Semantic searches on our databases

[mikehquan19]

Certain things in our databases are texturally rich so it's a good idea for the team explore if we can perform semantic search for API.

The user can query “An interesting class where I can learn about foundation of cybersecurity and networking” or “Elective class to take after we take machine learning”, or “Events that involve dancing and playing games together”. Then we can search for the vector database to get those classes or events.

If anyone wants to share their ideas or opinions on this potential feature, that would be appreciated.

[Kumud-Arora]

How about combining semantic search with structured filters (like department, prerequisites, etc.) so results don’t get too broad?

[mikehquan19]

Can you elaborate on what you mean by combination?

Is it like the user put in their description, and at the same time specify structured values? For example,

{
  "prompt": "The course that can do this, that, and the other",
  "filters": {
    "catalog_year": "2017",
    "offering_frequency": "Y",
    ...
  }
}

Or you mean we would have to parse the filters from the user's natural language prompt?

[Kumud-Arora]

Yes, I was thinking of the first approach you mentioned, user put in their description and at the same time specify structured values. I just thought that would be helpful in a situation where a user mentions they would like to take a "class similar to intro to machine learning" and semantic search might result in all AI/ML related classes offered regardless of whether they would be offered next semester or have additional prerequisites. I was thinking that if we combined semantic search with structured filters (e.g., department = CS, level = 4000, prerequisites satisfied = true, term = Fall 2026, etc) just like in your example, we can narrow results. Parsing the filters from the users nlp prompt could also work!

[mikehquan19]

Yeah the first approach would be something we seriously pursue because Mongo supports that.
We probably won't pursue the filters parsing approach because we don't want to prompt the model to parse the nlp query every endpoint lol. Considering that we are an api and not a website, someone could just ask me the key and spam that endpoint and the bill is going to skyrocket.

[mikehquan19]

I'm thinking of vector embedding the database but not sure how the results would come out. If you have some experience or would like to take a stab that would be cool.

We would probably just use MongoVector. It's still in preview, which means functionality might change in the future but that won't be much of problem I suppose