Search by meaning: embeddings and RAG

An embedding is usually several hundred to a few thousand numbers. A simple example uses just 2. Take the arrows (0.8, 0.6) and (0.6, 0.8). Multiply them number by number, then add: 0.8×0.6 + 0.6×0.8 = 0.96. Then divide by the two arrow lengths. Here both lengths are 1. So the score is 0.96. The two texts are very alike.

Think of a supermarket. Pasta sits near the sauces, far from the shampoo. Embeddings give each text a place on that kind of floor plan.

• Bad chunks. A table is split in half. Or a chunk says "this rises to 20%" but not what rises.
• Exact words. Embeddings catch the general meaning. An error code, a part number or a name can match the wrong page.
• One search only. Some questions need 2 searches. The second search needs the answer to the first.
• Who is asking. Meaning search does not check who asks. See permission-aware RAG.
• Hidden orders. Found text can carry prompt injection.

Priya's bot said the Pune hotel limit was ₹5,000 a night. It cited a real source. But that was the old travel policy. The new one scored a bit lower. The fix was not a smarter AI. She deleted the old file. She added "valid from" dates the AI could see.

Hybrid search. A keyword search, like the classic BM25, runs next to the meaning search. The two result lists are merged. Then a re-ranker model often puts the merged list in a better order. This helps when both meaning and exact words matter.

Agentic search. The AI gets search tools. It decides what to look up, reads, and looks again. This helps with questions that need several searches. It also helps with exact text, like code, logs and IDs. It costs more calls, tokens and time.

Other options. A small set of files that rarely changes can go straight into the prompt. A knowledge graph links people, teams and products, so questions can follow the links. Classic RAG still works well for simple look-ups in large sets of text. It needs no retraining (see Prompt, RAG or fine-tune?).

A list of numbers is also called a vector. A vector database stores them next to the text. It quickly finds the ones nearest to a question. With millions of chunks, it uses a fast shortcut that can miss a match now and then. As of October 2026, pgvector, Qdrant and Chroma are open source. Weaviate has an open-source core. Pinecone is a closed, hosted service.

Each embedding model makes its own kind of numbers. Numbers from 2 models can't be compared. So if you switch models, you must embed everything again.