v6.9.0 - Welcome Kobold!!

Welcome Kobold edition

Ask Jeeves!

• Exciting new "Ask Jeeves" helper who answers questions about how to use the program. Simply click "Jeeves" in the upper left.
• "Jeeves" gets his knowledge from a vector database that comes shipped with this release! NO MORE USER GUIDE TAB - just ASK JEEVES!
  • IMPORTANT: After running setup_windows.py you must go into the Assets folder, right-click on koboldcpp_nocuda.exe, and check the "Unblock" checkbox first! If it's not there, try starting Jeeves and see if it works. Create a Github Issue if it doesn't work because Ask Jeeves is a new feature.
  • IMPORTANT: You may also need to disable or make an exception for any firewall you have. Submit a Github Issue if you encounter any problems.

Scrape Python Library Documentation

• In the Tools Tab, simply select a python library, click Scrape, and all the .html files will be downloaded to the Scraped_Documentation folder.
• Create a vector database out of all of the .html files for a given library, then use one of the coding specific models to answer questions!

Huggingface Access Token

• You can now enter an "access token" and access models that are "gated" on huggingface. Currently, llama 3.2 - 3b and mistral-small - 22b are the only gated models.
• Ask Jeeves how to get a huggingface access token.

Other Improvements

• The vector models are now downloaded using the snapshot_download functionality from huggingface_hub, which can exclude unnecessary files such as onnx, .bin (when an equivalent .safetensors version is available), and others. This significantly reduces the amount of data that this program downloads and therefore increases speed and usability.
• This speedup should pertain to vector, chat, and whisper models, and implementing the snapshot_download for TTS models is planned.
• New Compare GPUs button in the Tools Tab, which displays metrics for various GPUs so you can better determine your settings. Charts and graphs for chat/vision models will be added in the near future.
• New metrics bar with speedometer-looking widgets.
• Removed the User Guide Tab altogether to free up space. You can now simply Ask Jeeves instead.
• Lots and lots of refactoring to improve various things...

Added/Removed Chat Models

• Added Qwen 2.5 - 1.5b, Llama 3.2 - 3b, Internlm 2.5 - 1.8b, Dolphin-Llama 3.1 - 8b, Mistral-Small - 22b.
• Removed Longwriter Llama 3.1 - 8b, Longwriter GLM4 - 9b, Yi - 9b, Solar Pro Preview - 22.1b.

Added/Removed Vision Models

• Removed Llava 1.5, Bakllava, Falcon-vlm - 11b, and Phi-3-Vision models as either under-performing or eclipsed by pre-existing models that have additional benefits.

Roadmap

• Add Kobold as a backend in addition to LM Studio and Local Models, at which point I'll probably have to rename this github repo.
• Add OpenAI backend.
• Remove LM Studio Server settings and revise instructions since LM Studio has changed significantly since they were last done.

Full Changelog: v6.8.2...v6.9.0

v6.8.2 - quality focus

Due to the growing of number of chat and vector models with larger contexts, this mini-release focused on extensive testing at longer contexts. From this point forward, 4k chat models will only be included if they're exceptional and 8k++ models if they're quality and/or offer unique characteristics (e.g. focused on coding etc.)

Added Models (all 8k++ context):

• LongWriter Llama 3.1 - 8b
  • • Exceptional at long responses where an unusual number of contexts are thrown at it.
• Yi - 9b
  • • Long context Yi 9b, replacing Dolphin-Yi 9b, which was under performing at long context.
• Solar Pro Preview - 22.1b
  • • Exceptional 4k model with a parent model that's 8k coming out in a few months; replaces Solar 10.7b.

