The goal of this project was to explore how IRIS Database implementation can satisfy search requirements leveraging AI.

Chat-with-AI is a familiar, simple textual interface, achieving higher productivity.

Question: What happens if a need cannot be easily expressed in textual form?

Hypothesis:
A conversation with a colleague could be:

“Hey what was that tune earlier?
The one that went “do do dee doo da”?

Instead of text input, can we explore “Sound” as input, and in a novel way apply new IRIS AI search powers to achieve this?

The Interface

The TOOT interface consists of a button.

TOOT Interface

Press the button
Whistle or sing a tune
Review the results

Results

AI collateral

Tailored tokenizer vocabulary
Unsupervised base model training
Supervised Similarity training with SentenceTransformer
Synthetic data generation for similarity training

TensorBoard SentenceTransformer

TensorBoard Unsupervised

IRIS product areas exercised:

Vector Embedding Search
Loading Data Embeddings with record insert
Extending the IRIS Embedding Interface
- HotModel parameter - Keeps transformer model loaded in memory between invocations of encoding record inserts ( approximate 2 to 4 times speedup )
- Used Cuda Dockerfile version to run faster insert speed in development
Python in BPL
Interop with shell utilities

SearchWorkflow

Docker recipie

There were a lot of dependencies packed in, to get this working out of the container.
Took a lot of rebuild and testing to get right.
Anticipate some reuse value in the Dockerfile investment.

Local packaged model
Encoding data
Production
Shell integrated utilities

Installation

Note: GitHub constrains max upload filesize to 25MB.
Due to time constraints am asking the developer please to use for example 7-zip to unpack the spanned archives.
Later can review and resolve this in a better way. Possibly direct download via Docker file.
Thank you for your patience.
Note: The dependencies involved result in a docker large image file ( 8.15GB )

Download the repository files
Unzip the spanning archives in “/src/data”
Unzip the spanning archives in “/model”

In the main directory on commandline run:

docker build -t toot .
docker run --rm --name toot -d -p 1972:1972 -p 52773:52773 -p 443:443 toot
docker exec -u root toot apachectl start

The docker appliance will take a long time to build. However the image will start and restart quickly.
The Ensemble Integration Production is registered to start automatically.
To explore the Python in BPL use the management portal.

Application access is provided by TLS. This is a security constraint of web browsers.
Only web pages delivered over HTTPS may use Audio and Video recording capabilities.

The ambition was high. Many discoveries. Would be good to write some new articles on the developer forum