-BR/rfm/adaptation).

## Your data

### What data does Avra need?

The records you want to score and the tables related to them, linked by consistent keys; for every row, when it was true and when your systems recorded it; your historical outcomes, each with its date; and enough history to evaluate a model on periods it didn't see. See [Preparing your data](/pt-BR/data-ingestion/preparing-data).

### Do I need to build features?

No. Send the tables as they are. The model reads the rows connected to each record across your tables, so there's no wide feature table to build. See [Relational data](/pt-BR/rfm/relational-data).

### How are missing values handled?

A missing value is simply left out instead of being filled in, so there's nothing to impute first. A missing key is different: a row whose key doesn't resolve breaks a relationship, and the ingestion pipeline's relationship check catches it when your contract declares it. See [Data contracts and checks](/pt-BR/data-ingestion/data-contracts).

### Does it read Portuguese?

Yes. Table names, column names, and text values are read as language, Portuguese included, so columns don't need renaming. See [Relational data](/pt-BR/rfm/relational-data#names-text-and-values).

### How much history do I need?

Enough past outcomes, over enough periods, to evaluate a prediction on dates it didn't see. How much that is depends on how often your outcome happens and on its window. Your Avra team checks it with you during data discovery.

### Which file formats can I send?

Parquet, CSV, or JSON, uploaded in the dashboard or read from an Amazon S3 or Google Cloud Storage bucket. Parquet fits best. CSV files need a header row. See [Preparing your data](/pt-BR/data-ingestion/preparing-data#file-formats).

### What if my data arrives late or gets corrected?

Send it with its real event time and the snapshot it arrived in. The ingestion pipeline keeps every version of a row, and a prediction only sees what had been recorded by its date, so late rows and corrections never leak into past predictions. See [Time and history](/pt-BR/data-ingestion/time).

## Running Avra

### Where does Avra run?

In Avra's cloud, or in your own environment: the data layer in your Kubernetes cluster, and Avra RFM in your AWS account on Amazon SageMaker. Avra GFM runs in Avra's cloud only. See [Deployment](/pt-BR/deployment).

### How do predictions reach my systems?

Through the API, one record at a time, when a decision comes up; through batch inference, for a whole portfolio on your schedule; or as embeddings, as features in the models you already run. In your AWS account, through SageMaker jobs. See [Avra Core](/pt-BR/core).

### How do I roll back a model?

Reassign the alias, such as `production`, to the previous version in the dashboard. Every version is immutable, and the change takes effect on the next call. See [Model versioning](/pt-BR/core/model-versioning).

### What happens when a delivery fails a check?

If the check blocks, the run is rejected with the reason, nothing invalid is published, and the previous version stays current. If it only alerts, the violation is reported and the delivery goes through. See [Data contracts and checks](/pt-BR/data-ingestion/data-contracts#when-a-check-fails).

### How do I get help?

Write to [support@avra.ai](mailto:support@avra.ai), and follow live status and incidents at [status.avra.ai](https://status.avra.ai). Report a vulnerability privately to `security@avra.ai`. See the [Vulnerability disclosure policy](/pt-BR/security/vulnerability-disclosure-policy).
