Changelog

STAN-1-Mini policy model upgraded — average compression improved from 45% to 50%.

DataAPI ingest endpoint now supports DOCX, PPTX, and XLSX out of the box.

Default retry policy is now exponential with jitter (was linear).

Rare race condition in streaming agent responses on Node.js 18.

ContextAPI parse endpoint occasionally returned empty IDENTITY blocks for prompts under 50 tokens.

Python SDK timeout default was incorrectly 30s — restored to documented 60s.

VexaAgent private beta opened. Request access via your account team.

New mode parameter on optimize — 'auto' | 'aggressive' | 'light'.

DataAPI CAR-1.5 distillation latency reduced by ~30% across all tenants.

Legacy /v0 endpoints scheduled for removal on 2026-09-01.

ingest endpoint now returns a job ID and runs asynchronously. Sync mode available with sync=true.

Idempotency-Key header support on all write endpoints.

Bun runtime officially supported in TypeScript SDK.

Rate-limit headers were missing on 200 responses for DataAPI query.

Edge runtime compatibility for Cloudflare Workers.

Engram filter — 7-signal hybrid re-ranking on CAR-1 retrieval.

Tenant-level token budget caps.

Quickstart guide rewritten with end-to-end TypeScript and Python examples.

Multi-region failover for managed cloud deployments (US, EU, APAC).

Streaming responses for ContextAPI optimize endpoint.

General availability. ContextAPI and DataAPI stable. Backwards compatibility guaranteed within 1.x.