However smart the search AI gets, squeezing everything into fixed-size number lists means some searches can never return the right results.

This explores why turning documents and queries into fixed-length vectors puts a hard ceiling on which search results a retriever can ever return, however well it is trained. The corpus treats this as a proven limit, not a tuning problem.

Here is the setup. An embedding retriever turns every document and every query into a list of d numbers, then returns the k documents whose lists sit closest to the query's. Do embedding dimensions fundamentally limit retrievable document combinations? uses communication complexity theory to show that for any dimension d, only a limited number of distinct top-k result sets can ever come back. That cap grows only polynomially with d. If people can ask for combinations of documents outside that set, some of those searches are unreachable, and no amount of training fixes it.

The surprising part is how little it takes to hit the wall. The limit shows up on trivially simple retrieval tasks. It also holds when the embeddings are optimized directly on the test data, so this isn't bad training data or an undertrained model. The geometry can't hold all the combinations. Where do retrieval systems fail and why? lists this as one of three structural failures in retrieval, next to fixed-interval retrieval triggers and embeddings that measure association rather than relevance. Its conclusion is that these need different retrieval approaches, and tuning won't do.

The corpus points to two ways out, and both stop squeezing everything through one vector. One is to skip the vector. Can direct corpus search beat embedding-based retrieval? trains an agent to search raw text with shell commands like grep. It beats dense embeddings on multi-hop questions with entity constraints, where embeddings blur similar-looking entities together. The other is to keep more information than one vector holds. Can verification separate structural near-misses from topical matches? uses cheap pooled-vector recall first, then a small verifier that reads the full token-to-token similarity map. It rejects structural near-misses that compressed matching can't tell apart, because it sees the whole interaction pattern rather than a summary.

A cousin of this limit turns up outside retrieval. Can state-space models match transformers at copying and retrieval? shows that state-space models are fundamentally limited at copying and looking things up, because they carry a fixed-size latent state. That result is about model architecture, not embeddings, so the link is my reading. But the pattern is the same: a fixed-size summary can't always support the queries you later ask of it, and the more reliable systems keep access to the full text.