Eagle Search is a website that uses semantic search to enable Boston College to search through all 2,080 courses for the Fall 2023 semester. First, in Python, I successfully replicated the HTTP OAUTH request sequence to parse course descriptions, titles, instructors, class times, locations, and seat counts from the BC registration site. Then, using OpenAI’s ada-embedding-002 model, I batch embedded and upserted 100 courses at a time to a Pinecone Vector Database. Each user’s query is then embedded with the same model and using Pinecone’s query function checked via cosine similarity against all 2,080 courses returning the top 100 semantic matches. The website has served over 250 students since launch less than 2 months ago and executed more than 600 search queries.
Reactive Resume is an interactive way to build multiple resumes with AI assistance. It functions as a platform for students with diverse interests to store their wide array of work, extracurricular, and research experiences. Then, they can simply click to add these experiences and generate a professionally formatted resume. OpenAI’s davinci-003 GPT-3 model is used to generate professional bullet points for users given the context of their position title and company. The prompt below is used where “investment banker” is replaced by the user’s inputted position title and “JPM Chase” is replaced by their inputted company. “/u2022” was used to reliably get GPT-3 to generate unicode bullet points needed for the rendering within the text area.
GitHub - emfastic/reactive-resume-nextjs: A website to build resumes for students.
Promptable AI is an open source repository to make LLM’s more composable, similar to Langchain, allowing users to swap out different foundational LLMs seamlessly in Typescript. I built the first version of the Data Loader Hub, which enabled support for embedding a user’s Slack Channel, Google Sheet, etc. I also provided developer support in Discord and created the first example of how to use Promptable in Nextjs when making API calls (shown in my approved PRs).
GitHub - cfortuner/promptable: Build LLM apps in Typescript/Javascript. 🧑💻 🧑💻 🧑💻 🚀 🚀 🚀
Pull requests · cfortuner/promptable
Final Form was built by myself, Nihaar Shah, and Steven Liss at the first NYC hackathon on Generative AI sponsored by Contrary Capital, Nomic AI (creators of GPT4All), Betaworks, and other notable investors/builders. Final Form is a Chrome extension and web app (unpublished but the unpack should work) that ingests text data on a user and is able to autofill text inputs on any website through a combination of semantic search and GPT-4. In our demo for the hackathon, we loaded all of Steve’s company data into the web app, which it then split using a Langchain text splitter, embedded these chunks with OpenAI’s embedding-ada-002 model, and uploaded them to a Pinecone Database. We then completed the YCombinator application as if we were Steve by just clicking our AI completion button next to each text input. At a high level, the button click triggers a semantic search for an answer to that question. However, below the surface, this implements HyDE from the “Precise Zero-Shot Dense Retrieval without Relevance Labels” paper in my literature review. This uses GPT-4 to generate a fake answer to the question which, when coupled with the question results in a higher, more relevant semantic match to a chunk of the uploaded data. This context is then fed into GPT-4 to generate a factual answer to the question.
GitHub - nihaarshah/MyFinalForm: Form based applications on steroid
GitHub - sliss/final-form-extension-m3: 🚀 Minimal Chrome extension template using manifest v3 🚀