Removed the TVs only vector store which was a sandpit dataset for a while - no longer needed.

Reduced permutations from 2 copies to 1 copy of training data to reduce training time

Tidied up comments

Internal chroma update

Adding the fine-tuning offline script

Added the local environment file and fixed a typo in the readme

Updated the readme

Added convenience function to pop N items from the end of a list

Fixed typo in comment.

Added the display of the question typed into the side by side setup.

Removed debug print statements which cause extra chaff in the logs

Fixed broken LinkedIn link

Added a side by side compare for the architectures

Updated the config to make the trimming less aggressive

Updated to refresh every 10 seconds when an endpoint is starting up

Added a page to see and control (pause/resume) the model inference endpoints

Deleted the now redundant literature review placeholder and added the latest draft document

Updated page as I learnt there should only be one final document

Updated front page so that I can toggle off security and the associated message on the front page will just disappear

Updated front page

Renamed the architectures for clarity in the app

Removed now unused code which calls huggin face via the free HTTP interface - moved everything to endpoints to be equivalent.

Updated the architectures config to use the new endpoints

Added user messaging when the endpoint is down.

Merge remote-tracking branch 'origin/main'

Added option to name an inference endpoint so it is human readable

Added a new architecture component which calls hugging face via a dedicated inference end point and not the HTTP interface - needed due to the size of the fine-tuned model

Added a new architecture component which calls hugging face via a dedicated inference end point and not the HTTP interface - needed due to the size of the fine-tuned model

Tidied up space in formatting

Added regex dependency as not installed by default on huggingface

Added and configured a component to trim the architectures to remove some common junk referencing the vector retrieved documents.

Internal chroma updates to the vector store

Fixed a display issue in the markdown with dollar signs in prices

Swicthed to use all_products as the sample vector store for the streamlit app

Added new all_products vector store

Enabled GIT LFS tracking sqlite3 files

Fixed bug aligning the names of the vector stores on train versus inference

Fixed bug aligning the names of the vector stores on train versus inference

Added the RAG training script to be run offline

Switched the list joiner to not lowercase the inputs

Added a couple of convenience methods to the classes to get the representations I needed, but put that in the classes

Added a utility to join lists of items as strings with comma/and separation

Added image for baseline architecture

Adding the image file missing from last commit.

Added option to include a diagram of the architecture under test on the architectures page.

Updating the requirements with better-profanity so it works on the server image.

Modified the input screener to be simple profanity check.

Updated naming convention for dataset databases to be clearer.

Tweaked to remove the deterministic thread as not relevant to my project, and added safety steps for the baseline model.

Internal chroma changes - maybe optimisations based on queries