---
title: Build with embeddings
description: Add the relationships to the models you already run with Avra's embeddings.
sidebar:
  label: Overview
---

## What embeddings are

An embedding is a list of numbers that describes a company and the context around it. Avra's embeddings come from [Avra GFM](/gfm): **1024-dimensional vectors** for Brazilian companies that capture network position, relationships, behavior, and how a company is changing over time.

They join your feature table like any other columns. Your model, your governance, and your monitoring stay as they are; they just see what your own rows can't show.

```mermaid
flowchart TB
  gfm["Avra GFM"] -->|"1024-dimensional vectors,<br/>live or batch"| store["Your warehouse or<br/>feature store, with version_id"]
  store --> table["Your feature table"]
  features["Your own features"] --> table
  table --> model["Your model"]
  class model accent
  classDef accent fill:#0A01FE,stroke:#0A01FE,color:#fff
```

:::info
Embeddings are the route for teams that run their own models. If you'd rather get scores than build a model, use [live predictions](/core/predictions) or [batch inference](/core/batch-inference) instead.
:::

## What they add

Embeddings carry signals that are hard to engineer by hand, without building a graph pipeline of your own: ownership networks and the patterns in them, regional and sector clusters, and each company's trajectory over time.

As data about a new company accumulates, its position in the embedding space can change. Even with little data at first, it can be placed near similar companies.

> **Thin-file ready**
>
> New and sparse companies still get a representation from the structure around them.

> **Shared across teams**
>
> Risk, growth, fraud, and analytics teams can use the same embeddings without rebuilding pipelines.

> **Your data stays yours**
>
> Your model trains on your side. Avra GFM is never trained on your data.

## What you receive

| Asset | Details |
|-------|---------|
| **Vectors** | 1024 dimensions per company; shorten them client-side to 512, 256, 128, 64, 32, or 16 |
| **Version** | Each vector comes from the version you call; store its `version_id` with the vector for provenance and reproducibility |
| **Notifications** | Webhooks when a model version is promoted, so you know when to refresh |

## Delivery

**Live API**

Synchronous retrieval for onboarding, underwriting, and interactive analytics. See the [technical guide](/gfm/embeddings/technical-implementation).

**Batch**

Embed large lists of companies asynchronously. See [Batch inference](/core/batch-inference).

**Feature store**

Persist embeddings in your warehouse or feature store, with their model version, for training and experiments.

## When to use embeddings

Use embeddings when your team already runs models and wants them to see relationships, when you need signal for companies with little history in your own data, or when several teams need a common representation without duplicating work.

## Get started

1. Read the [technical guide](/gfm/embeddings/technical-implementation) for specifications and integration patterns.
2. Browse the [applications](/gfm/embeddings/embedding-applications).
3. Plan version changes with [Model versioning](/core/model-versioning).
