Skip to content

v2.4.0 — Unified Runtime Wiring & Security Hardening

Choose a tag to compare

View all tags
@anulum anulum released this 02 Mar 22:37
· 2984 commits to main since this release
v2.4.0
8b141e5

What's Changed

P0 — Runtime Wiring

  • DirectorConfig.build_store() constructs VectorGroundTruthStore from config fields (vector_backend, chroma_*, reranker_*)
  • DirectorConfig.build_scorer(store=) auto-builds store and passes ground_truth_store=store
  • CoherenceAgent accepts _scorer/_store injection — backward-compatible
  • Server, gRPC, and CLI all share a single store → scorer → agent chain via config
  • CLI commands (review, process, batch) now respect DIRECTOR_* env vars

P1 — Security Hardening

  • Rate limiter default_limits wired (was a no-op)
  • Pydantic v2 body size limits: 100K chars prompt, 500K chars response (422 on violation)

P2 — Correctness

  • Batch result ordering preserved via index-based pre-allocation
  • ONNX batcher double-flush fixed; submit() returns [] below max_batch, explicit flush() added

P3 — Docker

  • ARG EXTRAS="server" enables lean vs NLI-enabled image variants

Tests

  • 1103 passed, 11 skipped, 0 failures
  • New tests for build_store, build_scorer, agent injection, env var override, rate limits, body size limits, batcher semantics
Assets 2
Loading