LangWatch
Sign Up
← Changelog

Go SDK 1.0 and Search in Datasets

The Go SDK reaches 1.0, with a rewritten guide for every provider it instruments natively. The dataset editor can now search its rows, and governance lists the agents your connected providers already know about.

LangWatch Team · September 13, 2026

What You Can Do Now

Go SDK 1.0

The Go SDK is 1.0, and the docs finally describe the SDK you actually install. The guide and reference are rewritten around NewExporter, and Anthropic, Gemini and Ollama are documented with their native instrumentations instead of the OpenAI-compatible shim, which means you get provider-accurate spans without routing through a compatibility layer.

go get github.com/langwatch/langwatch/sdks/go
go get github.com/langwatch/langwatch/sdks/go/instrumentation/openai
  • Each provider instrumentation is its own Go module, so importing one never pulls in another provider's SDK.
  • New pages for Amazon Bedrock and Firebase Genkit, and Go tabs on the quick start, metadata and labels, code examples and CLI pages.
  • Every Go snippet inside the app is rewritten for the new API, and Run via API has a Go tab for all three data sources.
  • The typed REST client covers prompts, datasets, traces, annotations and evaluations.

Search in datasets

  • Type in the search box in the dataset editor and the whole dataset narrows to the rows whose values contain the text, with the pager paging the matches. Matching is case-insensitive and ignores column names, so searching id does not return every row of a dataset with a conversation_id column.
  • The search runs on the server over the stored dataset, up to 50,000 rows or 100 MB. Past either limit it refuses with its own message instead of returning a partial answer that looks complete.

Agent discovery in governance

  • Connect a Databricks workspace or a Copilot Studio tenant and the agents that workspace or tenant already runs appear on the agents page, instead of only the ones somebody typed in by hand.
  • Inventory's catalog now reads the AI tool registry, so a card describes a tool you registered rather than the connector that fed it. Each card shows the rows that match how the tool is paid for, seats, subscriptions or tokens, and a summary strip above the tabs counts what is on screen.
  • Adding an Anthropic or OpenAI admin pull asks for less. The history date defaults to six months back for Anthropic and a year back for OpenAI, the report choice is a switch, and the advanced notes moved into a tooltip.

Test a connected agent with parameter overrides

  • Type name=value overrides under the Test agent message in the connected agent drawer, the same way you do in the run dialog. Names you leave out use the defaults the agent's code declares, and an unknown name or an option outside the declared list is refused in plain words.

The connected agent drawer, sending a test message with model=gpt-5 as an override

Personal ingest keys follow your CLI session

  • Personal ingest keys minted by a CLI session now live and die with that session. Logging out, revoking the device, signing in again from the same machine or letting the session expire retires that machine's keys, and nothing else revokes them, so setting up a new machine no longer breaks telemetry on another one.

Notes

  • Gateway budgets: a key at its budget limit kept answering budget_exceeded into the next period, because the cached spend described the period that had just ended. The gateway now re-reads spend when a period rolls over.
  • Traces: a feedback score of zero, a neutral thumbs rating included, was accepted on ingest and then dropped. Zeros are stored now, like any other number.
  • Data privacy: redaction was replacing about one in 57 trace ids because they looked enough like a bitcoin address to fool the detector. Opaque machine identifiers are left alone now.
  • Personal virtual keys: the list on your own settings page showed every personal key in the organization to an admin who can see other people's keys. It shows only yours now.
  • Docs: the introduction leads with agent testing, the Better Agents page is retired, and the agent testing section has a guide to grading runs with evaluators.
  • Go SDK 1.0 has one breaking change: Id, CreatedAt, UpdatedAt, ProjectId and TraceId on an annotation are values rather than pointers, so code that dereferences them no longer compiles. Read the field directly.
  • Shipped with Go SDK 1.0.0, typescript-sdk 1.14.0 and 1.15.0, MCP server 2.1.0, langevals 3.1.0 and agent-plugin 1.2.0.