Removed Models:

• Danube 3 - 4b - under performing at long context
• Dolphin-Qwen 2 - 1.5b - under performing at long context
• Orca 2 - 7b - superseded
• Neural-Chat - 7b - superseded
• Dolphin-Llama 3.1 - 8b - superseded
• Hermes-3-Llama-3.1 - 8b - superseded
• Dolphin-Yi 1.5 - 9b - redundant
• Dolphin-Qwen 2 - 7b - superseded
• Dolphin-Phi 3 - Medium - too difficult to work with and superseded
• Llama 2 - 13b - superseded
• Dolphin-Mistral-Nemo - 12b - too difficult to work with and superseded
• SOLAR - 10.7b - superseded

See Release 6.8 for full release notes, including how to upgrade old databases

Current Chat Models:

Current Vision Models

Current TTS Models

• Does not include Google's, which is online and hence no GPU or VRAM usage.

v6.8.1 - coding models!

Fix

• Added a single missing dependency from the last release.

See Release 6.8.0 for full notes, including how to update databases.

v6.8.0 - coding models!

Breaking Changes

Within the Manage Databases tab, what's displayed is no longer derived from parsing multiple JSON files. Rather, sqlite3 is used for much much faster response/latency.

As such, databases created prior to this release will not function properly. To migrate old databases rather than creating them anew, use the create_sqlite3.py attached to this release as follows:

1. Run the script and select the folder containing the old JSON files. The script will folder location is as follows:

• Go into the Vector_DB folder.
• Each folder within Vector_DB constitute a database.
• Within each folder constituting a database there is a folder named json.
• You need to run the create_sqlite3.py script selecting each json folder for each of your database folders.
• A new file named metadata.db should be created in each database folder.
• It's now safe to remove the json folder altogether.

2. Update the backup folder.

• The Vector_DB_Backup folder contains a mirror image of the Vector_DB - this is the "backup" folder.
• Delete all the contents within the Vector_DB_Backup folder and copy all the contents of the Vector_DB folder into it.

3. Start the program and it should run as usual.

These steps are unnecessary if you're creating a vector database for first time using Release v6.8.0, obviously.

New Chat Models for Coding Questions

• Deepseek-Coder-V2 - 16b (best)
• Yi-Coder - 9b (second best)
• CodeQwen1.5 - 7b (third, but still good)

Benchmarks will be forthcoming, but their metrics are comparable to similarly-sized models.

Misc.

• Now using a distil-whisper model for the voice recorder for an approximate 2x speedup.
• Added buttons to backup all databases at once or restored all backups at once (Tools Tab).

v6.7.1 - patch update

Patch

• Fixed Florence models not showing up as an option for vision models when a gpu was detected.

See v6.7.0 for all other release notes:

https://github.com/BBC-Esq/VectorDB-Plugin-for-LM-Studio/releases/tag/V6.7.0

v6.7.0 - LONG CONTEXT no see!

General Updates

• CITATIONS! with hyperlinks when searching the Vector DB and getting a response.

• Display of a chat model's max context and how many tokens you've used.

2X Speed Increase

Choose "half" in the database creation settings. It will automatically choose bfloat16 or float16 based on your GPU.

This results in a 2x speed increase with extremely low loss in quality.

Chat Models

Removed Internlm2_5 - 1.8b and Qwen 1.5 - 1.6b as under performing.
Removed Dolphin-Llama 3 - 8b and Internlm2 - 20b as superseded.
Added Danube 3 - 4b with 8k context.
Added Phi 3.5 Mini - 4b with 8k context.
Added Hermes-4-Llama 3.1 - 8b with 8k context
Added Internlm2_5 - 20b with 8k context

The following models now have have 8192 context:

Model Name	Parameters (billion)	Context Length
Danube 3 - 4b	4	8192
Dolphin-Qwen 2 - 1.5b	1.5	8192
Phi 3.5 Mini - 4b	4	8192
Internlm2_5 - 7b	7	8192
Dolphin-Llama 3.1 - 8b	8	8192
Hermes-3-Llama-3.1 - 8b	8	8192
Dolphin-Qwen 2 - 7b	7	8192
Dolphin-Mistral-Nemo - 12b	12	8192
Internlm2_5 - 20b	20	8192

Text to Speech Models

• Excited to add additional models to choose from when using whisperspeech as the text to speech backend - see the chart below for the various s2a and t2s model combinations and "relative" compute times along with real vram usage stats.

