Skip to content

CHANGELOG.md

Latest commit

 

History

History
53 lines (42 loc) · 2.53 KB

CHANGELOG.md

File metadata and controls

53 lines (42 loc) · 2.53 KB
Raw

Changelog

All notable changes to FerryAI. Format loosely follows Keep a Changelog.

[Unreleased]

[0.1.1] - 2026-07-09

Security

  • Harden .ai archive deserialization and native-binary integrity checks (ai, cpu-backend).

Changed

  • Reorganize repository root: move community-health files to .github/ and documentation into docs/.
  • Drop the PHP baseline to 8.3 (replace \Pdo\Sqlite with \PDO + getAttribute() + method_exists() runtime checks).
  • Remove Docker files (native FFI runtime needs no container sidecars).

Fixed

  • CI: auto-detect the VS2022 toolchain path, correct the release asset name and line-ending normalization, and exclude SharedMemoryManager from Psalm (ext-shmop Shmop class not recognized).
  • Documentation: correct the supported-versions table for the 0.1.x pre-1.0 line.

[0.1.0] - 2026-07-08

First public release. Inference-only runtime for PHP 8.3+ with a unified API over ONNX Runtime, llama.cpp and RubixML/Tensor engines.

Added

  • PostgreSQL + pgvector vector store (PostgresStore, PostgresCollection, PostgresVecIndex) with native ANN (<=>/<->/<#>) and HNSW/IVFFlat indexes; driver switch in AIFactory.
  • SQLite native KNN via the sqlite-vec (vec0) extension, opt-in, with brute-force fallback.
  • Observability wrapper wiring Metrics/Profiler/Logger into the facade (opt-in); ModelPool integration (pooled model loading, LRU eviction, real warmup).
  • NativeBinaryManager library resolution in AIFactory; RetryHandler+Logger in the model hub.
  • CPU backend: real pure-PHP tensor arithmetic and RubixML .rbm inference (isolated).
  • FFI CDEF generator (Core\FFI\CdefGenerator, bin/generate-ffi.php).
  • Facade config-wiring for embed/similarity; PSR-7 StreamResponse::create().
  • llama.cpp CPU + GPU inference via the ferry_llama wrapper, wired into LlamaBackend (AI::chat()/AI::stream()), with a native top-k pre-filter for fast sampling.
  • Strict grammar-constrained sampling (GbnfMatcher + GrammarSampler); sampler/grammar options on AI::chat().
  • Per-capability documentation under docs/ and this changelog.

Changed

  • AIFactory caches embedders per model; AI::chat pools its model (no reload per call).
  • SamplerFactory::forParams() selects greedy/nucleus by temperature; softmax honours temperature.

Fixed

  • RubixMLAdapter::isAvailable() used class_exists for the Estimator interface (now interface_exists).
  • All PHPStan level 8 findings resolved (gate fully green).