Vector Similarity Search

The foundational algorithm that trades a small amount of perfect accuracy for massive gains in query speed. Instead of comparing a query vector against every vector in the database—an O(N) brute-force scan—ANN algorithms use graph-based (HNSW) or clustering (IVF) indexes to quickly navigate to the most promising region of the vector space.

• HNSW (Hierarchical Navigable Small World): Builds a multi-layered graph where long-range links enable logarithmic scaling.
• IVF (Inverted File Index): Partitions the space into Voronoi cells, searching only the nearest clusters.
• Recall vs. Latency Trade-off: Tuning the ef_search or nprobe parameters directly balances result accuracy against query time.

Specialized data structures are required to overcome the curse of dimensionality, where traditional spatial indexes like k-d trees degrade to brute-force performance. Modern vector databases use purpose-built indexes.

• Quantization (PQ/SQ): Compresses high-dimensional vectors into compact codes, reducing memory footprint and accelerating distance calculations at the cost of minor precision loss.
• Disk-based ANN (DiskANN): Enables searching billion-scale datasets on a single machine by keeping the graph structure on fast NVMe SSDs and only the compressed vectors in RAM.
• Filtered Search: Combining vector similarity with structured metadata filters (e.g., 'color=red') requires pre-filtering or post-filtering strategies to maintain semantic relevance.

The effectiveness of a similarity search is entirely dependent on the quality and dimensionality of the upstream embedding model. The vector database must support the output dimensions of the model without re-indexing.

• Dimensionality: Modern models produce embeddings from 384 (MiniLM) to 4096 (Ada-002) dimensions. Higher dimensions capture more nuance but increase storage and compute costs.
• Multi-Modal Alignment: Advanced systems use CLIP-like models to embed text and images into a shared vector space, enabling cross-modal search (e.g., 'find a dress that looks like this photo').
• Fine-tuning: Domain-specific fine-tuning of the embedding model is critical for retail, where 'apple' must be disambiguated between a fruit and a brand.

For dynamic retail personalization, the vector index cannot be a static artifact. It must support online updates to reflect new inventory, trending products, and shifting user intent without requiring a full rebuild.

• Insert-Only Workloads: Adding new product embeddings as inventory changes must not degrade query performance.
• CRUD Operations: Updating or deleting vectors (e.g., for discontinued items) requires a log-structured merge-tree (LSM) or similar mechanism to avoid index fragmentation.
• Freshness: The delay between a product being added to the catalog and becoming searchable in the vector index should be measured in seconds, not hours.

Pure vector search can miss exact keyword matches that are critical for precision. Hybrid search combines the semantic understanding of vector search with the literal precision of sparse lexical search (BM25/SPLADE).

• Reciprocal Rank Fusion (RRF): A robust algorithm that merges result lists from vector and keyword indexes without requiring score normalization.
• Query Intent Classification: A lightweight model can route queries—sending 'Nike Air Max size 10' to keyword search and 'comfortable running shoes' to vector search.
• Metadata Boosting: Results can be re-ranked by combining the semantic similarity score with business rules, such as boosting items with high margin or current stock levels.

The most common distance metric in vector similarity search, measuring the cosine of the angle between two vectors rather than their absolute magnitude. A score of 1.0 indicates identical direction (perfect semantic match), 0 indicates orthogonality (no relationship), and -1 indicates opposite meaning. Unlike Euclidean distance, cosine similarity is magnitude-invariant, making it ideal for comparing text embeddings where document length should not influence relevance. In retail personalization, a user's session embedding is compared against product catalog embeddings using cosine similarity to rank the most semantically relevant items.

A class of algorithms that trade a small amount of accuracy for dramatic speed improvements when searching billion-scale vector datasets. Exact nearest neighbor search scales linearly with dataset size, becoming prohibitively slow for real-time personalization. ANN algorithms like HNSW (Hierarchical Navigable Small World) and IVF (Inverted File Index) use graph-based or clustering techniques to prune the search space, achieving sub-millisecond latency. In e-commerce, ANN enables a product catalog of 100 million SKUs to return semantically similar items in under 10ms.

A neural network that transforms unstructured data—text, images, or user behavior—into dense, fixed-length vector representations where semantic similarity is encoded as spatial proximity. Modern embedding models like text-embedding-3-large or Cohere Embed v3 produce vectors of 768 to 3072 dimensions. In retail, a product description embedding model encodes both a search query ('lightweight running shoes') and product titles into the same vector space, enabling semantic matching even when exact keywords differ. These models are typically fine-tuned on domain-specific data to capture retail-specific relationships like brand affinity or style compatibility.

A purpose-built database system optimized for storing, indexing, and querying high-dimensional vector embeddings at scale. Unlike traditional databases that index scalar values, vector databases use specialized index structures like HNSW graphs or IVF partitions to enable fast ANN search. Leading systems include Pinecone, Weaviate, Milvus, and Qdrant. Key capabilities include metadata filtering (e.g., 'only in-stock items'), hybrid search combining keyword and vector scores, and real-time index updates. For real-time customer segmentation, the vector database stores user session embeddings and retrieves the top-K most similar historical segments in milliseconds.

The foundational retrieval pattern where a query vector is compared against all vectors in a dataset to return the k most similar items. While exact k-NN is computationally expensive at scale, it serves as the conceptual basis for ANN approximations. In real-time personalization, a k-NN query retrieves the top-50 products most semantically similar to a user's current browsing session embedding. The parameter k balances diversity and relevance: a small k (e.g., 5) returns highly focused recommendations, while a larger k (e.g., 100) provides a broader candidate pool for downstream re-ranking by a business rules engine or CTR prediction model.

A search paradigm that understands the intent and contextual meaning of a query rather than matching exact keywords. Unlike lexical search (e.g., BM25), semantic search uses embedding models to map both queries and documents into a shared vector space where conceptually related items cluster together. A query for 'warm jacket for winter hiking' can surface products described as 'insulated parka' or 'thermal outerwear' without keyword overlap. In retail, semantic search dramatically improves recall for long-tail queries and reduces null results, directly impacting conversion rates. Modern implementations often combine semantic and lexical signals in a hybrid scoring function.