Can any devs help me find a solution for this?

Building a game and when a user guesses word X, I want to see if it's sufficiently similar to word Y. could I deploy some kind of mini, local LLM inside my codebase for this? ideally don't want to wait / pay each time for an openai api call.

founder of eventcaster (@event / events.xyz), hosting /latenightcrew, member of 🍊 dao

👨‍💻 /uno and /outside, @unofficial

yeah good idea. right now I have a dictionary of words and then use Levenshtein distance to determine it.

the issue is that it's highly subjective. for example, for some questions, answering "dog" should translate to "pet", but in cases where there are other animals as answers, it should not be accepted. So I would pass the question and other answers in when asking the LLM.

if you have a Postgres instance you could leverage ILIKE for semantic search unless you need it to be lexically similar.

guess

building /amp and /guess. prev: 2x founder, improbable, github, amazon. ntik.me

fortune cookie casts, memes, parenting, building frames, /framedl & /framedl-pro

it might be, but if i can just make one api call and move on idk if id call that overkill 😅

llm is an overkill, i think embeddings are best for this.

@nt built a fun game with word similarity as its premise — /guess — so i think he could give you poonters

yeah gonna get more creative with my dictionary here and see if it works

you can run a quantized, smaller model but also ilike, regex and levenshtein distance ma man

An LLM might be overkill here, you could just use word embedding vectors and check the similarity between the guess word and the target. Check out the usages here: https://radimrehurek.com/gensim/models/word2vec.html (grep for model.wv.most_similar)

you could load a pre-trained BERT model and get cosine similarity of words