Embeddings
Text turned into numbers, so software can find things by meaning, not keywords.
Embeddings are how software finds things by meaning instead of matching words. Every "chat with your documents" feature — semantic search, find-similar, the retrieval step in an AI assistant — runs on them. They are also the cheapest layer in the stack, so whether a vendor ships them tells you if you are buying a finished system or a starter kit.
You need this when
- You want AI to answer from YOUR manuals, tickets, contracts or product descriptions — not from general training data.
- Your people search in their own words and keyword search keeps missing obvious matches.
- You need "find similar" or deduplication across thousands of records: support tickets, listings, CVs, product lines.
You can skip it when
- Pure generation — write this email, summarize this page I pasted. Nothing to retrieve, so just prompt the model directly.
- A few dozen documents. Paste them into a long-context model instead of building an ingestion pipeline.
- Exact-match lookups on invoice numbers, SKUs or dates. Use a normal database query or keyword index — embeddings are worse at exact tokens.
The long version — open when you want the depth
What it is — in one coffee-break
An embedding turns a piece of text into a long list of numbers — typically 768 to 1,536 of them — arranged so that texts with similar meaning end up with similar numbers. "How to fix a flat tire" and "changing a punctured wheel" share no keywords, but their embeddings sit right next to each other. That one trick is what powers semantic search, "find similar documents," deduplication, and the retrieval step in every AI assistant that answers from your own files.
Embeddings are produced by an embedding model — a smaller, cheaper cousin of the chat models. You feed it text, it returns the numbers, you store them. Later, you embed a question the same way and ask "which stored texts are closest?"
When you actually need it (and when you don't)
You need embeddings the moment an AI system has to find the right information inside YOUR data — manuals, tickets, contracts, product descriptions — rather than the model's general training. That's the heart of every "chat with your documents" tool. You don't need them for pure generation tasks (write this email, summarize this page you're pasting in) — there's nothing to retrieve.
The practical decisions are smaller than the jargon suggests. Per the 2026 model comparisons: OpenAI's text-embedding-3-small ($0.02 per million tokens) is the safe default that's good enough for ~90% of applications; Jina v3 matches models nine times its price; and open-weight models like BGE-M3 — the one our own news pipeline runs locally, for free — are within a few benchmark points of the paid leaders. The differences between top models matter less than how you chunk your documents.
How to recognize good vs bad implementations
On the audits, the Embeddings element (Em) exposes a real strategy split among the big platforms. Gemini scores 9 — Google sells embeddings, storage and retrieval as one integrated stack. ChatGPT scores 8 with its own well-documented embedding models. Claude scores 3 — not an oversight but a choice: Anthropic expects you to bring your own embeddings, which is fine for developers and a trap for teams who assumed it was included. In workflow tools, look for whether embedding nodes support multiple providers (n8n scores 6.5 for exactly that flexibility) rather than locking you to one.
One gotcha every practitioner learns once: you cannot mix embeddings from different models. Switch models and you re-embed everything — so pick one and stick with it, and keep your ingestion pipeline able to re-run from scratch.
What this costs
Less than almost anything else in AI. Embedding 100,000 support articles (~150M tokens) costs about $3 on OpenAI's small model. Even heavy workloads rarely pass $100/month — the storage of the resulting vectors usually costs more than producing them. Self-hosting an open model drops the per-token cost to zero, at the price of running the hardware — the trade our own pipeline makes with BGE-M3 on a local machine.
Where to see it scored
Compare Em across platforms: Gemini (9), ChatGPT (8), Mistral (7), Claude (3 — bring your own). Embeddings feed the Vector store lesson — that's where the numbers live.
Flashcards
Check yourself
1. What makes two pieces of text "close" in embedding space?
2. You switch from one embedding model to another. What must you do?
3. Which of these tools scores lowest on the Embeddings element?
4. Roughly what does it cost to embed 100,000 support articles?
5. What affects retrieval quality most in practice?
Cheat sheet
- An embedding is 768-1,536 numbers that encode meaning, not the words themselves.
- Safe default: OpenAI text-embedding-3-small, $0.02 per million tokens.
- 100k support articles (~150M tokens) embed for about $3. Storage costs more.
- Never mix models: switching means re-embedding everything from scratch.
- How you chunk documents matters more than which top-tier model you pick.
- Claude scores 3 on Em (bring your own). Gemini scores 9 as one stack.
Who actually does this well
| Best on this element | Score | Why it scored that |
|---|---|---|
| Gemini / Google | 9 | First-party text and multimodal embedding models, plus auto-embeddings built into Vector Search 2.0 — a complete, current first-party embeddings stack. |
| ChatGPT / OpenAI Platform | 8 | text-embedding-3 family is strong, cheap and multilingual; a default choice for semantic search. |
| Glean | 8 | Retrains a custom BERT-based embedding model per customer monthly, a documented ~20% search-quality lift over six months — backend infrastructure, not an exposed embeddings API. |
| OpenEvidence | 8 | Semantic retrieval across 35M+ peer-reviewed publications is the product, and the proprietary retrieval algorithm now labels why each source was picked (Highly Relevant / Top Journ |
| Relevance AI | 8 | Embeddings are the house heritage and it shows: all-mpnet-base-v2 (768 dimensions) as the default, Cohere embed-english-v3.0 and embed-multilingual-v3.0 selectable, chunk sizes of |
| Eightfold AI | 7.5 | The skills graph is the product: matching runs on inferred skills across a stated 1.6 billion career trajectories and 1.6 million skills, and it surfaces adjacent roles keyword sea |
| Cursor | 7 | Codebase indexing and semantic search are embedding-powered and central to the product; solid in practice but not exposed as a standalone embedding capability. |
| Mistral AI | 7 | mistral-embed (1024-dim, multilingual, MTEB-competitive) plus Codestral Embed for code; the 8K input window and v23.12 vintage of the text embedder are showing age. |
And the other end of the same column:
| Weakest | Score | Why it scored that |
|---|---|---|
| monday.com | 3 | Embeddings power monday's own search/ranking internally (Sentence Transformers, self-hosted, per the official AI FAQ), but there is no customer-facing embeddings API or product. |
| Guild | 2.5 | Guild's privacy notice says AI helps identify education pathways matching a learner's goals, interests and market needs, which implies matching under the hood, but no embedding or |
| DeepSeek | 2 | No embedding endpoint anywhere in the API — a standard surface for model API platforms in 2026; you must pair a third-party embedder for any semantic search. |
| Tines | 2 | No embedding model or semantic-search feature ships. The honest signal is Tines' own Library, where the document-search example sends text to Pinecone over HTTP — you bring the emb |
Scored on 78 of 144 audited tools. Every score links to the full audit and its reasoning.