A vector database for meaning, not keywords.
Vector databases store embeddings—dense numerical fingerprints of meaning—so retrieval works across paraphrasing, synonyms, and intent. CogniWeave wraps a production-grade vector store in a full ingestion and retrieval pipeline.
Definition
A vector database is a specialised index for similarity search over dense vector embeddings. Each item—a sentence, a paragraph, an image—is mapped to a point in a high-dimensional space where distance corresponds to meaning.
How it works
- • Ingest: documents are parsed and split into semantically coherent chunks.
- • Embed: each chunk is encoded by a transformer model into a dense vector.
- • Index: vectors are stored in an ANN structure such as HNSW or IVF-PQ.
- • Query: incoming questions are embedded, then nearest neighbours are returned.
When to use one
Reach for a vector database whenever retrieval needs to generalise across vocabulary—RAG pipelines for LLMs, semantic search over a knowledge base, recommendations, clustering, or de-duplication.
Trade-offs vs. alternatives
Keyword indexes are faster for exact matches and cheaper to operate, but blind to paraphrasing. Pure vector retrieval can drift on rare entities or precise identifiers. The strongest systems combine both and rerank the union.
The CogniWeave advantage
CogniWeave hides the operational surface of a vector database. You upload files; we handle parsing, OCR, chunking, embedding, indexing, retrieval, and citation. Tenant isolation is per-workspace and answers carry source lineage by default.
Specs at a glance
- • Embedding dimension: 1536
- • Index: HNSW with cosine similarity
- • Target retrieval latency: under 200ms
- • Tenant isolation: per-workspace
- • Storage: AES-256 at rest, TLS 1.2+ in transit
Common questions
What is a vector database?
A vector database stores high-dimensional embeddings—numerical representations of meaning produced by transformer models. Instead of matching keywords, it finds entries that are geometrically close to a query vector, which corresponds to semantic similarity.
When should I use a vector database instead of a relational store or keyword search engine?
Use a vector database when retrieval must work across paraphrasing, synonyms, or multilingual content; when results need to feed an LLM as grounded context; or when keyword search returns too much noise. Hybrid setups that combine BM25 with vector recall usually outperform either alone.
How does CogniWeave compare to Pinecone, Weaviate, or Qdrant?
Pinecone, Weaviate, and Qdrant are general-purpose vector stores—you bring your own ingestion, chunking, and retrieval logic. CogniWeave ships the full pipeline (parse, OCR, chunk, embed, store, retrieve, cite) on top of its own memory engine, so you get cited answers out of the box.
What embedding model does CogniWeave use?
CogniWeave defaults to managed embeddings tuned for retrieval. Each chunk is encoded into a high-dimensional vector and indexed with HNSW for sub-200ms nearest-neighbour search.