Semantic Search

General definition

Semantic search is a retrieval approach that represents both queries and documents as embeddings, then finds the documents whose vectors are closest to the query vector. Because the embedding space is learned from vast amounts of language, it captures synonyms, paraphrases and conceptual relationships that keyword search would miss entirely.

• Dense retrieval - compare query and document embeddings using cosine similarity or dot product
• Hybrid search - combine a dense semantic score with a sparse keyword score (BM25) to get the best of both approaches
• Re-ranking - a second pass with a cross-encoder model to re-order the top candidates from the initial retrieval for higher precision
• Chunking strategy - how documents are split before embedding affects retrieval quality significantly

Semantic search is a core component of RAG pipelines, enterprise knowledge bases, and AI-powered customer support. It is also used in product search, content recommendation, and duplicate detection. The retrieved chunks are typically passed to an LLM alongside the original query to produce a grounded, accurate answer rather than a bare list of links.

In the Ethora ecosystem

The Ethora AI SDK gives you semantic search over your own content as a first-class feature. Index your product documentation, support articles or internal knowledge base, and the built-in RAG bot will retrieve the most relevant chunks before generating a response, so answers are grounded in your actual content rather than the model’s general knowledge.

With a self-hosted LLM agent, the embedding model and the vector store both run on your infrastructure. That means sensitive documents, patient records or proprietary content are indexed and searched locally, with no data sent to an external embedding API. The same OpenAI-compatible interface works whether you point to a cloud provider or a local Ollama or vLLM instance.