Embeddings · Vector DB
Embedding / Vector Database
What are embeddings?
- Turning meaning into numeric vectors (embeddings)
- A vector DB for fast similarity search
- A core component of RAG
What embeddings change is not the search results screen but the kind of work the search owner does. With traditional keyword search, a document was found only if the word 'annual leave' was literally in it, so owners had to hand-build synonym rules like 'vacation = annual leave = paid leave' over years. Now employees can ask in their own words and still get a match. The trade-off is that it becomes harder to explain why a document ranked first, because the ranking rests not on rules a person wrote but on distances between coordinates.
To get a feel for the numbers: embedding dimensions are typically 768 to 3,072, and documents are cut into 300–800 character chunks with 10–15% overlap at each edge so that sentences aren't cut off mid-meaning. In practice, vector search is rarely used alone; hybrid search, which runs keyword search (BM25) alongside and merges the results, is close to the default. That's because vectors are actually weaker on values that must match exactly, such as product codes, contract numbers, and clause numbers. A common setup adds metadata filters like department and year, then narrows the top 20 hits to 5 with a reranker before passing them to the AI.
The most common misconception is that 'installing a vector DB makes search smart.' A vector DB is built to always return the closest match even when nothing is relevant, so a similarity score like 0.82 does not guarantee a correct answer. Without a cutoff, irrelevant chunks get attached as evidence and produce plausible wrong answers. Also, swapping the embedding model puts new vectors in a different coordinate space from the existing ones, so everything must be re-indexed. Teams that change only the model often see search quality collapse. Failing to delete vectors of revised or retired documents causes similar problems. A policy retired last year gets cited as evidence in today's answers, and without permission filters, chunks from documents the user is not allowed to view end up in answers.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayEmbedding, which turns words into vectors that capture meaning, was popularized by word2vec, which Google released in 2013. It later evolved into sentence embeddings and vector databases and became the backbone of RAG.
Consider a hypothetical case. Company A makes industrial valves, with 420 employees and annual revenue of KRW 98 billion (≈ US$73,320,400). To answer client inquiries, its technical sales team had to dig through 3,800 PDFs of specifications and installation case studies spanning 15 years, averaging 40 minutes per inquiry. Search only worked if you knew the exact product code, so new hires simply asked senior colleagues instead.
- When a senior manager on the technical sales team pulled 200 failed-search logs from the past six months, 78% were conversational questions like 'What gasket do we use for high-temperature lines?', and most of them returned zero results.
- A manager on the information systems team split the 3,800 PDFs into 600-character chunks with 100-character overlap, embedded roughly 42,000 chunks, and loaded them into a vector DB.
- In the first test, a catalog discontinued in 2016 ranked near the top, and the head of sales pointed out, "We don't even sell that product," so they added metadata filters for document year and product line.
- Searches for codes like 'V-40S' kept ranking low, so the team switched to a hybrid approach that runs keyword search alongside vector search, then passed the top 20 results through a reranker to narrow them to 5 as the evidence for answers.
- After assigning someone to re-index regulation updates once a week, by month three the time spent finding materials per inquiry fell from 40 minutes to 7 minutes, and internal requests asking seniors for help dropped from 60 to 18 a month.