The companies that go DEEP on embeddings in interviews are the ones whose product directly depends on embedding quality. Three categories cover most of them.
First, **embedding API and model vendors**. Cohere, OpenAI, Voyage AI, Anthropic, and Google all ship embedding models as commercial products. Their engineers need to understand contrastive training, MTEB benchmarks, and Matryoshka representation learning at a deep level, and they expect candidates to as well.
Second, **vector database vendors**. Pinecone, Weaviate, Qdrant, Milvus, and pgvector all build infrastructure that stores and searches millions of embeddings. They care about HNSW indexes, quantisation, recall@k, and the trade-offs between flat search and approximate methods.
Third, **AI-product companies whose core feature is retrieval**. Perplexity, Glean, Notion AI, Slack AI, Harvey, and many others all live or die on embedding quality. Their interviews go deep because a bad embedding choice tanks their product.
What doesn't make the list: companies with no AI surface area, and teams within larger companies whose specific role doesn't touch embeddings. The Photoshop team uses CLIP for some features but they're hiring graphics engineers, not retrieval specialists.
Four categories of companies go deep on embeddings in 2026 interviews.
**Embedding model vendors** treat embedding internals as their core competency: Cohere, OpenAI, Voyage AI, Anthropic, Google, plus open-source teams behind BGE-M3, Jina, Nomic Embed, and E5. They expect candidates to discuss contrastive losses, hard-negative mining, multilingual training, and MTEB.
**Vector database vendors** focus on the infrastructure side: Pinecone, Weaviate, Qdrant, Milvus, MongoDB Atlas Vector, pgvector. Interviews probe HNSW parameters, quantisation strategies, recall-latency curves, and hybrid retrieval.
**Retrieval-heavy AI product companies**: Perplexity, Glean, Notion AI, Slack AI, Harvey, Hebbia, Box AI, and the RAG-focused teams inside larger orgs. They go deep because their product quality is bottlenecked by retrieval quality.
**Big-tech research and applied teams that ship embeddings**: Meta AI (DPR, Contriever), Google DeepMind (Gemini Embedding), Microsoft (E5). These ask research-level questions.
The two distractors in this question (Adobe Photoshop specifically and a no-AI startup) fail the test. Adobe uses embeddings in some image-AI features, but the Photoshop team specifically isn't an embedding-deep interview pipeline. A no-AI company won't probe embedding internals at all.
Quick reference
TL;DR
Embedding-deep interviews come from vector DB vendors, embedding API providers, and AI-product teams whose core feature is retrieval, not from companies with no AI surface area.
Think of a job interview as the company asking about what they touch every day. A bakery quizzes you on bread. A locksmith quizzes you on keys. A car-rental shop with no AI? They might ask if you can drive, not if you understand how their search ranks results, because they don't have one. The companies that go deep on embeddings are the ones whose entire business depends on them: the folks who SELL the search tech, the folks who SELL the vector storage, and the folks who BUILD apps where finding the right answer fast is the whole product. Anyone else might ask a surface question; the experts will dig in for an hour.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Preparing for an interview at a company whose product has no retrieval surface. Going deep on InfoNCE and HNSW is wasted preparation. Spend that time on what the team actually builds: backend systems, ML basics, or domain knowledge.
The team within a large company doesn't touch retrieval directly. FAANG and big-tech roles vary wildly by team. An ad-targeting role at Google won't probe embedding internals even though Google Search Quality might.
The role is a product or business-side position, not engineering. Embedding internals are an implementation detail. Product roles care about user-facing quality metrics, not loss functions.
Quick scan · 2–4 min
Everything important, quickly.
Categorise: model vendors, vector DB vendors, retrieval-product companies, big-tech applied teams. Name 2 or 3 examples per category. Eliminate the no-AI distractor on principle. Eliminate Photoshop by team vs company logic. Close with how to predict depth from public signals.
Real-world examples
Real products, models, and research that use this idea.
1Cohere ships embed-v3 as a commercial API and runs interviews focused on multilingual contrastive training.
2Pinecone, Weaviate, and Qdrant interviews probe HNSW tuning, quantisation, and hybrid retrieval design.
3Perplexity and Glean engineering interviews probe full RAG pipelines including chunking, retrieval, and rerankers.
4OpenAI's text-embedding-3 team and Voyage AI's v3 team go deep on contrastive losses and MTEB benchmarking.
5Harvey and Hebbia (legal-AI retrieval) interview for domain-specific embedding fine-tuning and evaluation.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QIf you're interviewing at a non-vendor AI-product company, what embedding topics matter most?
Sign in or go Pro to see the approach for this follow-up.
QHow does the interview depth differ between a model-provider company and a vector DB vendor?
Sign in or go Pro to see the approach for this follow-up.
QWhat signals should you look at to predict embedding depth in an interview?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Picking Adobe Photoshop because "images use embeddings." The Photoshop interview pipeline is graphics and UX, not embedding internals: the wrong team gets asked the wrong questions.