v0.4.8 - Multilingual support, security update, context bugs

-- Configurable evidence sources — new Settings → Advanced tab lets you pin the number of retrieved evidence snippets per query (4–30). Local/Ollama users who were seeing only 4 sources can now increase this independently of the model's context window. Default remains Auto (scaled dynamically).

-- Multilingual library support — three new embedding models available in Library settings: BAAI/bge-m3 (100+ languages, ~2.3 GB), BAAI/bge-large-zh-v1.5 (Chinese-optimised, ~1.3 GB), and paraphrase-multilingual-MiniLM-L12-v2 (50+ languages, ~470 MB). Download on first use, no extra dependencies. Fixes BM25 tokenization for Chinese, Japanese, and Korean text, which was producing near-zero retrieval scores.

-- Follow-up query regression fixed — on turn 2+, fresh retrieval evidence was silently discarded, causing GPT-4o and similar models to generate a new question rather than an answer. Evidence is now correctly injected into every follow-up turn.

-- Conversation history no longer drops first-turn evidence — the context trimmer now pins the first exchange (which carries the bulk of library evidence) and only evicts middle turns when the budget is exceeded.

-- Query condenser reliability — fixed four bugs: false positives on short comparison questions, missed anaphora detection when queries end with ?, condensation firing before any assistant reply exists, and a hallucination guard that falls back to the original query when the condensed version drifts too far.

-- Security — 12 dependency updates including starlette 0.49.3 → 1.0.1 (malformed Host header CVE), fastapi 0.121.2 → 0.134.0, axios 1.13.2 → 1.17.0, react-router-dom 7.9.6 → 7.17.0, and several Python package patches.

v0.4.7

v0.4.7

A mix of new features and bug fixes. The biggest additions are cloud embedding support and a bunch of fixes for empty model list issues that several people ran into.

What's new

Cloud embeddings (OpenAI)
You can now use OpenAI's text-embedding-3-small or text-embedding-3-large as your embedding backend instead of a local model. This is useful if you don't have the RAM for a local model or want higher-quality embeddings. Head to Library → Embedding Model to try it — you'll need your OpenAI API key configured first. There's a privacy warning before you switch so you know what you're signing up for (your PDF text leaves your device).

Source card buttons
The icon-only buttons on each source card (Scholar, Books, Semantic Scholar) have been replaced with labelled text buttons with favicons. Should be much easier to tell apart at a glance.

Bug fixes

• LM Studio / single-provider users (#28): If you only had one provider configured, the model dropdown would often stay empty. Fixed — models are now fetched fresh whenever settings change, and there's a ↻ refresh button if you ever need to force a reload.
• OpenRouter / cloud provider users (#26): Enabling a cloud provider while Ollama was set as the active provider could leave you stuck with an empty model list and a connection error pointing at Ollama. The app now auto-switches to the first enabled provider at startup and on settings save.
• Google provider: An invalid API key was being silently swallowed and treated as a transient error. It now correctly surfaces as an auth failure.
• Metadata filters without a provider: Year range queries like "papers after 2020" now work even if no LLM provider is configured — extracted via regex as a fallback.
• Chat window top border: There was a small visible strip at the top of the chat panel at all times. Gone.

v0.4.6

GitHub Models provider — Connect via a dedicated GitHub Models option using your GitHub Token (no manual endpoint config needed); contributed by @ComePerin
Fix: ZotMoov / linked-file PDFs — PDF paths for files stored outside the default Zotero directory now resolve correctly (storage:, absolute, and attachments: path types all handled)
Fix: Sources panel crash — Missing apiFetch import caused the Sources panel to fail to load
Fix: OpenAI-compatible model listing — Added fallback for non-standard model list responses; normalises Azure ML resource-path IDs to simple model names

RAG Assistant for Zotero v0.4.5

New

• Custom PDF storage directory — set a non-default path to your Zotero PDF storage folder in Settings → Zotero Configuration
• Semantic Scholar search button in the Sources panel (alongside existing Google Scholar)

Fixed

• Syncing stuck at 0%: when Zotero's database is locked, the error message now explicitly tells you to close Zotero before syncing
• Sync failures no longer disappear silently — errors are now visible in the UI
• OpenAI model list: the available models dropdown now accurately reflects what your API key has access to, rather than showing all curated GPT models unconditionally
• Changing API key or base URL now clears the cached model list so it reloads correctly
• Windows: startup no longer fails prematurely while PyInstaller is still extracting on launch
• Windows: port cleanup on startup no longer risks killing unrelated applications

Removed

• Ollama status indicator from the top navigation bar
• Windows 32-bit (ia32) build targets — x64 only going forward
• Intel Mac (x64) builds were already dropped in v0.4.5; arm64 / Apple Silicon only (Rosetta 2 works for Intel users)

v0.4.4 UX Improvements & Fixes

As luck would have it, Github had an outage at the same time I was trying to release a new version, but this seems to be working now! Sorry for the confusion.

This is a quick patch to address some UX issues.

