Semantic search
Docs, tickets, catalog.
What is an embedding? Vectors and semantic similarity
An embedding is a dense numeric representation (a vector) of text, an image, or an entity, placed in a space where proximity reflects semantic similarity. Popularized by Word2Vec (Mikolov et al.) and generalized by encoders like BERT, embeddings power semantic search, clustering, and RAG. They do not “understand” alone—they measure useful distances.
In one sentence
An embedding turns content into a vector to compare meaning, not just words.
Key points
- Semantically close texts have close vectors (cosine / dot product).
- Foundation of semantic search and RAG with LLMs.
- Quality depends on the encoder model and domain.
- You must store and index vectors (vector DB) to scale.
Term at a glance
- Embedding
- Vector embedding · Dense representation
- English term
- Embedding
- Domain
- Artificial intelligence
- Category
- Representation / NLP
- Level
- Intermediate
What does “embedding” mean exactly?
Word2Vec showed vectors learned from co-occurrence capture analogies. BERT and modern encoders produce contextual embeddings: the same word changes with the sentence.
In SME practice: split documents, compute a vector per chunk, index, then retrieve chunks close to a question before passing them to an LLM.
Pitfalls: similarity ≠ truth; off-topic but lexically close docs can rank if chunking is poor.
How do you implement embeddings?
- 01
Prepare the corpus
Clean, chunk, add metadata (source, date).
- 02
Choose the encoder
Multilingual model if FR/EN/ES; test on your queries.
- 03
Index
Vector database + metadata filters.
- 04
Evaluate retrieval
Precision@k on real business questions.
A concrete embedding example
An industrial distributor indexes 4,000 tech sheets. A buyer asks for a “steam-resistant food-grade gasket.” Lexical search misses synonyms; embedding search surfaces the right sheet. Sales saves minutes per call.
What are embeddings used for?
RAG
Relevant context for an LLM.
Dedup / clustering
Group similar tickets or products.
Recommendation
Similar content or items.
Benefits and limits of embeddings
- Robust to synonyms
- Foundation for RAG
- Scalable with ANN indexes
- Reusable across use cases
- Indexing and re-indexing cost
- Quality tied to model and chunking
- Similarity ≠ factual accuracy
- Sensitive data in the index must be protected
How do embeddings differ from keyword search (BM25)?
| Embedding | Lexical search (BM25) | |
|---|---|---|
| Signal | Vector semantic proximity | Term overlap |
| Synonyms | Often handled better | Weak without explicit synonyms |
| Exact match | Sometimes weaker (SKU, codes) | Excellent on identifiers |
| Practice | Often hybrid (vector + lexical) | Simple and predictable |
Why embeddings matter for Quebec SMEs
Your knowledge is buried in PDFs and tickets. Embeddings make it findable by intent, not only exact words—a prerequisite for a credible document assistant.
Frequently asked questions
Do we re-index often?
Yes when the corpus changes; automate the pipeline for new docs.
What vector dimension?
Whatever the chosen model uses; bigger is not always better for your domain.
Multilingual?
Use a multilingual encoder and test FR↔EN on real pairs.
Where to store?
Managed vector DB or pgvector; encrypt and control access like any sensitive DB.
Related terms
Sources and references
Want search that understands intent? We implement embeddings, indexing, and evaluation on your corpus.
Build semantic search