v2.0.0

🚀 Major Update

📦 Easier Setup & Modern Foundation

• The program now works with Python 3.11, 3.12, and 3.13.
• Upgraded to the latest versions of all the major AI libraries, including PyTorch with CUDA 12.8 support.
• The setup process now automatically detects whether you have an NVIDIA GPU and installs the right components.
• Internal settings are now validated automatically, so typos or invalid values in the configuration won't silently break things.

---

🧠 Rebuilt Database Engine — More Stable, Faster, and Crash-Resistant

• The process of turning your documents into a searchable database has been completely rewritten. The biggest win: if something crashes partway through a large job, the program can now pick up where it left off instead of starting over.
• New performance presets (Minimal, Low, Normal, High, Maximum) let you tell the program how hard to push your computer.
• Switched to a faster, more efficient way of storing and searching the document database. Queries return quicker, files take up less space on disk, and large databases (tens of thousands of documents) are much more manageable.
• Embedding (the math that turns text into something the AI can search) is now smarter about batching similar-sized chunks together, which speeds things up without you having to do anything.
• Better support for very large embedding models like Qwen3-Embedding 8B, with proper handling for FlashAttention when your GPU supports it.
• Document text is now cleaned up more carefully before being added to the database — invisible characters, weird spacing, and other quirks that used to cause problems are now handled gracefully.

---

🖼️ Refreshed Lineup of Vision Models

• The old image-description models have been replaced with a brand-new lineup that's significantly smarter and more capable:
  • Liquid-VL — three sizes (480M, 1.6B, 3B), fast and efficient
  • InternVL3 — four sizes (1B, 2B, 8B, 14B), excellent quality
  • Qwen VL 3B — strong all-rounder
• Larger models are now automatically loaded in a compressed format that uses much less VRAM, so you can run bigger, smarter models than before on the same hardware.
• New side-by-side comparison tool: pick a single image, select multiple vision models, and the program will run all of them so you can see which one describes your images best. You'll get a nicely formatted report showing the descriptions, how long each took, and how much they wrote.

---

🎨 Refreshed Lineup of Search Models

• The catalog of "embedding models" (the models that power document search) has been overhauled with the latest and greatest:
  • Qwen3-Embedding in three sizes (0.6B, 4B, 8B) — currently among the best available
  • Snowflake Arctic Embed v2 — strong quality at a reasonable size
  • Google EmbeddingGemma — Google's new compact model
  • IBM Granite Embedding — small, fast, commercially friendly
  • infly's inf-retriever — high-end retrieval models
  • Updated BGE and E5 families
• Each model now shows you helpful info up front: how big it is, how much memory it needs, what license it's under, and a quality ranking so you can make an informed choice.

---

✨ Polished Interface & New Themes

• 24 built-in color themes to pick from — everything from clean defaults to fun ones like Matrix, Tron, Monet, and Hyperbolic.
• Helpful tooltips have been added throughout the program.
• Selecting documents got much better: you can now pick an entire folder at once (with the option to include subfolders), and the program will efficiently link to all compatible files.
• New "Chunks Only" query mode: instead of asking the AI a question, you can preview exactly which pieces of your documents would be sent to it. Great for fine-tuning your settings before running real queries.
• The query view now shows clickable citations with similarity scores and PDF page numbers (grouped intelligently, like "p.3-5, 8"), so you can jump straight to the source of any answer.

---

🔧 Other Quality-of-Life Improvements

• PDF safety check: before building a database, the program now scans your PDFs and warns you if any of them are scanned images that need OCR first — no more silently empty entries in your database.
• Cleaner AI responses: when using reasoning models (like Deepseek-R1) that "think out loud," that internal thought process is now hidden from the chat view automatically, so you just see the actual answer.
• Updated KoboldCpp downloader to match its current naming and download options. If you don't have an NVIDIA GPU, the program now correctly steers you toward the Vulkan/CPU build.

---

⚠️ Important Notes Before Upgrading

• Existing databases need to be rebuilt. The new database format is faster and more efficient, but it's not compatible with databases created in older versions. Your source documents are untouched — you'll just need to run them through database creation again.
• Fresh install recommended. The cleanest way to upgrade is to delete your old virtual environment folder and re-run the setup script. The setup process is now smoother and faster than before, so this isn't a big lift.
• The bundled KoboldCpp file naming has changed. If you have a script or shortcut pointing to a specific Kobold binary, double-check the new names listed in the README.

v1.3.0 - MiniCPM v2 is awesome!

Release notes

• Added the MiniCPM v2 vision model, which is currently the highest quality publicly available one IMHO.
• Added bitsandbytes dependency in order to run MiniCPM v2.

v1.2.1 - bug fix

Improvements

• Added the correct download_model.py file that release v1.2.0 accidentally forgot to include.
• Removed a de-bugging print statement from gui.py.

The transcriber functionality and image summary generator should hopefully work properly now along with everything else. If not, please submit an issue.

v1.2.0 - add whisper models

Improvements

• Added all distil whisper models and removed the esoteric quants; nobody really used them.
• Fixed a bug in gui.py.

v1.1.0 - faster installation

Much faster installation using the amazing uv library written in Rust to speedup regular pip install commands.

v1.0.1 - update moondream2

Using a newer version of Transformers requires using a newer version of Moondream2's source code.

v1.0.0 - Kobold VectorDB Plugin

Welcome and enjoy!

To use just start Kobold in server mode and then experiment with the program. Feel free to hit me up as Vic49 on the Kobold Server or feel free to submit an issue here or contribute as well.

v0.0.1

v0.0.1