News Center

We are excited to unveil the latest addition to our model collection: o3 on Azure! This advanced reasoning model, now available as an Azure-hosted option on Dataster, brings unparalleled capabilities to your AI endeavors. Whether you're tackling complex problem-solving, analyzing visual data, or exploring new horizons in coding and science, o3 is here to elevate your work. As always, you have the flexibility to bring your own Azure-hosted o3 model if you prefer. Sign in, send some tokens, and experience the powerful capabilities of o3 today!

We are delighted to unveil the latest gem in our model collection:anthropic.claude-3-7-sonnet-20250219-v1:0 on AWS! This sophisticated model, now available as an AWS-hosted option on Dataster, brings the timeless beauty and structure of sonnets to your language processing endeavors. Whether you're composing poetry, generating eloquent content, or exploring new creative horizons, Claude 3.7 Sonnet is here to elevate your work. As always, you have the flexibility to bring your own AWS-hosted Claude 3.7 Sonnet if you prefer. Sign in, send some tokens, and experience the poetic mastery of Claude 3.7 Sonnet today!

We are thrilled to announce the launch of our latest feature: Model End of Life Planning! This tool empowers builders to seamlessly navigate the retirement of AI models by identifying potential successors and evaluating them with ease. Say goodbye to the uncertainty and disruption that often accompanies model deprecation. With Model End of Life Planning, you can proactively plan for the future, ensuring your projects remain cutting-edge and efficient. Get ready to future-proof your AI endeavors and experience the peace of mind that comes with strategic planning!

We're excited to introduce the latest addition to our model catalog: claude-3-5-haiku-v1:0! This innovative model, now available as an AWS-hosted option, brings the elegance and precision of haiku to your language processing needs. Whether you're crafting poetry, generating concise content, or exploring new creative possibilities, Claude 3.5 Haiku is here to inspire. Sign in, send some tokens, and experience the artistry of Claude 3.5 Haiku today!

We are thrilled to unveil the Latency Cost/Performance Indicator, a new feature designed to help builders make informed decisions about the cost-efficiency of their models and RAGs. This graph plots each model and RAG according to its relative latency index and cost index, derived from specific use case token usage and public pricing. By visualizing the trade-off between latency and cost, builders can determine whether investing in a more expensive model is justified by the latency improvements it offers. This enhancement ensures that resources are allocated effectively, maximizing the performance and economic viability of your Generative AI applications.

In line with our recent enhancements for latency jobs, we've also improved the visualization of human evaluation job results. Now, you can see each LLM, RAG, system prompt, and their combinations displayed on a bar chart, making it easier to assess output quality.

We are excited to announce a new feature for Latency jobs! For each LLM, RAG, System prompt, and their combinations, we now provide a graph that plots the median latency, as well as the minimum, maximum, and first and third quartiles. We are confident that this enhancement will empower builders to make the best data-driven decisions for their Generative AI applications.

We are thrilled to introduce a powerful new model to our catalog! With the addition of the latest GPT-4o variant, known as gpt-4o-2024-11-20, our users can now evaluate whether updating their version of GPT-4o will result in improvements or regressions for their specific use cases.

We are excited to introduce another vector store option for our users to build RAG applications and evaluate their accuracy or latency. Previously, users could only build RAGs with grounding data stored in Azure AI Search. Now, with the addition of Amazon OpenSearch Serverless, we offer a compelling new option for RAG applications. You can now create a fully AWS-hosted RAG using Titan Text Embeddings V2, OpenSearch Serverless, and Claude 3.5.

To empower our users to build and test their RAGs entirely on AWS, we are thrilled to announce support foramazon.titan-embed-text-v2:0. Previously, only the latest OpenAI embedding models were supported. This update underscores our commitment to being truly multi-cloud and providing our users with a wide range of options for their GenAI applications.

We are dedicated to creating the most versatile and feature-rich prompt gallery available. Now, you have the power to re-order your prompts, allowing you to present them in the sequence that best suits your needs, rather than the order they were created in. Experience the freedom to customize your workflow like never before!

In the realm of LLM and RAG output evaluation by humans, minimizing bias is crucial, if not eliminating it entirely. It's easy to develop a preference for a specific model or provider, but this can skew the objectivity of the evaluation process. Dataster offers robust bias removal mechanisms for human evaluation. Historically, it has been possible to mask the model, RAG, and even the system prompt used to generate a particular output. With today's release, we introduce an additional layer of bias reduction by randomizing the order in which input and output pairs are presented to evaluators. This makes it virtually impossible to determine which model, RAG, or system prompt produced the output until the entire evaluation is complete. Happy blind testing!

We're continuing our Large Language Model catalog expansion with the addition of cohere.command-r-plus-v1:0 as an AWS hosted option; bring your own or use ours. Sign in, send some tokens and try it out!

Last month, we introduced our first A21 Large Language Model with Jamba 1.5 Mini. Building on this momentum, we are excited to announce the addition of ai21.jamba-1-5-large-v1:0 to our catalog! This model is now available as an AWS-hosted option. As always, you have the flexibility to bring your own AWS-hosted Jamba 1.5 Large if you prefer. Give it a try and see what it can do!

For specific sensitive use cases, users may want to delete their prompts from the database. We're introducing a feature that allows for the easy deletion of a large collection of prompts from the catalog and the underlying data store, while retaining the ones you wish to keep. Deleting a prompt or bulk deleting prompts will effectively remove all text and associated metadata from our database instantly. However, prompt content may still appear in our secure logs for up to seven days.

Today, we're thrilled to unveil the latest addition to our model catalog: ai21.jamba-1-5-mini-v1:0! This cutting-edge language model is now available as an AWS-hosted option. Plus, you still have the flexibility to bring your own AWS-hosted Jamba 1.5 Mini if you prefer. Dive in and explore the possibilities!

We're excited to announce the addition of another powerful language model to our collection: gpt-4o-mini-2024-07-18. As always, you have the flexibility to bring your own OpenAI GPT-4o mini if you prefer.

Dataster has a human evaluation feature where users can rate a list of completions as good or bad, or alternatively, on a scale from 1 to 5. We are excited to release a new capability in beta that automates the binary evaluation process. Dataster can now automatically rate all your candidate models and RAGs against your use case, helping you make the best decisions when building your GenAI-backed application.

We are adding the capability to optionally save a ground truth answer corresponding to a user question in a user prompt. This will underpin our upcoming Auto Eval feature, which will allow you to launch automated evaluation jobs to determine how your RAGs perform against a large and representative set of user questions. This will supplement the existing Human Eval feature, which lets users manually assign grades to each response and rank their models and RAGs for a specific use case.

We are adding support for gpt-4o-2024-08-06 hosted on OpenAI. This model was previously supported on Dataster only as an Azure OpenAI hosted model. Users can now compare the model’s performance across both providers. With this release, Dataster now supports 10 models across three different providers.