Braize

Product

Changelog

What shipped in Braize, newest first. We keep this honest — features land here when they are real and tested, not when they are announced.

  1. June 16, 2026

    v0.1.0

    Release

    First production release

    • Braize is live on a region-pinned single cell, served end-to-end: connectors → per-workspace knowledge base → cited answers over REST, OpenAI-compatible API, and the dashboard playground.
    • Manual, approval-gated deploys via GitHub Actions with semantic versioning starting at v0.1.0.
    • Daily database backups with 30-day retention to a separate bucket, plus a documented disaster-recovery runbook.

  2. June 2026

    Improved

    Answer accuracy & numeric fidelity

    • Multi-value completeness repair: detects when a multi-figure question dropped a value and runs one targeted strong-model repair when it strictly increases coverage.
    • Verbatim numeric contract with $/comma/trailing-zero, magnitude-word, and table-denomination folding so figures are copied, never recomputed.
    • Typed model routing escalates hard financial / multi-value / ambiguous-entity questions to the stronger model.
    • Measured query accuracy on real corpora: IRS 93.9, Apple 88.9; agent-native retrieve recall >90% at ~1.5s.

  3. June 2026

    Improved

    Agent-native retrieval & OpenAI-compatibility

    • POST /v1/retrieve returns permission-aware ranked passages with no generation latency — the same ACL pre-filter as /v1/query.
    • Drop-in OpenAI-compatible /v1/chat/completions over the same grounded pipeline.

  4. May–June 2026

    Infra

    Security & resilience hardening

    • ACL enforced as a pre-filter inside every retrieval SQL leg, proven by a dedicated leak suite.
    • Deferred-embedding resilience: an embedding-provider outage degrades retrieval to lexical-only instead of failing an upload or a query.
    • SSRF egress guard, prompt-injection sanitation and quarantine, and processing_policy inference routing with hard-fail.