Glossary · AI & automation
Vector embeddings meaning
Vector embeddings are lists of numbers that capture the meaning of text or records, so similar items sit close together. How they power semantic search and RAG.
How do vector embeddings work?
An embedding model reads a chunk of text and outputs a fixed-length list of numbers, often hundreds or a few thousand long. You cannot read the numbers individually, but their position matters: "salary credit" and "monthly pay received" land near each other, while "salary credit" and "cheque bounce" land further apart. To find related content, the system turns the question into a vector the same way and measures which stored vectors are nearest, commonly with a measure called cosine similarity.
A vector database is simply a store built to hold millions of these vectors and return the nearest ones quickly.
Keyword search vs semantic search: what is the difference?
| Keyword search | Embedding (semantic) search | |
|---|---|---|
| Matches | Exact words | Similar meaning |
| Handles synonyms and Hinglish | Poorly | Better, depending on the model |
| Exact codes and numbers | Strong | Weak |
| Explainability | Easy: the word matched | Harder: the vectors were close |
Good systems use both, so that an account number or GSTIN is found exactly while a loosely worded question still finds the right clause.
A worked example in a loan file (illustrative)
An analyst asks: "Has the borrower had any EMI returns?" The bank statement never uses that phrase; it has narrations like "ACH DR RTN" and "NACH return charges". A keyword search for "EMI return" finds nothing. An embedding search finds those lines because their meaning is close, and passes them to the model, which lists the dates and amounts with page references. The narrations and the question are illustrative. Retrieval like this is one building block of credit appraisal software for NBFCs that drafts a memo from statements, GST returns and bureau data, with the credit officer still deciding.
What are the common mistakes with embeddings?
- Mixing models. Vectors from two different embedding models cannot be compared. Re-embed everything when you change models.
- Chunks too large or too small. A whole policy in one vector blurs meaning; single lines lose context.
- Relying on similarity for numbers. "Rs 50,000" and "Rs 5,00,000" can look close. Use exact lookups for figures.
- Forgetting that vectors hold data. Embeddings of personal or financial documents are still sensitive and need the same access control and data residency as the source, which is one reason lenders weigh on-premise vs cloud AI.
- Embeddings vs knowledge graph. Embeddings capture similarity; a knowledge graph stores explicit relationships. They answer different questions.
To see why structured, repeatable analysis still matters alongside search, read manual vs automated credit underwriting.
Related terms
Read next
Where this shows up in our work. AutoCredit: Statements, GST, bureau and AA into one file; a memo against your policy bands; same-day field verification. Book a 30-minute working session with an engineer.
Questions
What is a vector embedding in simple words?
It is a list of numbers that an AI model produces to represent the meaning of a text, image or record, arranged so that things with similar meaning get similar numbers and can be found together.
What is a vector database used for?
A vector database stores embeddings and quickly finds the ones closest to a query. It is commonly used for semantic search and as the retrieval step in retrieval-augmented generation.
Are embeddings the same as a large language model?
No. An embedding model turns content into vectors for search and comparison. A large language model generates text. RAG systems typically use both: embeddings to find passages and an LLM to write the answer.