Find answers by meaning, not by the words you typed.
Semantic search uses transformer embeddings to retrieve passages that address your question—regardless of whether they share a single keyword with it. CogniWeave runs hybrid retrieval over your private corpus and returns cited answers.
Definition
Semantic search is retrieval based on the meaning of text, not its surface form. It maps queries and documents to a shared vector space where related concepts sit near each other.
How it works
- • Encode each document chunk into a dense embedding at ingest time.
- • Encode the user's query at request time with the same model.
- • Find the nearest neighbours of the query vector via an ANN index.
- • Optionally rerank with a cross-encoder for higher precision.
When to use it
Use semantic search when users ask questions in natural language, when synonyms and paraphrasing matter, when content spans multiple domains or languages, or when you need to feed an LLM grounded context for RAG.
Trade-offs vs. keyword search
Keyword search is cheaper, fully interpretable, and unbeatable for exact identifiers like SKUs or error codes. Semantic search costs more compute per query and can hallucinate near-misses on rare entities. Hybrid retrieval generally beats either alone.
The CogniWeave advantage
CogniWeave runs hybrid retrieval (BM25 + vector) over your private corpus and returns citations linked to the original passage. Tenant isolation, audit logs, and your-data-stays-yours posture are on by default.
Example queries
- • ‘How do we handle late shipments?’ → matches a policy passage that never uses the word ‘late’.
- • ‘Ways to cool a room without AC’ → returns content about fans, ventilation, and shading.
- • ‘Who signed the MSA with Acme?’ → returns the signature page of the relevant PDF.
Common questions
What is semantic search?
Semantic search retrieves results by meaning rather than literal keyword overlap. Queries and documents are encoded into vectors, and the system returns the documents whose vectors are closest to the query—even when they share no words.
How is semantic search different from keyword search?
Keyword search (BM25, TF-IDF, Elasticsearch defaults) matches tokens. Semantic search matches intent. Keyword search wins for exact identifiers and rare names; semantic search wins for natural-language questions and concept overlap.
Does semantic search replace keyword search?
Not usually. The best retrievers fuse both signals—a BM25 candidate set reranked by vector similarity, or vice versa—because each compensates for the other's failure mode.
How accurate is CogniWeave's semantic search?
CogniWeave uses retrieval-tuned managed embeddings indexed with HNSW. Recall and latency depend on corpus size and chunking, but typical deployments serve top-10 results in under 200ms with grounded citations attached.