Glossary
The terms used across Avra's documentation, what each one means, and where to read more.
The terms you’ll meet across these docs, in alphabetical order.
| Term | What it means | Read more |
|---|---|---|
| Adaptation | Self-supervised learning of a copy of Avra RFM on your database. The copy is exclusive to your workspace | Adaptation and fine-tuning |
| Alias | A name, such as production, that points at one version of a model. Reassigning it promotes or rolls back without code changes |
Model versioning |
| API key | The credential your systems send as a bearer token. Each key belongs to one workspace | API keys |
| Avra Core | Delivers Avra’s models into your systems: live predictions, batch inference, embeddings, and the controls around them | Avra Core |
| Avra GFM | The Graph Foundation Model. Pre-trained on Avra’s knowledge graph of the Brazilian economy, it returns embeddings and probabilities of default for Brazilian companies, from Avra’s cloud only | Avra GFM |
| Avra RFM | The Relational Foundation Model. Pre-trained on diverse relational data, it predicts over your database without a training project | Avra RFM |
| Backfill | An ingestion run over a window of past snapshot times, which loads history without moving the sources’ regular checkpoints | Time and history |
| Backtest | Predictions made at past dates and scored on what happened after them | Evaluation |
| Batch | An asynchronous job that scores every record in the files you upload | Batch inference |
| Bitemporal | Keeping two times for every row: when it was true and when it was recorded | Time and history |
| Checkpoint | The position each ingestion source keeps, so a run reads only what arrived since the last one | Data ingestion |
| CNPJ | The Brazilian company registration number, and Avra GFM’s input | Avra GFM |
| Correlation ID | The UUID you send in the X-Correlation-Id header to trace a request |
Live predictions |
| Data contract | The pipeline definition agreed with your Avra team: the sources, their keys and time columns, the checks, and the output | Data contracts and checks |
| Data layer | The Helm charts that run the ingestion pipeline and the Avra agent in your Kubernetes cluster | Install the data layer |
| Embedding | A vector that represents a company, from Avra GFM. It has 1024 dimensions and can be shortened | Build with embeddings |
| Event time | When a row was true at the source: when the payment happened or the address took effect | Time and history |
| Fine-tuning | Learning one task in a copy of Avra RFM. The copy is exclusive to your workspace | Adaptation and fine-tuning |
| Horizon | The end of a prediction’s window: when its answer is known | What you can predict |
| Identity key | The column or columns that identify one row of a table you send | Preparing your data |
| In-context prediction | Avra RFM reads your history as context when a prediction is made. No weights change | How predictions work |
| Ingestion pipeline | How your data reaches Avra’s models: checked against your data contract, then published as a new version | Data ingestion |
| Labelled example | A past instance of the same prediction, with its known outcome, read as context | How predictions work |
| Leakage | Information from after a decision reaching the prediction about it | Time and leakage |
| Model version | One immutable version of a model, with its own model_version_id. Only active versions serve predictions |
Model versioning |
| Organization | Your company’s account in Avra. It holds your workspaces, and models and roles apply across it | Dashboard |
| Prediction date | The date a prediction is made for. It sees only what was known by then | What you can predict |
| Probability of default (PD) | Avra GFM’s credit outputs, at five horizons: pd_30d, pd_60d, pd_90d, pd_180d, and pd_365d |
Credit |
| Quality gate | The ingestion stage that runs your contract’s checks before anything is published | Data contracts and checks |
| Risk band | A group of records whose outcomes stay distinct and stable, for setting policy | Risk bands |
| Score | What Avra RFM returns for a yes-or-no outcome. It ranks and compares records, and isn’t a calibrated probability | What you can predict |
| Shadow | Running a new version beside the current one on live traffic, with no effect on decisions | Evaluation |
| Snapshot time | When your systems recorded a row: the capture of the source it came from | Time and history |
| Webhook | A notification Avra sends to your endpoint when an event happens, such as a batch ending | Webhooks |
| Workspace | An isolated area of your organization that holds API keys, batches, files, and logs | Dashboard |