Acceldata
AIO

Last updated: Nov 17, 2026 08:34 UTC

Version 26.11.0

Date: 17th November 2026

What's New

This section consists of the new features and enhancements introduced in this release.

Tracing

  • Tags on Traces, Sessions, and Spans: AI Observability (AIO) now supports key:value tags on traces, sessions, and spans. You can set tags in code with aio.set_tags() or aio.using_tags(), turn span attribute keys into tags on newly ingested spans in the project's Settings > Attribute Tags, or add a tag manually with Add tag on a trace, session, or span. Tags appear as chips on trace and session pages and in the span pane. Use the Tags filter on the Traces, Sessions, and Spans pages to show rows that have any of the selected tags, or select Exclude these to hide them. Adding or removing manual tags and turning attribute tags on or off requires permission to edit the project. For more information, see Tag Traces.
  • Infrastructure Traces Hidden by Default: The Traces page now hides HTTP and MCP infrastructure traces by default, so your application's own traces appear first. For more information, see Explore Traces.
  • Trace and Session Owner: Trace and session pages now show who started the trace or session and when. If no user is recorded, only the start time appears. For more information, see Explore Traces.

Data Quality

  • Assets Used in Traces: AIO now identifies the tables, files, and endpoints that each trace reads or writes and matches them to assets in the ADOC catalog. AIO detects database assets from the SQL that your application runs, and you can declare other assets in code with aio.declare_assets(). Trace and session pages show each asset and its data quality policy results in an Assets used panel, the Traces, Sessions, and Spans pages include an Assets column, and the Traces page includes an Assets filter. If your application connects to a database at a different address than the one registered in ADOC, such as localhost through a tunnel, map the address with data_source_hosts in aio.init() or with the AIO_DATA_SOURCE_HOSTS environment variable. For more information, see Record the Data Your Agent Uses.
  • Quality Coverage: The new Quality coverage tab in a project ranks the assets that the project's traces used, shows which of them data quality policies check, and lists the columns that no policy checks. For more information, see Check Quality Coverage.
  • Add Quality Checks with ADM: Select Add quality checks with ADM on an asset that isn't checked in the Quality coverage tab to get a prompt that you paste into ADM chat. ADM then drafts policies for you to review. For more information, see Check Quality Coverage.

Overview

  • Home Dashboard: AIO now opens on Home, a summary of every project you can access. Opening AIO or selecting the Acceldata logo takes you there, and you can still reach the Projects list from the navigation menu or the Projects breadcrumb. Home shows cost, tokens, LLM calls, models, sessions, and users, compares cost with the previous period of the same length, and includes usage over time, rules, and failed traces. Filter Home by project, user, and a time range of up to 90 days, and select a row to open the matching project, session, or trace. For more information, see Home Dashboard.
  • Open AIO from ADM Chat: After you ask a question in ADM chat, you can open AIO from a link in that conversation.
  • Navigation Improvements: Breadcrumb items that are cut off now show their full label in a tooltip, shortened IDs show the full ID, the back arrow shows Back to <parent>,, and the user menu shows your name or email address. For more information, see Find Your Way Around.

Projects

  • Sample Applications for New Projects: If you have permission to create projects, you can now select Explore Application Catalog on the Projects page and pick a sample application, and AIO creates a project for it. The project's connection page includes a See in GitHub link, commands to clone the application or add it to an existing folder, and the remaining setup steps from its README. In the Projects list, these projects show the application's name without the code that AIO adds to it. For more information, see Start from a Sample App.
  • Projects Page Columns and Settings: The Projects page now includes a Columns picker that remembers your selection and a new Description column that is hidden by default, and the Project column is renamed Name. Select the gear icon on a row to open that project's Settings. When you open a section that needs a project, the list asks you to pick one and then returns you to that section. For more information, see Find Your Way Around.
  • Default Retention of 90 Days: The Retention (days) field on the Create project form now defaults to 90 days instead of 60. When no project is selected, or a project has no retention set, the Time selector offers ranges of up to 90 days. For more information, see Data Retention.

