What is semantic search?
What is it?
Semantic search finds content by meaning, not only by exact keyword matches. It understands that different words can express the same idea.
Traditional keyword search: query must contain the same words as the document.
Semantic search: query and document can use different words if their meanings align.
Why does it exist?
Users ask questions naturally. They may not know the exact terminology in your manuals or policies. Semantic search improves recall — finding relevant content even when vocabulary differs.
How does it work conceptually?
Semantic search in RAG typically uses embeddings and vector similarity:
Indexed chunks → chunk embeddings stored in vector database
User query → query embedding
Compare → similarity search (nearest vectors)
Results → most semantically similar chunksExample
Manual text: "Perform a full service inspection at 500 operating hours."
| Query | Keyword search | Semantic search |
|---|---|---|
| "500 hour maintenance" | May match (shared words) | Matches |
| "half thousand hour service" | Likely no match | Likely matches |
| "when to service the machine" | Likely no match | Likely matches |
Semantic search vs keyword search
Many production systems use hybrid search — combining keyword and semantic signals. Noulek's MVP retrieval is embedding-based semantic search through the vector database, not a separate keyword index exposed in the admin UI.
Future concepts in this path will cover hybrid approaches and reranking.
Where does it fit in a RAG pipeline?
Semantic search is the mechanism inside retrieval:
Query → Embed query → Vector similarity search → Ranked chunksIt powers both Noulek Search (display results) and Chat (select context).
How this relates to Noulek
Noulek Search uses semantic retrieval over indexed Knowledge Base chunks:
- Documents must be Ready in Knowledge Base.
- Embeddings are created with OpenAI models during indexing.
- Similarity search runs against your Weaviate collection.
If Search returns nothing useful, the issue may be missing content, provider misconfiguration, or a query with no semantically close indexed material — not necessarily "broken" search.
Try Search the Knowledge Base for the product workflow, and Search or Chat no results for diagnosis.