Skip to content

CHANGELOG.md

Latest commit

 

History

History
67 lines (60 loc) · 3.82 KB

CHANGELOG.md

File metadata and controls

67 lines (60 loc) · 3.82 KB
Raw

Changelog

[Unreleased] - 2026-04-03

Added

  • LlamaLoraAdapterConfig resource and LlamaModelConfig.lora_adapters
  • LlamaMultimodalConfig resource and LlamaModelConfig.multimodal_config
  • LoRA adapter loading through the stable llama.cpp adapter API during session open
  • optional libmtmd build wiring behind GODOT_LLAMA_ENABLE_MTMD
  • LlamaSession capability introspection for loaded LoRA adapter count plus image/audio multimodal support
  • LlamaSession.generate_multimodal_async() for multimodal generation with image/audio inputs
  • LlamaSession.generate_multimodal_messages_async() for message-templated multimodal generation
  • LlamaSession.image_to_media_input() convenience helper for Image -> PNG media dictionaries
  • LlamaSession.get_multimodal_token_count() for completed multimodal requests
  • in-memory PackedByteArray media input support via data key in media input dictionaries
  • MultimodalInput.data field for passing image/audio bytes without a file path

Changed

  • Multimodal generation methods are now fully bound and documented
  • LlamaMultimodalHandle::evaluate_prompt accepts both file-path and in-memory buffer media
  • Media input dictionaries now accept either path or data (or both, with data taking precedence)
  • Multimodal submission now validates media dictionary shape, readable file paths, and media-marker counts before enqueuing worker requests
  • completed stats now include multimodal_token_count for multimodal requests
  • Updated API.md with full multimodal method documentation and media input dictionary format
  • RAG storage now builds against the vendored libSQL amalgamation instead of the old bundled SQLite path
  • RAG retrieval now defaults to exact cosine SQL search over libSQL vector columns
  • RagCorpusConfig no longer exposes vector_metric; the current public RAG surface is cosine-only
  • Legacy RAG schema v1 embedding_blob databases now migrate to schema v2 embedding_vec F32_BLOB storage on open

[0.2.0] - 2026-04-02

Added

  • first-class local RAG support
  • RagCorpusConfig, RagCorpus, and RagAnswerSession Godot classes
  • SQLite-backed persistent corpus storage with schema versioning
  • deterministic UTF-8 normalization and token-aware chunking
  • dedicated embedding pipeline through libllama
  • exact dense retrieval with metadata/source filtering, dedupe, and MMR
  • grounded prompt assembly with packed-chunk citations
  • mock-embedder-backed RAG unit tests and SQLite integration tests
  • offline evaluation executable and RAG documentation set
  • updated demo project with local corpus ingestion, retrieval preview, and streamed grounded answers

[0.1.0] - 2026-04-01

Added

  • Initial project structure with CMake/Ninja build system
  • CMakePresets.json with dev, release, asan, tsan presets
  • llama.cpp integration via pinned submodule (commit 825eb91a6)
  • godot-cpp integration via pinned submodule (4.6-stable, tag 10.0.0-rc1)
  • src/llama/: Thin C++ wrapper layer around libllama
    • LlamaModelHandle: shared model ownership with tokenize/detokenize
    • LlamaContextHandle: move-only context with decode and embeddings
    • LlamaSamplerHandle: move-only sampler chain with configurable parameters
  • src/core/: Async runtime
    • InferenceWorker: background jthread for non-blocking generation
    • Request queue with cooperative cancellation
    • Structured Error type with code/message/context
    • UTF-8 validation and codepoint counting
  • src/godot/: GDExtension binding layer
    • LlamaModelConfig Resource with editor-friendly properties
    • LlamaSession RefCounted with async generation, tokenization, embeddings
    • Signal-based result delivery via poll()
  • Unit tests (Catch2) for UTF-8, Error, and parameter types
  • Minimal Godot demo project
  • .gdextension manifest for Linux, Windows, macOS
  • Documentation: BUILD.md, ARCHITECTURE.md, API.md