• Save & Test Button: API key testing now automatically saves settings first, eliminating the confusing "test failed because you didn't save" issue

• Custom Endpoints: Added optional custom endpoint configuration for OpenAI-compatible providers - enables Azure OpenAI, proxies, and custom deployments

• Bug Fix: Resolved TypeScript errors preventing sound notification settings from saving properly

• License Correction: Fixed macOS DMG to include correct Apache 2.0 license (was incorrectly showing MIT)

v0.4.3 - Trademark Compliance

RAG Assistant for Zotero v0.4.3

Name Change & Migration Required

This release includes a name change to comply with Zotero's trademark requirements. The application is now called "RAG Assistant for Zotero" instead of "ZoteroRAG".

Why This Change?

Zotero's trademark guidelines require third-party apps to use the format "for Zotero" rather than starting with "Zotero". We've updated:

• Application display name → "RAG Assistant for Zotero"
• Executable name → rag-assistant
• Package name → rag-assistant
• Repository → RAG-Assistant-for-Zotero

Migration Steps (One-Time Only)

Your profiles, vector databases, and settings will be preserved if you run these commands before launching the new version:

macOS

# Close the app if running, then rename the data folder:
mv ~/Library/Application\ Support/ZoteroRAG ~/Library/Application\ Support/RAG\ Assistant

Windows

# Close the app if running, then rename the data folder:
# Open PowerShell and run:
Rename-Item -Path "$env:APPDATA\ZoteroRAG" -NewName "RAG Assistant"

Or using Command Prompt (cmd.exe):

rename "%APPDATA%\ZoteroRAG" "RAG Assistant"

Linux (Debian/Ubuntu/AppImage)

# Close the app if running, then rename the config folder:
mv ~/.config/zotero-rag-assistant ~/.config/rag-assistant

What Happens If I Don't Migrate?

If you skip the migration and launch v0.4.3 directly:

• The app will create a new data folder with the new name
• You'll see empty profiles and need to re-index your library
• Your old data will remain in the old folder (safe to manually migrate later)

If you encounter any issues with the migration, please open an issue at: https://github.com/aahepburn/RAG-Assistant-for-Zotero/issues

v0.4.2

Added provider-aware dynamic caps to cater to both local and online models of various sizes.

v0.4.1

Full Changelog: v0.4.0...v0.4.1

v0.4.0

Zotero RAG Assistant v0.4.0 - Intelligent Filtering & Enhanced UI

Major Features

Metadata Filtering System
Filter your research by Zotero's rich metadata - publication date ranges, tags, and collections. The new Reciprocal Rank Fusion (RRF) algorithm intelligently balances filtered results with pure semantic search, ensuring you never miss relevant papers even with strict filters. There are multiple modes for this -- filtering can be done manually or via an LLM call.

Multi-Tab Sidebar
Reorganised interface into dedicated Chat, Library, and Prompts tabs for cleaner navigation and better workflow.

LM Studio Support
Official integration with LM Studio for running local models. Automatic model discovery via /v1/models endpoint with optimized prompts for local inference.

External Models Support
Users can now enter their API keys and use online models from Claude, OpenRouter, OpenAI, etc.

Dynamic Model Discovery
Ollama and LM Studio now automatically fetch available models from their APIs, eliminating manual configuration.

Improvements

• Settings Redesign: Tabbed interface (General, Advanced, About) for better organization
• ChromaDB Schema v2: Enhanced metadata storage enables efficient filtering queries
• Metadata Migration Overhaul: Pre-cached lookups, bidirectional sync, graceful error handling
• Search Mode Clarity: Renamed "Original" to "Semantic only" with clearer descriptions
• New Icon: Redesigned with Zotero's official color scheme (white background, red/purple accents)
• Linux Packaging: Simplified to .deb only for more reliable installation

Fixes

Linux Stability (Ubuntu 24.04)

• Fixed immediate renderer crashes (SIGKILL) on startup
• Added comprehensive crash handlers with helpful error messages
• Implemented safe GPU/sandbox defaults for Linux
• Command-line flags (--no-sandbox, --disable-gpu) now work correctly
• Post-install script sets proper chrome-sandbox permissions

v0.3.2

Changed

• Provider Integration: Replaced Perplexity provider with Mistral AI
• Removed PerplexityProvider class and all Perplexity-specific code
• Added MistralProvider with OpenAI-compatible API
• Updated frontend, tests, and documentation
• Available models: mistral-large-latest, mistral-medium-latest, mistral-small-latest, open-mistral-nemo, open-mixtral-8x7b
• CI Build Process: Updated macOS Intel builds to handle PyTorch compatibility
• Intel Mac CI runners use torch==2.2.2 for building (last version with x64 wheels)
• Bundled apps for Intel Macs continue to work correctly
• Security risk isolated to CI build environment only

Security

• PyTorch Version: Updated to torch>=2.6.0 to address CVE-2025-32434
• Fixes serious vulnerability in torch.load() function
• Breaking change for Intel Mac developers: PyTorch 2.6+ does not provide x86_64 wheels
• Intel Mac developers must manually install torch==2.2.2 (with known security risk)
• Intel Mac users (GUI): Unaffected - DMG builds continue to work normally
• App uses safetensors format for model loading which mitigates the vulnerability
• See README for Intel Mac development setup instructions