Meaning as numbers
How “refund chahiye” finds the right English sentence.
An embedding turns text into a long list of numbers, placed so that similar meanings land close together. The language doesn't matter; the meaning does.
Each dot is one of the 12 sentences: 3,072 numbers each, flattened to 2D. Rings mark the best matches.
- My parcel still hasn't arrived.0.778
- Mera order abhi tak nahi aaya.0.736
- मेरा पार्सल अभी तक नहीं आया।0.715
- I want my money back for this.0.587
- Paise wapas kab milenge?0.577
- Please refund my broken charger.0.569
Cosine similarity between the query's embedding and each sentence's. No shared words needed.
Real gemini-embedding-2 embeddings, 2026-10-21
“refund chahiye” finds “Paise wapas kab milenge?” first, then the English refund sentences, without a single shared word. This is how agents search documents, memories and even tools.
Why does “package is late” find the Hindi sentence about a missing parcel?