Changelog

All notable changes to this project will be documented in this file.

[0.4.1] - 2026-06-02

UI Model Configuration: Exposed select dropdowns for preferred Embedding Model and Vision/LLM Model (for OCR) in the Settings tab of the control panel UI (port 4733).
Backend Model Discovery: Implemented a new /api/models GET endpoint mapping active downloaded models in LM Studio.

Embedding Model Verification: Enforced matching check between query embedding model and index embedding model, automatically creating/checking embedding_model.json database metadata to prevent dimension mismatches and search query corruption.
Vision OCR Model Selection: Updated OCR processing to respect the user-selected vision model from settings, with automatic download loading and LLM fallback logic.

LanceDB Apache Arrow Compatibility: Converted raw Apache Arrow Vectors into native JavaScript arrays before performing MMR and K-Means cosine similarity calculations. Fixes fatal NaN distance ranking crash causing "Cannot read properties of undefined (reading 'vector')" errors.
Control Panel CSS: Fixed a CSS styling issue where dropdown text within the Search Settings menu was rendering invisibly (white text on white background).
Documentation Overhaul: Updated and to properly separate end-user benefits from developer architecture, and added a critical section regarding "Best Practices & Local Model Quirks" (pronoun binding, helpfulness bias/hallucinations, and tiny model rigidness).

Build System: Fixed an esbuild ES2024 unhandled target warning by properly bumping the tsconfig.json target to ES2022.
Git Ignore: Flattened ui/.gitignore into the root .gitignore to resolve nested ignore file errors during lms push bundling.

Build Isolation: Reverted the ncc standalone bundler logic. @lancedb requires platform-specific native .node binaries loaded dynamically. Shipping node_modules and utilizing standard tsc ensures the plugin doesn't crash on Windows/Mac/Linux with MODULE_NOT_FOUND errors for native bindings.

Extracted Diagnostics: Refactored the runAutomatedTest functionality out of src/index.ts and into src/diagnostics/ for cleaner module segregation.
Documentation: Added a "Quickstart: Recommended Models" section to the README.md advising users to load DeepSeek-OCR, EmbeddingGemma, and Granite-4-h-tiny concurrently for optimal 10-12GB VRAM usage.

Build Isolation: Fixed a bug where tsc was attempting to compile root-level tests/ during the production build step by adding it to tsconfig.json's exclude block.
Cleanups: Completely removed tracking of legacy files (failed.json, test scripts, .vscode/) and corrected UTF-16 encoding corruption.