Additive Manufacturing Chatbot

BLUF - I coauthored a paper published in a peer-reviewed journal pioneering a chatbot specific to laser powder bed fusion

AMGPT is an chatbot built on an open-source model enhanced with retrieval augmented generation (RAG) to boost accuracy of responses, enabling smaller models to generate truthful responses akin to a subject matter expert. RAG combines the most recent findings in a field with a LLM pretrained on older documents, making it useful for protecting privacy and staying updated with the avant garde. I coauthored a paper that was accepted by the Journal of Additive Manufacturing for the following contributions:

Developing RAG setup using an LLM for additive manufacturing queries.
Achieving higher performance with LLM-RAG system than GPT-4 in domain knowledge and specificity.
Evaluating LLM-RAG parameters including sampling temperature and top_k similarity.
Building a chat interface with Streamlit, to store chat history for RAG-enhanced LLM.

The project pipeline is summarized below:

Procuring a corpus of additive manufacturing papers, or any single topic in a domain
Embedding the documents that are first converted to .txt files or .tex to preserve equations in the paper to save memory
Chunking and caching the documents into tokens in VectorStoreIndex
Retrieve the tokens with LlamaIndex when passing a query from a user
Generate a prompt to ask the LLM based on proprietary or offline information
Feed the query and and fetched tokens into a Llama 2-7B LLM
Record the response and compare to one generated by ChatGPT-4 with an identical prompt
Collect subject matter expert blind evaluations after compling both chatbots' outputs per prompt

Manuscript

For in-depth details, please see the published paper or reach out to me!

AMGPT: A large language model for contextual querying in additive manufacturingGeneralized large language models (LLMs) such as GPT-4 may not provide specific answers to queries formulated by materials science researchers. These …