AI search in local documents

[PangaeaProxima]

The cost of training an AI model is still quite high as far as I know. However, it is already very useful to be able to use AI for searching in your own local data. I have come across a tool that can be installed locally to search PDFs and other types of documents and that I found to work quite well: https://github.com/BBC-Esq/VectorDB-Plugin

Search in audio should also be possible, but I was not able to test this, since I don't have NVIDIA graphics.

Simply follow the installation instructions, you have to install quite a bit of supporting software and libraries. I had to install several libraries manually because their install timed out or they were still missing after the install script had run through.

Once you have launched the program GUI window, first click "Models" and select one (for the start it doesn't matter that much which) and then click "Download Selected Model". After the download is complete (Watch the console output for status) click "Create Database" and select the model in the drop down box. Enter a database name in the text field. Then click "Choose Files" and select one or more documents. If there is an error message because of lack of access rights, end the program and start it again from a console window with administrator rights. Then click "Create Database" and wait until vectorization is complete. Then click "Query Database" and select the database from the drop down menu.

Now you can enter your question in the text box; then click "Submit Question". The vector database will be searched and one or more text selections (so called "chunks") from the document(s) that most fit your question will be delivered. If you select "Chunks Only" they will be displayed directly, otherwise they will be given as input to a LLM that will formulate an answer from them. You may select an LLM vis LM Studio which has to be installed separately; if you use the selection "Local Model" one will be downloaded if you use it the first time.

[Archie]

Thanks for sharing. I have not tried it out yet, but I think everyone who has tested the off-the-shelf LLMs has noticed that it's highly biased against Holocaust revisionism. I assume this is because it's mostly synthesizing and summarizing mainstream sources which all say revisionism is bunk. Obviously, we need to find ways to teach it without it reverting to baseline. That reversion has been the biggest issue that I have found. Sometimes you can force concessions out of it on some specific point but I find there's no way to carry this "knowledge" over for long. If you could get it to "remember" everything then you might be able to make some progress, i.e., it would get more expert in whatever topic you trained it on.

Over at Unz.com, Ron has added some AI features. There is an AI summary button available now for most of the articles as well as a few custom chat bots available. Just getting it to summarize books and articles and things accurately is quite useful.

[PangaeaProxima]

Did you try https://gab.ai ? I don't think Andrew Torba has the resources to create a new model from scratch, but it seems they were able to tweak an existing model to be less biased:

[Stubble]

This may be the ridiculous holocaust claims index I need. Thank you.

I will test it on my files and see if it can help me find the random stuff I can't remember the exact place of.