Vector Databases
Specialized databases optimized for storing, indexing, and retrieving high-dimensional vector embeddings. Enable efficient similarity search (e.g., nearest neighbor queries) for applications like semantic search, recommendation systems, and LLM-powered retrieval.
Core Functionality
- Approximate Nearest Neighbor (ANN) Search: Uses algorithms like HNSW, IVF, or FAISS for scalable similarity matching.
- Embedding Support: Stores vectors generated from text, images, or other modalities via models like BERT, CLIP, or Sentence Transformers.
- Scalability: Handles millions/billions of vectors with low-latency queries.
Key Limitations
- Model Constraint: Requires the same embedding model for both vector generation and retrieval (e.g., BERT embeddings must be searched with BERT).
- Semantic Rigidity: Struggles with complex relationships beyond vector similarity (e.g., hierarchical or causal links).
GraphRAG: Flexible Alternative
- GraphRAG leverages knowledge graphs and LLMs to query structured data, eliminating the model constraint of vector search.
- Flexibility: Uses different models for graph construction (e.g., GNNs) and retrieval (e.g., LLMs), enabling richer context.
- Advantage: Better handles complex queries involving relationships (e.g., “Show me products similar to X that are also used with Y”) compared to pure vector similarity.
Integration Ecosystem
- RAG Pipelines: Primary component in Retrieval-Augmented Generation for LLM context retrieval.
- Knowledge Graphs: Knowledge Graphs enhance vector search with structured relationships (e.g., Neo4j + vector DBs).
- Embedding Models: embedding-models must align with vector DB capabilities (e.g., dimensionality, metric).
References
- 2026 04 14 IBM Explainer creating GraphRAG
Source Notes
- 2026-04-23: Claude · ▶ source
- 2026-04-10: Structured AI Context Beyond RAG Limitations with Map First Architectu · ▶ source
- 2026-04-20: Larql Querying and Modifying LLM Internal Database Structures · ▶ source