Embeddings are the numerical vectors an AI model produces to represent text, images or other data, placed so that items with similar meaning end up close to each other. A supplier name, an invoice line or a contract clause becomes a long list of numbers that software can compare.
Closeness is usually measured with cosine similarity. The authors of Sentence-BERT showed in 2019 that finding the most similar pair among 10,000 sentences took about 65 hours with BERT and about 5 seconds with SBERT.
In finance, embeddings are used to:
- match a supplier name written differently in the ERP, the bank statement and the invoice
- find the purchase order line an invoice line describes in other words
- retrieve the contract clause relevant to an invoice before a price check
Similarity is not equality. A high score says two records are about the same thing, not that two amounts, IBANs or invoice numbers are identical: embeddings find candidates, exact rules confirm them. Vectors from two different models are not comparable either, so changing model means encoding the whole index again.
Embeddings are not anonymised data. A 2023 study by Morris and colleagues recovered 92% of 32-token text inputs exactly from their embeddings, and full names from clinical notes. Vectors computed from invoices or payroll files deserve the same protection as the files, a point for GDPR compliance in AI.
A vector database stores and searches embeddings, semantic matching uses them to pair records, and retrieval-augmented generation uses them to find the passages a model should read. An embedding model can itself be adapted through fine-tuning.
Phacet's AI Match combines exact and semantic matching, with vectorization at line level, confidence scores and plain-language justifications. The same engine runs 3-way matching, bank reconciliation and French-style account matching.