Humanoid Approach for Large Language Models
The Problem, The Solution and The Business
The Problem
ChatGPT , is one of the major breakthrough in Artificial Intelligence and NLP technologies. Its power relies basically on recent advances in Conversational-AI and Generative Models, which generates human-like texts to answer human-like questions.
However, one of the major issues is that ChatGPT is being outdated !! Yes you have read that right ! ChatGPT used data till September 2021 , which means it uses two years old data , by the time of writing this article.
Let’s do a quick check to this problem:
The Solution
The are many solutions to this problem. One classical solution is Fine-Tuning , which is training the last set of layers of the neural network model(s) comprising the LLM models.
However, Fine-Tuning is not the best option of we want to Democratize AI , in another words, making AI technologies available to everyone. The reason is that Fine-Tuning requires domain and machine learning experience to select the right per-trained model, select the right tuning method and parameters. Furthermore, the AI Engineer or Researcher should take care of several problems like over-fitting and forgetting (Check this paper for more information)
On the other hand, a recent technique called “Retrieval Augmented Generation” (RAG) has gained a lot of momentum to solve these kind of problem. Instead of Partial-Training the model, we combine “Generative Models” which is the “Large Language Model” (LLM) with “Information Retrieval” methods to Generate results over a given set of documents.
This is related to a well know problem in Machine Learning, Artificial Intelligence and Information Retrieval which is called “Open Book Question Answering”, which can be defined as follows (definition quoted from this paper abstract):
Open Book Question Answering (OBQA) : is a subset of question answering tasks where the system aims to find answers in a given set of documents (open-book) and common knowledge about a topic.
So how LLM+RAG and OBQA approaches can solve the out-of-date problem in ChatGPT. Simply instead of Partial-Training the model of recent training , these methods do the following:
- Split documents into a set of “chunks”
( Check our mini-article about text splitting for NLP and LLM). - Retrieve “relevant” N chunks using information retrieval techniques , like
word embedding and vector search. - Run LLM to generate Results “through” the retrieved chunks.
(Check this blog for wonderful detailed explanation)
One more desirable effect of applying LLM+RAG approach is “sustainability” , which means reducing the carbon footprint of running the AI algorithm, as it is computationally is much more effective.
The Business
Given this new technology , there are a set of Business opportunities that you can leverage to either start a new product or expand your set of products. The core idea is to apply what we call the “Humanoid-LLM approach” as follows:
- Utilize the Power of LLMs : Check these two article 1 , 2 to learn more about the existing open source LLMs that you can you use , for free in your business
- Utilize the power of Similarity Search : Check this awesome lecture notes about the basics of Information Retrieval.
Also there is a set of wonderful tools that makes it easy with few lines of code : FAISS , Pinecone - Utilize Domain experience for document pool selection : Domain experts will select set of recent, trustworthy set of confidential or public set of documents to “steer” LLM modes in the right direction.
Why we think “Humanoid-LLM” process is the best for business ?
When it comes to real-life business applications , precision matters. One way to improve the “precision” of an answer is to provide the model with the “right” set of information sources. Despite the hype still ChatGPT can’t just eliminate 100s and 1000s of year of cumulative humanity experience. Hence, the best results can come by combining the best of both, not eliminating each other, and the Humanoid-LLM approach applies this concept.
Another advantage of the “Humanoid-LLM” approach is “explainability” of AI models. Despite that AI Engineers and Researchers focus only on accuracy of the mode, business stake holders focus also on “understandability” of the model, to be able to TRUST its results.
The Business
Given the opportunity to provide more “precise” and “business-steered” answers based on LLM and Retrieval based methods applied to recent, trustworthy and business related documents, the are some business opportunities that can lead to AI powered application that leverage these technologies :
1 ) Trouble shooting
For many personal-scale and industry-level appliances, precise trouble shooting is a must. These devices , are usually accompanied with a user manual and frequent updated versions of trouble shooting documents.
Relying only on the available information on the internet (Forums and so on), besides the problem that ChatGPT relies on 2 years old data, will not lead to precise answer. Hence, applying the Power of LLM + RAG will lead to more trustworthy precise answers.
Hint : Check our article to see how can we apply LLM+RAG concept using our AnswerMe API
2) Financial Documents Analysis
Understanding financial reports is a critical task for any investor or investment manager. However, it is a daunting task that takes time, even from experts. By combining the access to high quality financial reports and the power of LLM+RAG , we can get precise answers about company’s growth , cash flow status, future plan etc…
Hint : Check our article to see how can we apply LLM+RAG concept using our AnswerMe API
3 ) Customer Support
Customer support comes at the core of customer satisfaction and retention activities to any business. By combining all company’s documents , customer chat history, policies etc.. in a set of documents then feeding them to LLM+RAG when answering customer complaints, we can save huge amount of money and effort and provide high quality, precise and actionable insights to customers.
For more informational articles, LLM development and Support :
Follow us on X (Twitter) , LinkedIn
Or send us an email : info at betaflow dot ai
Leave a comment below about thought or another possible applications for LLM + RAG
