---
title: Adaptive Embeddings
description: How our embeddings automatically adapt to your model's complexity without losing predictive power.
---

Avra delivers embeddings as 1024-dimensional vectors from the Graph Foundation Model. Each vector is **adaptive**: you can slice it client-side to smaller dimensions (512, 256, 128, 64, 32, 16) while preserving semantic meaning and predictive power. One API call, one storage footprint, every dimension you need — no retraining, no separate endpoints.

## Why adaptive embeddings
- **One training pass, many footprints** — derive lighter embeddings for simpler models without retraining.
- **Consistent semantics** — lower-dimensional slices preserve the ordering and relative distances of the full vector.
- **Compose with your RFM** — workspace-specific representations update every slice simultaneously.

## Selecting a dimensionality

**1024 dimensions**

Maximum fidelity for deep learning, semantic search, or graph analytics where
latency and storage are secondary.

**512 dimensions**

Balanced trade-off between richness and compute for most production
workloads.

**256 dimensions**

Optimized for real-time scoring and streaming analytics on CPU-bound services.

**128 dimensions**

Ideal for dashboards, feature stores, and propensity models where storage is a
concern but accuracy must remain high.

**64 / 32 / 16 dimensions**

Reserve for edge deployments or extreme latency constraints. Treat the size as
a hyperparameter—step down gradually while monitoring KPIs.

:::info
All Avra API calls return the full 1024-dimensional embedding. Refer to the [API Reference](/api-reference) for examples of slicing these vectors client-side to your desired dimensionality.
:::

## Best practices
- Treat dimensionality as a hyperparameter: start at 256-d for tree-based models (XGBoost, LightGBM), evaluate performance, then adjust up or down based on validation metrics.
- Normalize vectors before computing cosine similarity or feeding them into distance-based models.
- Cache embeddings alongside the model snapshot ID so experiments remain reproducible.
- Avoid applying additional dimensionality reduction for production workloads—it degrades the learned semantics. Reserve techniques like PCA or t-SNE for visualization only.

## Trajectory signals (temporal dynamics)

Embeddings aren’t static. As new signals arrive, an entity’s position moves—its “trajectory.” Reading this movement adds foresight to your models without changing your integration.

- Direction of change: is the entity moving toward patterns historically linked to risk or growth?
- Magnitude vs. noise: we compare movement to typical volatility to avoid false alarms.
- Aggregation: roll up movements at portfolio, sector, or region level to spot drift early.

How to use
- Credit: trigger reviews when customers move toward risky clusters before KPIs deteriorate.
- Growth: prioritize accounts whose movement matches your best customers’ pre‑expansion patterns.
- Ops: watch suppliers whose trajectories resemble prior disruption events.

Access
- Returned as movement indicators with relevant products (for example, Credit Score) and accessible via embeddings over time for custom analytics.