Updates

Updated dependencies to be current:

• jupyterlab==4.1.2
• requests==2.31.0
• dvc==3.48.1
• dvc-s3==3.0.1
• pandas==1.5.0
• pyarrow==9.0.0
• gradio==4.19.2
• torch==2.2.1
• sentence-transformers==2.5.1
• bertopic==0.16.0

LLMSherpa

Pulled in LLMSherpa to help with parsing multiple types of documents (specifically for PDFs) into chunks best to use for RAG.

To work locally, run the docker container locally docker run -p 5010:5001 ghcr.io/nlmatics/nlm-ingestor:latest

ChromaDB

Switched the storage of the document embeddings to use chromadb instead of the custom approach we built initially. Interesting metrics:

• Abstract Embeddings stored by numpy uses 814M of storage
• Same embeddings stored in chromadb along with the text value uses 8G of storage

spaceapps2022

Huntsville AI submission for Space Apps 2022

First thoughts:

1. Get something that minimally meets the requirements of the challenge:

   1. Download a bunch of PDFs from the NASA server, initially throw into a /data folder here. Probably follow a kedro type layout, even if not using kedro.
   2. Figure out how to extract text from PDF - currently looking at several options for this - leading candidate is PyPDF2 since it provides an easy way to also pull images
   3. Crank up a summarization, topic mapping, word cloud, etc to create something from this word soup
   4. Spin up a streamlit to display - have streamlit host based on this repo

2. Get fancy - lots of thoughts here - feel free to throw more in

   1. Graph representation with documents and attributes and some fancy (already built & free) way to visualize and traverse
   2. Start grabbing a few videos/recordings from the same NASA server - transcribe and store text - then follow similar approach to PDF
   3. Extract images from PDFs and run object detection - add these to the list of keywords for the document.

3. Somewhere, we'll have to make a video to present - not sure how to go about that except for doing something cheap or stupid on my phone.