Skip to content

Latest commit

 

History

History
202 lines (159 loc) · 11.4 KB

README.md

File metadata and controls

202 lines (159 loc) · 11.4 KB
Raw

🚀 Supercharged Vector Database!

Requirements  •  Installation  •  Using the Program  •  Request a Feature or Report a Bug  •  Contact

This repository allows you to create and search a vector database for relevant context across a wide variety of documents and then get a response from the large language model that's more accurate. This is commonly referred to as "retrieval augmented generation" (RAG) and it drastically reduces hallucinations from the LLM! You can watch an introductory Video or read a Medium article about the program.

At a Glance

Feature Details
General text extraction .pdf .docx .epub .txt .html .enex .eml .msg .csv .xls .xlsx .rtf .odt
"Vision" models to create image summaries .png .jpg .jpeg .bmp .gif .tif .tiff
Transcribe audio files to text .mp3 .wav .m4a .ogg .wma .flac and more...
Type or speak your query Using a powerful WhisperS2T voice recorder
Get a response from an LLM LM Studio Local Models Chat GPT (coming soon)
Text to speech playback of the LLM's response Bark WhisperSpeech ChatTTS Google TTS
CPU and Nvidia GPU support Looking for testers or contributors for AMD and Intel GPUs as well as Metal/MPS/MLX

Requirements

🐍 Python 3.11  •  📁 Git  •  📁 Git LFS  •  🌐 Pandoc  •  🛠️ Compiler

The above link downloads Visual Studio as an example. Make sure to install the required SDKs, however.

EXAMPLE error when no compiler installed:
EXAMPLE of installing the correct SDKs:

Back to Top

Installation

Step 1

Download the latest "release," extract its contents, and open the "src" folder:

  • NOTE: If you clone this repository you will development versions, which may or may not be stable.

The last attempt to support 🐧 Linux and 🍎 MacOS is Release v3.5.2. Make sure and follow the readme.md instructions there.

Step 2

Within the src folder, create a virtual environment:

python -m venv .

Step 3

Activate the virtual environment:

.\Scripts\activate

Step 4

Run the setup script:

Only for Windows for now.

python setup_windows.py

Step 5 - 🔥IMPORTANT🔥

In order to use the Ask Jeeves functionality you must:

  1. Go into the Assets folder;
  2. Right click on koboldcpp_nocuda.exe;
  3. Check the "Unblock" checkbox
  4. Click OK.

If the "unblock" checkbox is not visible for whatever reason, another option is to doubleclick koboldcpp_nocuda.exe, select the .gguf file within the Assets directory, and start the program. This should (at least on Windows) attempt to start the Kobold program, which will trigger an option to "allow" it and/or create an exception to "Windows Defender" on your computer. Select "Allow" or whatever other message you receive, which will enable it for all future interactions. Please note that you should do this before trying to run the Ask Jeeves functionality in this program; otherwise, it may not work.

Submit a Github Issue if you encounter any problems as Ask Jeeves is a relatively new feature.

Back to Top

🖥️Usage🖥️

🔥Important🔥 for more detailed instructions just Ask Jeeves!

Activate the virtual environment and start the program

Every time you want to use the program you must activate the virtual environment:

.\Scripts\activate

python gui.py

Download the vector model

  • Select and download a vector/embedding model from the Models Tab.

Create a vector database

This program extracts the text from a variety of file types and puts them into the vector database. It also allows you to create summarizes of images and transcriptions of audio files to be put into the database.

Entering General File Types

In the Create Database tab, select files you want to add to the database. You can click the Choose Files button as many times as you want.

Entering Images

This program uses "vision" models to create summaries of images, which can then be entered into the database and searched. Before inputting images, I highly recommend that you test the various vision models for the one you like the most.

To test a vision model:

  1. From the Create Database tab, select one or more images.
  2. From the Settings tab, select the vision model you want to test.
  3. From the Tools tab, process the images.

After determining which vision model you like, add images to the database by selecting them from the Create Database tab like any other file. When you eventually create the database they will be automatically processed.

Entering Audio Files

Audio files can be transcribed and put into the database to be searched. Before transcribing a long audio file, I highly recommend testing the various Whisper models on a shorter audio file as well as experimenting with different batch settings. Your goal should be to use as large of a Whisper model as your GPU supports and then adjust the batch size to keep the VRAM usage within your available VRAM.

To test optimal settings:

  1. Within the Tools tab, select a short audio file.
  2. Select a Whisper model.
  3. Process the audio file.
  4. Within the Create Database tab, doubleclick the transcription that was just created.
  5. Skim the page content field to get a sense of whether the transcription is accurate enough for your use-case or if you need to selecta more accurate Whisper model.

Once you've obtained the optimal settings for your system, it's time to transcribe an audio file into the database:

  1. Within the Create Database tab, delete any transcriptions you don't want entered into the database.
  2. Create new transcriptions you want entered (repeate for multiple files).

    Batch processing is not yet available.

Actually Creating The Database

  • Download a vector model from the Models tab.
  • Within the Create Database tab, create the database.

Manging the Database

  • The Manage Database tab allows you to view the contents of all databases that you've created and delete them if you want.

Query a Database (No LM Studio)

  • In the Query Database tab, select the database you want to use from the pulldown menu.
  • Enter your question by typing it or using the Record Question button.
  • Check the chunks only checkbox to only receive the relevant contexts.
  • Click Submit Question.
    • In the Settings tab, you can change multiple settings regarding querying the database. More information can be found in the User Guide.

Query a Database with a Response From LM Studio

This program gets relevant chunks from the vector database and forwarding them - along with your question - to LM Studio for an answer!

  • Perform the above steps regarding entering a question and choosing settings, but make sure that Chunks Only is 🔥UNCHECKED🔥.
  • Start LM Studio and go to the Server tab on the left.
  • Load a model.
  • Turn Apply Prompt Formatting to "OFF."
  • On the right side within Prompt Format, make sure that all of the following settings are blank:
    • System Message Prefix
    • System Message Suffix
    • User Message Prefix
    • User Message Suffix
  • At the top, load a model within LM Studio.
  • On the right, adjust the GPU Offload setting to your liking.
  • Within my program, go to the Settings tab, select the appropriate prompt format for the model loaded in LM Studio, click Update Settings.
  • In LM Studio, click Start Server.
  • In the Query Database tab, click Submit Question.

Back to Top

Request a Feature or Report a Bug

Feel free to report bugs or request enhancements by creating an issue on github or contacting me on the LM Studio Discord server (see below link)!

CONTACT

All suggestions (positive and negative) are welcome. "bbc@chintellalaw.com" or feel free to message me on the LM Studio Discord Server.