AI Accelerator - Pipelines 1.0.0 release notes

Suggest edits

Released: 20 November 2024

This is the GA release of EDB Postgres AI - AI Accelerator - Pipelines. Since the technical preview, a number of enhancements have been made.

Highlights

  • Integrated model runtime.
  • Support for external models with OpenAI API.
  • Secure management of API credentials via postgres "user mapping" feature.
  • Model registry that allows configuring internal/external models.
  • pgfs/volumes allows working with external data from S3 object stores and file systems.
  • Low-level primitives for models: encode, transform for image and text data.
  • Enhanced retriever concept for data retrieval.
  • Support image and text data, including: Postgres tables, external storage, and ad-hoc queries.
  • Support for building fully custom AI workflows/pipelines.

Enhancements

DescriptionAddresses
Integrated model runtime

The tech preview had a separate model runtime. Now, the model runtime is integrated into the product.

Support for external models with OpenAI API

The tech preview only supported internal models. Now, external models can be used with OpenAI API.

Secure management of API credentials via postgres "user mapping" feature

Improved from the tech preview which had no secure credential management.

Model registry that allows configuring internal/external models

A new concept in Pipelines over the tech preview which had only retrievers.

PGFS allows working with external data from S3 object stores and file systems

The tech preview could only work with S3 in retrievers. Now, the volume concept allows working with external data in general.

Low-level primitives for models: encode, transform for image and text data

Low-level primitives for modes have been added to complement the existing high-level functions.

Enhanced retriever concept for data retrieval

Retriever concept has been enhanced:

  • Calls can now return rich table results with data such as vector distance.
  • Flexible interface allows "retrieving" just the key or the source data.
  • Enables integration with custom workflows/pipelines.
Support image and text data

Images can be retrieved from

  • External storage or Postgres tables.
  • Images can also be retrieved from "bytea columns" in Postgres tables
  • Also able to be retrieved ad-hoc in a query.
Support for building fully custom AI workflows/pipelines
  • direct SQL access to models (encode, transform, …) even allowing batch-execution
  • flexible interfaces in our high-level composite functions (like retrieve)
  • allowing direct access to all low-level interfaces that we use in high-level functions; such as exposing functions like "embed_single_row"
← Prev

EDB Postgres AI - AI Accelerator 1.0.0 release notes

↑ Up

EDB Postgres AI - AI Accelerator 1.0.0 release notes

Next →

Open Source Licenses for software incorporated in EDB Postgres AI - AI Accelerator - Pipelines

Could this page be better? Report a problem or suggest an addition!