Embeddings On Demand with Embedding Export

Today, LGND subscribers can directly access raw geo-embeddings via S3 buckets.

Previously, LGND-generated embeddings lived inside our platform, where they could be searched and compared using our custom methods. We recognize, however, that some technical users may prefer to reference embeddings in their own modeling stack – now they can.

How It Works

LGND now automatically provisions a dedicated, LGND-owned S3 bucket for all Developer tier accounts. Embedding collections may be exported to this bucket as hive partitioned geoparquet datasets optimized for use with cloud-native clients like DuckDB, GeoPandas, or Sedona. The parquet files contain the embeddings themselves and several top level metadata fields like the geometry and image acquisition date. Access follows a bring-your-own-credentials model: provide your IAM role or user ARN and LGND configures the S3 bucket and KMS key policies on your behalf.

Visit our developer resources repo for examples.

Why Direct Access Matters

The LGND API emphasizes abstraction for most common workflows: search by example or text, find similar locations. Still, some workflows need to operate on them as raw vectors. 

Here’s what we’ve heard from users:

• Continuous regression. Predicting above-ground biomass, crop yield, soil moisture, or population density isn't a similarity problem — it's a regression problem. Direct embedding access lets you train regression heads against your ground-truth targets.
• Custom feature engineering. Clay geo-embeddings are 1024-d. For some, that’s too much information. With raw embeddings, users can customize embedding dimensions to reduce data size and focus on the vector indexes that matter most.
• Model fusion. Concatenating embeddings across models or across spatial and temporal lags can help unlock new insights and model uplift.  

Get Started

Embedding export is available now for all subscribing users. Not subscribed yet and want to evaluate it? Write to us at sales@lgnd.ai.