The simple version
When you add a paper to Sciwand, the app can convert its text into a list of numbers that capture the meaning of the content. These number lists are called embeddings.
Think of it like a fingerprint for ideas. Two papers about "gene editing in cancer therapy" will have similar fingerprints, even if they use completely different words. This lets Sciwand understand what your papers are about, not just what words they contain.
In short: Embeddings turn text into meaning. This lets the AI find the right papers, passages, and notes - even when your search terms don't exactly match the original wording.
Why this matters for your research
Without embeddings
Search only finds exact word matches. Asking the AI about your papers means sending the full text of every article each time - slow and expensive.
With embeddings
Search understands meaning. The AI instantly identifies the most relevant passages and only sends those to the model - faster and far cheaper.
Embeddings are generated once when you add a paper. After that, every search and every AI chat query uses them instantly - no re-processing needed.
What embeddings unlock
Chat with many papers at once. Without embeddings, asking a question means the AI has to read through your entire articles from scratch. With embeddings, Sciwand instantly finds the top relevant passages across hundreds or even thousands of papers and sends only those to the AI. You can chat with your entire library in seconds.
Much lower AI costs. Embedding models are extremely cheap - roughly $0.01–0.15 per million tokens (a million tokens ≈ several hundred papers). After that one-time cost, every AI chat session sends a small fraction of your library rather than the full text. This can easily reduce your per-question AI costs by 90% or more.
Faster responses. Because less text is sent to the AI model, you get answers faster. Instead of waiting for the model to process dozens of full papers, it processes a handful of relevant paragraphs.
How much do embeddings cost?
Embedding models are among the cheapest AI services available. Here's a rough guide:
| Model | Price / 1M tokens | ~100 papers | Notes |
|---|---|---|---|
| OpenAI text-embedding-3-small | $0.02 | ~$0.01 | |
| Google text-embedding-004 | $0.025 | ~$0.01 | |
| Voyage voyage-3-lite | $0.02 | ~$0.01 | 200M tokens free |
| Ollama / LM Studio | Free | Free | Runs locally on your computer |
For most researchers, the entire library can be embedded for well under $1. Local models (Ollama, LM Studio) are completely free but require a capable computer.
When you might want embeddings off
Disabling embeddings is a valid choice too. Here's what changes:
Benefits of disabling
- Full article text is sent to the AI every time - nothing is filtered out
- Potentially more thorough answers if using a large, capable model (GPT-4, Claude, Gemini)
- No setup required - works immediately
- Zero embedding costs
Trade-offs
- Higher AI usage costs per question (full text sent each time)
- Slower responses (more text to process)
- Can only chat with a few papers at a time (model context limits)
- No semantic search across your library
Rule of thumb: If you mostly work with a few papers at a time and use a powerful model with a large context window, disabling embeddings can give you thorough, full-text answers. If you have a large library and want to search and chat across many papers quickly and cheaply, turn embeddings on.
Quick summary
Embeddings ON - cheap one-time processing, instant semantic search, chat with your whole library, lower per-question AI costs, faster responses.
Embeddings OFF - no setup, full text goes to AI every time (more thorough if using a top-tier model), but higher costs, slower, and limited to a few papers per chat.
Most users with more than a handful of papers will benefit from enabling embeddings. You can always turn them on or off later - your library data is never affected.