Google ships AI Studio observability
Google upgraded Vertex AI’s Studio with granular Logs and Datasets features so developers can reconstruct sessions and analyze per‑action model behavior — a clear push into observability as first‑class developer tooling. The additions close gaps for debugging and governance by enabling session‑level trace reconstruction and per‑action analytics for model fine‑tuning. (engagecoders.com)
Logging can be turned on from the AI Studio Build dashboard with a single “Enable logging” click that starts capturing GenerateContent and StreamGenerateContent calls for the selected billing-enabled Cloud project without code changes. (blog.google)) AI Studio imposes a default per‑project storage limit of up to 1,000 logs and will expire logs not saved into datasets after 55 days, forcing explicit curation for long‑term analysis. (ai.google.dev)) Selected log rows can be turned into named Datasets inside AI Studio and exported as CSV, JSONL, or directly to Google Sheets for offline review or for feeding back into evaluation workflows. (ai.google.dev)) Datasets created from logs can be re‑used to re‑run interactions via the Gemini Batch API and optionally shared with Google as demonstration examples to support model improvements. (blog.google)) Vertex AI’s request‑response logging writes samples to a BigQuery table and supports sampling configuration (for example, a sampling_rate of 1 logs all requests while 0.1 logs 10% of requests), with support extending to Gemini models and certain partner models like Anthropic when using compatible endpoints. (docs.cloud.google.com)) Google notes the feature is live in AI Studio Build mode and available in regions where the Gemini API is offered, and Vertex docs flag request‑response logging as a Pre‑GA offering with possible limitations on support. (blog.google))
