The biggest problem with retrieval is actually semantic relevance. I think most embedding models don't really capture sentence-level semantic content and instead act more like bag-of-words models averaging local word-level information.

Consider this simple test I’ve been running:

Anchor: “A background service listens to a task queue and processes incoming data payloads using a custom rules engine before persisting output to a local SQLite database.”

Option A (Lexical Match): “A background service listens to a message queue and processes outgoing authentication tokens using a custom hash function before transmitting output to a local SQLite database.”

Option B (Semantic Match): “An asynchronous worker fetches jobs from a scheduling channel, transforms each record according to a user-defined logic system, and saves the results to an embedded relational data store on disk.”

Any decent LLM (e.g., Gemini 2.5 Pro, GPT-4/5) immediately knows that the Anchor and Option B describe the same concept just with different words. But when I test embedding models like gemini-embedding-001 (currently top of MTEB), they consistently rate Option A as more similar measured by cosine similarity. They’re getting tricked by surface-level word overlap.

I put together a small GitHub repo that uses ChatGPT to generate and test these “semantic triplets:

https://github.com/semvec/embedstresstest

gemini-embedding-001 (current #1 on MTEB leaderboard ) scored close to 0% on these adversarial examples.

The repo is unpolished at the moment but it gets the idea across and everything is reproducible.

Anyway, did anyone else notice this problem?

I’m not sure what the “biggest” problem is, but I do think diversity is vastly underappreciated compared to relevance.

You can have maximally relevant search results that are horrible. Because most users (and LLMs) want to understand the range of options, not just one type of relevant option.

Search for “shoes” and only see athletic shoes is a bad experience. You’ll sell more shoes, and keep the user engaged, if you show a diverse range of shoes.

I liked how Karpathy explained part of this problem as "silent collapse" in his recent Dwarkesh podcast. Meaning the models tend to fall into a local minima situation of using a few output wording templates for a large number of similar questions, and this lack of entropy diversity it becomes a tough hard to detect problem when doing distillation or synthetic data generation in general. These algorithms as nice python functions are also useful repurposed for labeling parts of ontology and topic clusters etc [1]. Will definitely star and keep an eye on the repo !

[1] https://jina.ai/news/submodular-optimization-for-text-select...

Nice, I actually read that Jina article when it was published, but forgot they use facility location as well! The saturated coverage algorithm looks pretty interesting, I'll have a look at how feasible it would be to add that to Pyversity.

Producing a diverse list of results may still help in a couple of ways here.

* If there are a lot of lexical matches, real semantic matches may still be in the list but far down the list. A diverse set of, say, 20 results may have a better chance of including a semantic match than the top 20 results by some score.

* There might be a lot of semantic matches, but a vast majority of the semantic matches follow a particular viewpoint. A diverse set of results has a better chance of including the viewpoint that solves the problem.

Yes, semantic matching is important, but this is solving an orthogonal and complementary problem. Both are important.

This seems like a good template to generate synthetic data, with positive/negative examples, allowing an embedding model to be aligned more semantically to underlying concepts.

Anyways, I'd hope reranking models do better, have you tried those?

Might also be useful for dataset curation, or even just prompt engineering. For example when training a classification task and picking a diverse set of examples for training or evaluation.

True, I think that's also a great usecase! Though these algorithms likely won't scale to very large datasets (e.g. millions of samples), but for smaller datasets, like fine-tuning sets, I think this would work very well. I've worked on something similar in the past that works for larger datasets (semantic deduplication: https://github.com/MinishLab/semhash).

It would be better if there are some real-world performance tests and comparisons with other embedding-only search methods.

That's indeed something I plan to add in the near future. I'll probably add a tutorial as well to showcase how you can use this with e.g. sentence transformers. There's some pretty good benchmarks in the paper that I used as inspiration for some of these algorithms: https://arxiv.org/pdf/1709.05135, I'll most likely try to reproduce some of these.