v2.16.0

Compatibility Notes

• [openai] Raise the minimum required openai package version to 7.5.0. (875e285)
• [rpt] Deprecate sap-rpt-1-small and sap-rpt-1-large.
  These model names remain functional until their retirement date (2026-12-31). (ba5b1b4)
• [rpt] The server default changed from for parse_data_types changed from true to false.
  Pass parse_data_types: true explicitly if you relied on the old default. (ba5b1b4)
• [rpt] The RptClient constructor no longer has a default model name.
  Pass one of the known model names explicitly, e.g. 'sap-rpt-1.5'. (ba5b1b4)
• [rpt] Widen TargetColumnConfig.prediction_placeholder to accept null.
  The type is now string | number | null (was string | number). (ba5b1b4)

New Features

• [core, llm-batch] Add LlmBatchModel type for models supported in LLM batch processing.
  BatchCreateRequest.spec.model now uses this type for IDE autocomplete. (0da56e2)
• [core] Add gemini-3.5-flash-lite to the available model list.
  Remove deprecated model mistralai--mistral-large-instruct (retirement date: 2026-09-30). (4b2c014)
• [rpt] Add a confidence_interval field to PredictResponsePayload predictions for regression tasks. (ba5b1b4)
• [rpt] Default gzip compression level to 1 for RPT predict requests. (ba5b1b4)
• [rpt] Add an explanations field to prediction_config; read feature importance scores per query row from the response. (ba5b1b4)
• [rpt] Expand ColumnType from 3 to 16 values (integer, int16, int32, int64, uint8, decimal, double, boolean, largestring, uuid, time, datetime, timestamp).
  Map the new numeric variants to number; all other new variants (including datetime and timestamp, which are full ISO strings) map to string. (ba5b1b4)
• [rpt] Add context_mode to PredictionConfig and PredictResponseMetadata (from RPT spec v1.6.0). (21556db)

Fixed Issues

• [core] Set the default http agent socket timeout for AI Core requests to 1201 seconds (1201000 ms).
  This overrides the Cloud SDK default of 5 seconds, which is too short for chat and streaming completions.
  The value covers the orchestration server-side maximum of 1200 seconds plus 1 second of leeway.
  Keep-alive is now disabled on these requests.
  This avoids reusing stale sockets that a load balancer may have closed during the longer timeout.
  Both settings apply only to the service-binding destination.
  You can still override them with a custom destination's agentOptions or CustomRequestConfig.httpsAgent. (3483acb)
• [langchain] Normalize invalid tool message content to empty string to prevent AI Core 400 errors when MCP tools return empty results. (4cda527)

Improvements

• [core] Added mistralai--mistral-medium, sap-rpt-1.5, sap-rpt-1.5-large, sap-rpt-1.6, sap-rpt-1.6-large, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna to the available model list.
  Remove deprecated models gpt-4.1-nano (retirement date: 2026-10-14), gemini-2.5-flash (retirement date: 2026-10-16) and gemini-2.5-pro (retirement date: 2026-10-16). (94b36e5)
• [document-grounding] Update document grounding specification. (478fc53)