Modern Search Explained: FTS, Embeddings, RAG and More

Introduction

Modern AI systems such as chatbots, search engines and RAG applications are built on a combination of several technologies. Terms like FTS, embeddings and re-ranking come up again and again.

This article explains the most important fundamentals – technically accurate, but easy to follow.

1. FTS (Full-Text Search)

Full-text search is the classic way of searching text content. It is based on:

• Tokenization (breaking text apart)

• Stemming (identifying word stems)

• Stopwords (ignoring irrelevant words)

FTS is particularly good for:

• exact terms

• names, IDs, technical vocabulary

2. BM25

BM25 is a ranking algorithm for FTS.

It scores documents based on:

• how often a term appears

• the length of the document

• how rare a term is

BM25 decides which document is „most relevant“.

3. Embeddings / Vector Search

Embeddings turn text into numeric vectors:

"cancel contract" → [0.23, -0.88, ...]

That makes semantic search possible:

• similar meaning is recognized

• not just exact words

4. Hybrid Search

Hybrid search combines:

FTS + embeddings

Why?

• FTS → precise words

• embeddings → meaning

Together they deliver considerably better results.

5. Top-K Retrieval

Top-K describes the selection of the best results:

Top 5, Top 10, Top 20

These are then handed to the LLM.

6. Re-Ranking

Re-ranking is a second step after the search:

Top 50 results → reorder → Top 5

Often using:

• LLMs

• cross-encoders

• specialized models

→ a large gain in quality

7. Chunking

Documents are split into smaller pieces:

• better search

• more precise answers

What matters:

• not too large

• not too small

• semantically coherent

8. HNSW (ANN)

HNSW is an algorithm for fast vector search.

It belongs to:

• Approximate Nearest Neighbor (ANN)

Advantages:

• very fast

• scales well

9. RAG (Retrieval-Augmented Generation)

RAG combines search with AI:

search → context → LLM → answer

Advantages:

• current data

• your own documents

• fewer hallucinations

How the Components Work Together

Documents
→ Chunking
→ Embeddings
→ Storage (PostgreSQL + pgvector, for example)
→ Hybrid Search (FTS + vector)
→ Top-K Retrieval
→ Re-Ranking
→ LLM (RAG)
→ Answer

Conclusion

Modern search is not a single feature – it is a system made of several components:

• FTS & BM25 → words

• embeddings → meaning

• hybrid search → the combination

• re-ranking → quality

• RAG → the AI application

The quality of your AI depends directly on your retrieval stack.