---
title: "Vector embeddings & language: how models turn words into geometry"
url: https://daily.dev/posts/vector-embeddings-language-how-models-turn-words-into-geometry-io9eekrza
source_url: https://redis.io/blog/text-embeddings-how-language-becomes-vectors
type: article
source: "Redis"
published: 2026-08-13T01:01:25.714Z
updated: 2026-08-13T01:01:52.033Z
tags: ["nlp", "rag", "redis", "vector-search"]
reading_time: 8
upvotes: 0
comments: 0
language: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Vector embeddings & language: how models turn words into geometry

**[Redis](https://daily.dev/sources/redislabs)** · 8 min read · 0 upvotes · 0 comments

## Summary

An explainer on how text embedding models translate words into numerical vectors so that semantically related terms land close together in space, unlike keyword matching. Covers the distributional hypothesis, why language is harder to embed than pixels due to ambiguity and context, and traces the evolution from static models (word2vec, GloVe) to contextual models (BERT, Sentence-BERT). Also explains what embedding geometry encodes (analogies, relationships, bias) and touches on ANN search (HNSW) for fast retrieval at scale, closing with a pitch for Redis as a real-time vector storage and retrieval layer via Redis Iris.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://redis.io/blog/text-embeddings-how-language-becomes-vectors>

## Questions this post answers

### Why do embeddings for 'not good' end up similar to 'good' instead of opposite?

Negation is a known weakness in text embedding models because phrases like 'not good' tend to appear in similar surrounding contexts as 'good', so models trained on co-occurrence patterns often produce similar vector representations for phrases that mean the opposite. This is a documented limitation even in modern embedding models, not just older ones.

_Anyone debugging semantic search relevance issues can find related discussions curated on daily.dev._

### What is the difference between word2vec and BERT for generating word embeddings?

Word2vec (2013) produces one static vector per word regardless of context, so a word like 'bank' collapses onto a single point blending its financial and river senses. BERT (2018) reads a whole sentence bidirectionally and generates a distinct contextual vector per word occurrence, so 'bank' next to 'pond' differs from 'bank' next to 'loan'. GloVe (2014) is similar to word2vec but factorizes global co-occurrence counts across a corpus instead of predicting neighboring words.

_Developers choosing an embedding approach for search or RAG can track these comparisons on daily.dev._

### How many dimensions do dense vector embeddings typically have?

Dense vector embeddings typically have between 50 and 1,000 dimensions, according to standard NLP references, with modern models varying widely in size depending on how much semantic detail they encode. Individual dimensions usually carry no clear human-interpretable meaning; what matters is the relative position of points in the space, not any single value.

_Engineers sizing embedding models for their retrieval stack can follow this kind of detail on daily.dev._

## Similar posts on daily.dev

- [Embeddings Explained: How AI Understands Meaning](https://daily.dev/posts/embeddings-explained-how-ai-understands-meaning-ixv4nljq5) · Medium · 0 upvotes · 0 comments

---

Tags: [#nlp](https://daily.dev/tags/nlp), [#rag](https://daily.dev/tags/rag), [#redis](https://daily.dev/tags/redis), [#vector-search](https://daily.dev/tags/vector-search)

[View this post on daily.dev](https://daily.dev/posts/vector-embeddings-language-how-models-turn-words-into-geometry-io9eekrza)