How We're Building AI Search Engines using LLM Embeddings

A lot of people have asked us for ideas of how they can leverage Large Language Models (LLMs) for their business applications. A common example is to use the native language comprehension capabilities of LLMs to find matching content. This makes LLMs an excellent tool for search!

In this video, ThinkNimble CTO William Huster demonstrates a prototype application that enables searching for job descriptions using an unstructured, English-language description of a job seeker.

The code for this demo can be found here:

Chapters
00:00 Intro - Why Build an LLM-based Search Engine?
01:00 Demo of Searching Job Descriptions
01:46 What is an Embedding?
03:06 Search by Meaning, not Content
03:52 Search with Unstructured Data
05:10 How Search with Embeddings Works
06:01 Set Up Database, Data Models, and Data
08:33 Generating Embeddings for JDs
11:04 How the Search Code Works
12:05 Creative Ways to Use Search Results
12:37 Outro - Other Use Case Examples
13:40 Outro - Final Words

Technologies used in this demo:
- Django
- PostgreSQL + pgvector
- Python sentence-transformers library

Links and Resources:

[engKanani]

excellent video, much better than tons of other long "bla bla" videos out there, thanks!

[mrdatapsycho]

Short and compact. Excellent video to get an Overview of LLM-based search.

[bracodescammer]

Awesome. I learned how to build inverted indices and this here now seems so simple, yet versatile in comparison.

[WishyIwish]

Very nice video - cool to get another perspective on RAGs and how to implement them with a very different stack.

[jeromeeusebius]

Thanks for sharing. Great video. This is a useful self contained template for a search usecase. I plan to apply this to one of my use case. I watched the video twice now and the second time around I got a much better understanding. Another interesting part, like you mentioned, is using another LLM call to potentially get an explanation for the output. One question I had, which addressed it towards the end, is how do you logically split the document to ensure consistency, not splitting in the middle of toughts/ideas, etc. One could even try different schemes, e.g., using another higher-level ML model to evaluate different splitting schemes.

Thanks once again.

[shubhamroy]

Great video man. Can you upload more videos like this explaining a bit more on the code side?

[jpops]

Thanks for the vid!! Have you guys heard of the Bittensor / Opentensor foundation?

[mvasa]

lot of potential - research papers / legal etc. Nice job is explaining . Now how would you productize this so that a customer would simply need to install it on a target system or drive where they have their files stored - you have the ability to automatically (efficiently) consume those ... and inference on any new adds?

[devd]

Great video, but I am unable to find the code in the given github link, could you please add the python script!

[SiD-hqfo]

this maybe weird question but, being beginner I'm thinking about getting into it, do you have any other platform like discord where you share and interact with user like me ? :) thank you

[RustemShaimagambetov]

To be honest, this is not a search through LLM, embeddings that generate large language models, in videos just used sentences transformers (all-MiniLM-L6-v2)

[khatharrmalkavian]

This honestly seems like the dumbest application of this technology.

[wildfotoz]

wow - there are so many things wrong with this video it's not funny. First, Excel documents are structured data, not unstructured. If the data was a bunch of resumes in word or pdf format, then you'd have unstructured data. Second, your csv files are not csv. Csv stands for comma separated values. You showed html files. I feel sorry for your clients.

[timonweb_com]

Hey, nice idea about chunks! Thanks a lot for the video!