Python SDK

  • Database and HTTP Instrumentation: Version 1.1.3 of the acceldata-aio-tracer SDK extends aio.init() to instrument database drivers, HTTP server frameworks, and HTTP clients, so database queries and HTTP calls appear in your traces. Support for psycopg 3 is built in, and drivers without their own instrumentation are covered by a generic wrapper. To instrument more drivers, install the db extra. For more information, see Supported Providers and Frameworks.
  • Model Input Redaction for Part of an Application: Use the @aio.redact_model_inputs decorator or a with aio.redact_model_inputs(): block to replace the input messages and system instructions of each model call inside it with a SHA-256 fingerprint. This works with or without a redaction policy and doesn't change model outputs or other spans, such as tools and graph nodes. In generator functions, use the with block instead of the decorator. For more information, see Hide System Prompts and Tool Definitions.
  • SQL Literal Redaction: Set redact_sql_literals=True on aio.redaction_policy to replace string and number literals in SQL statements on database spans with ?. Statements, table and column names, and bind placeholders are kept, and if redaction fails, the SQL text is withheld. For more information, see Redaction Overview.
  • LangGraph Node Inputs and Outputs: Decorate a LangGraph node function with @aio.langgraph_node to record the state that the node receives and the value it returns. To choose what is recorded, pass the input= and output= helpers. Nothing is recorded when content capture is off. For more information, see Python SDK.
  • Export Compression and Queue Size: The SDK now compresses its exports with gzip, and its span queue holds up to 10,000 spans. To change these defaults, set OTEL_EXPORTER_OTLP_COMPRESSION (or a per-signal compression variable) or OTEL_BSP_MAX_QUEUE_SIZE. For more information, see Python SDK.
  • Warnings for Existing OpenTelemetry Setups: If aio.init() takes over a TracerProvider that it didn't create, the SDK now logs a warning the first time you use aio.redact_model_inputs(), because span content can't be withheld in this setup. Log records are still covered. A logger provider that isn't an SDK LoggerProvider now gets its own error message.To avoid both, call aio.init() before you configure OpenTelemetry. For more information, see OpenTelemetry.

Rules

  • Last Check Filter: Use the Last check filter on the Rules page to show rules by their latest result, such as Failed, Could not decide, or Passed. The filter is saved in the page URL, so you can share the filtered view. For more information, see Rules Overview.
  • Longer Judge Responses: Judge and classifier rules can now return responses of up to 4,096 tokens, up from 512. For more information, see Rules Overview.

Deprecations and Breaking Changes : This section lists changes to existing behavior that might require action.

  • SQL Values Removed When Content Capture Is Off: When content capture is off, the SDK now replaces literal values in SQL statements with ? before export and keeps table and column names, so AIO can still detect assets. To keep full SQL values in your traces, keep content capture on. For more information, see Data Quality Overview.
  • One Rule per Template per Project: Each project can now use a rule template only once. Templates that a project already uses show as In use and can't be selected, and Save & enable is disabled if the selected project already has a rule from that template. To change that rule, edit the existing rule. For more information, see Create a Rule.
  • Deleting Projects: You can't delete Acceldata ADM projects. Their trash icon is disabled, and its tooltip explains why. If a project deletion fails, the confirmation dialog now closes and an error notice appears. To retry, select the trash icon again. When the server doesn't respond, project and rule error messages now say so. For more information, see Create a Project.
  • Sessions Search: Search on the Sessions page now matches session IDs and no longer searches the first message of a session. For more information, see Explore Traces.
  • Rules Filter Wording: With the Rules filter on, the Traces and Spans pages now say "rules" instead of "checks", for example Clear rules filter. While the Rules filter is on, the Tags filter and the check box for hiding infrastructure traces are hidden. For more information, see Explore Traces.
  • Parameter Examples Removed from Rule Settings: Parameter fields on the rule configuration form no longer list example values that you can select to insert. The parameter description still appears below each field. For more information, see Create a Rule.

Resolved Issues

This section lists the issues that have been resolved in this release.

  • Resolved an issue where session rules evaluated each turn of a conversation as a separate conversation. Session rules now evaluate the whole conversation, so Conversation ran long triggers as expected and a later successful turn no longer changes an earlier failed turn to Pass.
  • Resolved an issue where a Latency regression result replaced the previous window's result for the same cohort.
  • Resolved an issue where rows at the end of a batch were not evaluated by rules. These rows are now evaluated in the next pass.