Inspiration
In recent years there have been tremendous advancements in Natural Language Processing and Language Models through the advent of Deep Learning, Self-Supervised Learning and appropriate Compute. However, these text-based models do not provide the liveliness that speech represents and remain restricted to a text-based prompt making a conversation with them unmemorable.
With this motivation, we present DashaWise, an AI Assistant powered by the capability of a language model pretrained on the internet to enliven the experience with AI language models as well as make them accessible to the visually impaired.
What it does
This app enables you to query via speech command to the most powerful language models for factual and non-factual questions. You can run it access it via a web app or via call. The speech impaired can also access it via text commands. To run it locally you just need to download the code, install dependencies and enter your OpenAI key and you will be good to go.
How we built it
The tech stack heavily relies on the tools by DashaAI - Their speech and NLU/NLG workflow powered by (few-shot trainable) Machine Learning models, the DSL scripting language and builds on their example app using Dasha. We found their tools to be very easy to use and still be flexible enough for providing freedom to developers. Our web app uses HTML, CSS and JavaScript (Node, NPM and Express).
Our Knowledge (Language Model) Prompt Engineering uses Prompt Engineering to steer OpenAI's GPT-3 towards general-purpose usability and compatibility. Our prompt engineering consists of changing the zero-shot language modelling to few-shot for increasing the performance along with tuning the stochastic beam decoder for temperature, nucleus sampling.
Challenges we ran into
The main challenges were faced in prompt Engineering of GPT-3 in conjunctions to co-align with Dasha's Neural Models (particularly the Voice API). I really enjoyed the steep learning curve with the nicely documented (and amazing tutorial) for Dasha tools
Accomplishments that we're proud of
DashaWise is remarkable at answering both factual and non-factual questions. It can also perform basic dialogue, chatting, recipe/advertisements/taglines suggestion and numerous other language generation tasks. It finally enables the especially impaired to also experience the power of language models trained on the internet's text.
What we learned
- The capabilities of Dasha's conversational-AI-as-a-service platform
- The Dasha's workflow and DSL.
- Prompt Engineering over GPT-3.
- Integrating Third-party Neural Models with Dasha
What's next for DashaWise
The direction that excites us the most is towards the dynamic generation of Dasha's DSL nodes and transitions using the GitHub Co-Pilot Code-NLP model. This allows for even faster creation of Dasha applications and will further push the ability to create Conversational-AI using Dasha towards No-Code Regime.
Log in or sign up for Devpost to join the conversation.