Google made a bunch of announcements at its recent Google I/O event, including the release of its Vertex AI models for language generation - This is a direct competitor to OpenAI’s API which was released a few months ago, and has already taken the world by storm.
I decided to take 2 APIs - OpenAIs “GPT3.5” and Vertex AIs “text-bison@001” for a test drive to compare its results, specifically for a combination of a vector database enabled semantic search + LLM prompt and response. I believe knowledge retrieval and summarization is one of the fundamental use cases for large language models, and it will be built on top of a data stack consisting of something like this: Vector Databases + LLM libraries + LLM APIs + Applications
Setup:
I embedded approximately 1000 pages of documents from the Economic Survey of India 2022-23, the Global Economic Prospects report from the World Bank 2023, and the World Economic Outlook 2022 into a Chroma DB vectorstore.
Next, I created the backend infrastructure required to call the 2 APIs - langchain for OpenAI, and direct python scripts for Google Vertex AI
Finally, I created a simple frontend for this application using streamlit, which compares the query results from the 2 APIs.
The results:
I asked a simple question to my application: What are the economic prospects of different countries of the world?
The vector DB does its job and returns 3 results (as instructed) based on similarity search from the different documents:
OpenAI GPT3.5’s response is quite comprehensive, provides sources and does a great job overall.
Vertex AI’s response is also great, but it's slightly less detailed than the OpenAI response on this one:
The one area where Google is consistently faster (as of now) is its API response time:
Let’s try again, with a more specific question:
In this particular question, GPT3.5 really shines - the depth of the response and the detail in its source explanation is great
On the other hand, Vertex AI tends to quickly summarize the results and doesn't delve into as much detail. The result is still accurate though
First conclusions: Advantage OpenAI (but only slightly)
It looks like text-bison@001 has got off to a great start, but it's slightly behind GPT3.5 in language generation for this particular use case. I can only guess for the moment, but some of the reasons could be:
It isn’t as comprehensively trained as GPT3.5 - remember, OpenAI has also had 6 months of extensive user data + approximately 1.5 years since the API has been live. That’s quite a bit of a head start
The Google API also has a smaller token output at the moment (1024 tokens vs 4096 for GPT3.5) which might be causing it to shorten its outputs
On the flip side, Google has a distinct edge when it comes to speed of response compared to GPT 3.5 - but how long this will last is anybody’s guess.
Finally, it’s worth noting that I’ve not even tested the GPT-4 8k or 32k API against Vertex. We all know that GPT-4 is a full generation ahead of GPT3.5 in terms of capabilities and will provide much superior results compared to GPT3.5. The text-bison@001 isn’t even in the race
The next few months will be very interesting to see - especially on how Google tries to play catch up on its models and how this shapes up.
Hope you enjoyed this read.
留言