10.03.2024

New Releases, Enhancements, + Changes

What's New

Embedding Traces

With Embeddings Tracing, you can effortlessly select embedding spans and dive straight into the UMAP visualizer, simplifying troubleshooting for your genAI applications. How this works:

  • Users can now select embedding spans and go to the embedding visualizer (also available on project level nav for all generative models)

  • All embedding spans get pulled over for UMAP + clustering

  • Users can select an individual point/cluster and the embedding span attributes get pulled over

  • "Color by" span attributes is also available, and whichever attribute users "color by" can also be viewed as a column in the table

Note: this functionality is only viewable for embedding spans after October 2, 2024.

Learn more about our embeddings visualizer here.

Experiments Details Visualization

Users can now view a detailed breakdown of labels for their experiments on the Experiments Details page.

Note: this feature is currently only available for label-based evaluations. Support for score-exclusive evaluations will be coming in a future update.

Enhancements

Prompt Playground Improvements

  • Full OpenAI Models Support: We've added full support for all available OpenAI models in the playground including the o1-mini and o1-preview.

  • OpenAI Function/Tool Calls: With OpenAI's addition of function calling in their models, we have added support in Arize starting from the trace all the way to the playground. View function/tool calls in your traces, and open the playground to further experiment and test.

  • Full-screen Data Mode: Enter full screen mode to view your data more easily.

  • Features For Datasets:

    • Prompt Overriding: When using a dataset to test out different prompts, users can now replace the prompt template with the dataset.

    • Pop Up Windows for Long Outputs & Variables: Viewing longer outputs is now easier with pop up windows.

Along with these features, we've added improvements such as better input variable behavior, autocompletion enhancements, support for mustache/f-string input variables, and more.

Filters Updates

  • Filter History: We now store the last three filters used by a user! Users can easily access their filter history in the query filters dropdown, making it simpler to reuse filters for future queries.

Traces Page Updates

  • Parent Spans: We’ve enhanced the Traces page to give users a better experience when filtering spans. Now, even if the filter doesn’t match the parent span, users will still see the the parent span with the relevant spans nested under it.

  • Quick Filters: We’ve introduced quick filters, allowing users to apply filters directly from the table by hovering over the text to reveal the filter icon.

New Arize-Otel Package for LLM App Tracing

We heard you, and we made it much simpler to add automatic tracing to your applications! It's now just a few lines of code to use OpenTelemetry to trace your LLM application. Check out our new quickstart guide which uses our arize-otel package.

Workflow Improvements

  • Support for new line creation while reviewing prompts.

  • Easily add spans to a dataset from the Traces page using the "Add to Dataset" button.

  • Quickly create an evaluation task within a trace if no evaluations currently exist using the "Setup Task" button.

📚 New Content

The latest video tutorials, paper readings, ebooks, self-guided learning modules, and technical posts:

🧑‍⚖️ Selecting the Right Model for LLM-as-a-Judge Evaluations

🤖 Leveraging Agent Evaluation and Memory to Build Robust Agentic Systems

🔭 Exploring OpenAI’s o1 Models

🪞Breaking Down Reflection Tuning

Last updated