Current Chat and Vision Models

v6.6.0 - 8192 CONTEXT!

General Updates

• Ensured that vector model pulldown menu auto-updates.
• Made the vector model pulldown menu more descriptive.

Local Models

• Added Internlm v 2.5 1.8b. In the last release, version 2.0 of Internlm's 1.8b model was removed. However, the quality increased noticeably with their version 2.5 so I'm re-adding it.

Vector Models

• Excited to add Alibaba-NLP/gte-base-en-v1.5 and Alibaba-NLP/gte-large-en-v1.5. These vector models have a context limit of 8192, which is automatically set within the program. With a conservative estimate of 3 characters per token, that means that you can set the chunk size to approximatly 24,576!!
• Removed Stella as it was under-performing and too difficult to work with. There is no love loss since the prior release marked it as "experimental" anyways.

Current Chat and Vision Models

v6.5.0 - Llama 3.1 & MiniCPM v2

General updates

• Remove triton dependency as cogvlm vision model is also removed.
• Redid all benchmarks with more-accurate parameters.

Local Models

Overall, the large amount of chat models was becoming unnecessary or redundant. Therefore, I removed models that weren't providing optimal responses to simplify the user's experience, and added Llama 3.1.

Removed Models

• Qwen 2 - 0.5b
• Qwen 1.5 - 0.5b
• Qwen 2 - 1.5b
• Qwen 2 - 7b
  • Redundant with Dolphin Qwen 2 - 7b
• Yi 1.5 - 6b
• Stablelm2 - 12b
• Llama 3 - 8b
  • Redundant with Dolphin Llama 3 - 8b

Added Models

• Dolphin Llama 3.1 - 8b

Vision Models

Overall, two vision models were removed as unnecessary and MiniCPM-V-2_6 - 8b was added. As of the date of this release, MiniCPM-V-2_6 - 8b is now the best model in terms of quality. I currently recommend using this model if you have the time and VRAM.

Removed Models

• cogvlm
• MiniCPM-Llama3

Vector Models

• Added Stella_en_1.5B_v5, which ranks very high on the leaderboard.
  • Note, this is a work in progress as currently the results seem to be sub-optimal.

Current Chat and Vision Models

v6.4 - stream responses

Improvements

• All "local models" now stream their responses for a better user experience.
• Various small improvements.

Local Models

• Fixed Dolphin Phi3-Medium
• Added Yi 1.5 - 6b
• Added H2O Danube3 - 4b
  • Great quality small model.
• Removed Mistral v.03 - 7b
  • The model is gated so it's difficult to implement in a program. Plus, there are a plethora of other good models.
• Removed Llama 3.1 - 8b
  • Same as with Mistral.
• Added Internlm 2.5 - 7b
• Fixed Dolphin-Mistral-Nemo

Vision Models

• Added Falcon-vlm - 11b
  • Great quality. Uses Llava 1.6's processor.

Falcon-vlm, Llava 1.6 Vicuna - 7b, and Llava 1.6 Vicuna - 13b have arguably surpassed Cogvlm and are faster for less VRAM. Thus, Cogvlm may be deprecated in the future.

Misc.

• Most, but not all, models should now download to the Models folder so you can take your folder with you. FYI, ensuring that all models do so is a work in progress, the goal being to carry all of the necessary files + program on a flash drive.

Current Chat and Vision Models

6.3.0 - whisper upgrade

NOTE

This release has been deleted a few times because of errors but this one should work now....

Updates:

• Added the large-v3 whisper model and removed large-v2.
• Added all three distil whisper model sizes.
• Ensured that all whisper model files are downloaded to the Models/whisper folder in the source code folder.
• Added error handling in metrics bar for if/when the numbers go over 100% - e.g. a model overflows the vram.
• Modified gui.py to specify the multiprocess type earlier in the script to avoid some errors.