AI - I Don't Want a Chatbot... What Can I USE????

Chatbots are fun, and I like to use LLMs to help me write articles. BUT, while that is cute and all, what can I use to really help my business out? What is going on out there, BESIDES chatbots??? Interestingly enough, a quiet revolution is occuring in the area of Search.

Finding information, simple data retrieval is becoming a bigger and bigger challenge. For the longest time, we have been relying on keyword retrieval. You know what I'm talking about... Put your terms in the search bar and pray some good results come back.

That was improved with metadata refining, the good old Amazon checkboxes. However using a metadata based retrieval was reserved for only those who spent time learning query languages. Not the normal person.

In walks in AI and a new way of indexing information. The Vector search!

Alright, so what is a vector search? LLMs have the ability to look at the files, images, video, text, etc. and transform them into a series of numbers. Our content, in the search, is now represented as numeric values rather than the bits of encoded data that they once were. The search engine now can use comparisons of those numbers to determine what is closest to what is being asked for.

Let's do an example. Let's look at apples and oranges.

When an apple is transformed by the LLM in to a vector, it looks at what it has and the context around it. The apple is red, the apple is round, the apple is a fruit, the apple comes from a tree, the apple has an outer skin, the apple is edable. All that stuff goes in to the formation of the Vector number sequence. Let's say that is 1,2,3,4,5 (tasty, red, apple, edible, fruit).

Now we put in an orange. That goes in as 1,3,7,4,5 (tasty, orange, orange, edible, fruit).

What this enables the system to do is to very quickly search on what is the most similar to the query being asked for. Our query can now be based not on keywords, but on what we want. Where before if we put in "Fruit that is like an orange but not an orange" the keyword search would be confused and find us a bunch of oranges. The Vector search, however, changes the query in to a vector representation, and looks for what is most similar in the index.

In our example, the vector for the query comes out to be something that matches with apples because they have similar qualities to oranges, tasty, edible, fruit, but aren't oranges.

Now instead of 5 properties, think of hundreds. This is how the vector search is able to measure similarities and dissimilarities.

The next innovation to come recently is Semantic Ranking. That sounds really fancy... because it is. Microsoft and other AI leaders have been pushing searches through Machine Learning (ML) for a good long time now. They have created models that are able to look at the context of a query, along with its vector similarities and rank results so that the most relevant searches in terms of vector result AND language context.

How does this work? Well... Let's look at the term "Capital." It has many different meanings depending on the context in which it is used. Capital in a financial discussion has a completely different meaning in a criminal discussion, and still a different meaning when talking about letters. Capital's meaning becomes even more muddled when just talking about Politics. Is it the building? The city? Capital of a state? Of a country?

Semantic Ranking helps with this issue. It looks for context and relatedness to rank matches higher that make the most sense for the query.

Using all of these methods together allows us to create relevant, natural language queries on data, and get accurate search returns.

For more of a technical explanation, Check out this awesome article from Alec Berntson at Microsoft, on how Azure AI Search Hybrid Search outperforms the other search methods.

So... breaking it all down. What are some of the MOST exciting things that you can start to use AI for, that aren't chatbots? AI Search can increase your users productivity by enabling them to find relevant items quicker and more accurately, while use natural language queries to do so.