Your new AI Generative Digital Assistant designed for Oracle Fusion (ODA + GenAI)

In a previous post I covered how to use local LLMs to provide - quick automated translations for Oracle Fusion using both the new browser built-in window.ai powered by gemini nano and then a custom models downloaded into the browser using transformers.js from hugging face.  

Translate the Oracle Fusion interface with a local LLM
In one my last posts I covered how you can now use LLMs locally within the browser without the need to connect to a LLM Service - using different approaches - ie the new internal browser LLM accessed via window.ai and also using custom LLMs ran locally as a

Over the last couple of months; I've been working on and off - on a new approach and productised solution with the Fishbowl Team called GDA (Your New Generative Digital Assistant) that enhances Oracles Digital Assistant for Fusion with local LLMs to improve response times and making conversations feel more natural using the latest LLMs designed for intent recognition and entity extraction that can be enhanced and retrained if needed on targeted datasets.

Essentially expanding and improving the conversation capabilities of ODA with improved LLMs with the added ability to allow you to bring and introduce your own trained models.

Why use local LLMs?

You don't need to use local LLMs and depending on the complexity of your conversation flow you can configure this solution to interact directly with a third party LLM either your from your own server or Azure Open AI, Cohere etc.. GDA is designed as an added layer on top of ODA and handles conversation from the user to ODA and from ODA back to the user.

In my opinion local LLms are perfect for chatbots/ Digital Assistant using the latest AI capabilities that are now part of the browser that maintain privacy and security for your conversations.

Tell me more about browser LLMs

Gemini Nano, the smallest version of Google's Gemini AI family, is designed for on-device applications, making it suitable for tasks like intent classification and entity matching, particularly when performance and efficiency are critical. The Gemini Nano model is built to run on mobile devices with low latency and high efficiency, which can be a significant advantage for applications needing real-time processing, such as intent recognition in chatbots or entity matching in mobile apps.

Intent Classification:

Gemini Nano performs well for intent classification, as it is designed to handle text-to-text tasks efficiently. While it may not achieve the same level of accuracy as larger models like Gemini Ultra or Gemini Pro, it is optimized for local, on-device processing. This makes it particularly valuable in scenarios where privacy, offline access, or reduced latency is a priority, such as personal assistant apps or mobile chatbots​

Entity Matching:

For tasks like entity matching, Gemini Nano's text processing capabilities can provide fast and accurate results. Its ability to execute on-device makes it a practical choice for applications where cloud-based models may be impractical due to latency or cost concerns. Additionally, with its support for fine-tuning via LoRA blocks, you can adapt Gemini Nano to more specialized entity-matching tasks, improving performance for domain-specific use cases​.

However, for more complex or large-scale skills, you might want to consider using one of the larger models, as they offer stronger benchmarking results on tasks requiring more sophisticated reasoning and processing​.

In summary, Gemini Nano is a great option for on-device applications that prioritize efficiency and privacy, and need a tool for intent classification and entity matching tasks.

Initial Feedback on this approach

If you work with the Oracle Digital Assistant platform - then you know how much fun they can be with weekly conversational log reviews; retraining and enhancing the skill to try to improve the overall user experience and flow interaction..

Using Generative AI in-front of ODA..
.. gets rid of all that pain!!!

Lets' forget about the weekly log reviews, planning, setting up and managing huge utterance datasets that are used to help target user intents which are often poorly implemented by developers that don't fully understand conversational interactions.

Oh.. did I mention this approach also handles all your multilingual needs!! The LLM in-front of ODA acts as your translation engine - so all you need to do is create and manage your flow in English and let the LLM auto-detect the language and act as your personal 2way translator..

Benchmarking

It's coming soon.. but ask the Fishbowl Sales Team for time to see a demo of it in action for yourself..

How does it work?

Under the covers the interface is built using Oracle JET; their Javascript Extension Toolkit which is the framework used for Visual Builder Studio and the Fusion Interface today.

The app runs on-top of the Oracle Digital Assistant SDK headlessly providing the flexibility to control the entire user experience and brand - as the first release was designed for Fusion - we opted to use the Redwood Design System and will be exposing themes in future releases to enable the component to be deployed and branded on B2B, B2C Portals and Sites.

Behind the scenes we are using multiple LLM interactions and depending on your conversational flow use case you can change the models used or provide additional fine-tuning via LoRA blocks, for more specialized entity-matching tasks, improving performance for more domain-specific use cases​ (We are still experimenting with this approach server side)

Can I use it with my existing skills

Yes!.. but when they say less is more this actually comes true with the GDA app, the front-end apps local LLM is pre-trained on your exported YAML Flows and custom entities for validation.  

Utterances are cut right back down in ODA with the GDA app using a custom pre-prompt model that we have designed - this means that the local LLM understands the designed user conversational flow from ODA - this can be tweaked and customised to provide more flexibility if issues do arrive that we may not have caught.

All you need are 2 Utterances per Intent!

Does it work with the OOTB ODA HCM Fusion Solution?

Yes.. You first have to deploy the Fishbowl GDA app to the Fusion interface - which can then interact with the Fusion ODA Services once configured...

..however if you have customised the OOTB UI these updates will be lost and replaced with the new GDA interface.

The GDA Roadmap

We are near our first release and hoping to officially launch in the next week or two..🤞  

Phase 2

Will be incorporating Fishbowl Solutions other Gen AI offerings:

Generative AI Solutions | Fishbowl Solutions

What does this mean?
We will be going further than skills and conversational flows incorporating with our own AI Agents for content based solutions within the GDA Interface and interacting directly with those services.

The Fishbowl Solutions team have 2 core offerings around Content RAG and Generative AI one is designed around search ingestion allowing you to index content from multiple sources and allows for your teams to have conversations with Generative AI around the content indexed which could be from any document management platform or repository ie sharepoint, Jira, google or other third party..

.. And the second is around our new AI Platform and interface designed for WebCenter Content with the new Oracle WebCenter Content Fusion adapters allowing you to bring WebCenter Content Cloud into Fusion and enabling users to more effectively return their content within a chat interface.

Phase 3

MultiChannel and Branding.

Integration with MS Teams and as standalone desktop app and enabling brand support to allow customers to include the GDA within their B2B or B2C Sites or Portals or employee desktop. Exposing a conversational experience from Fusion such as Oracle HCM to the desktop so you no longer need to open and access a browser simply launch the app and start talking with your AI Agent.

The future of GEN AI with Digital Assistants is coming

Although it's not out today you can see Oracle are heavily committed to this approach - first with their AI-Powered Clinical Digital Assistant:

And I'm sure elements of what comes out with the Oracle Clinical Digital Assistant will be making its way into the Oracle Digital Assistant.. So be sure to follow updates from that team to get an idea of what could be coming in the near future..

Interested in seeing the new GDA app running in Fusion

Reach out to the Fishbowl Solutions Sales team and check out the latest new Generative AI Solutions they are working on:

Generative AI Solutions | Fishbowl Solutions
Contact Us | Fishbowl Solutions

... Keep an eye out for my next post and demo of Agents working with ODA to provided a guided conversational experience built with flow designer and enhanced with AI Agents.