ANN with Filters

Online retailers use filtered vector search to combine semantic understanding of product descriptions with hard business constraints. A user searching for "comfortable running shoes" triggers a dense retrieval of shoe embeddings, but the results are constrained by real-time filters for in_stock = true, price <= 150, and brand IN ('Nike', 'Adidas'). This is typically implemented using pre-filtering on a bitmap index for metadata before performing the HNSW graph traversal, ensuring sub-100ms latency for millions of products.

Law firms and corporate legal teams deploy ANN with filters to search across millions of documents. A query like "termination clauses with non-compete" finds semantically similar clauses via bi-encoder embeddings. Boolean filters are applied concurrently to restrict results to documents where document_type = 'contract', effective_date > 2020-01-01, and party_name = 'VendorCorp'. This conjunctive query ensures compliance and relevance, often using post-filtering strategies when filter selectivity is low.

In drug discovery, researchers search massive libraries of molecular compounds encoded as graph neural network embeddings. A filtered ANN search finds compounds with similar binding affinities but is constrained by crucial biochemical properties: molecular_weight < 500, logP < 5, and synthetic_accessibility = 'high'. This multi-stage retrieval pipeline uses pre-filtering with a Bloom filter for fast exclusion of toxic or non-synthesizable compounds before the expensive similarity search.

Banks perform real-time similarity search on transaction embeddings to identify fraudulent patterns. A new transaction is compared to a database of known fraud vectors. Metadata filtering is critical: transaction_amount > 10000, merchant_category = 'high-risk', and user_account_age < 30 days. The system uses filter pushdown to evaluate these predicates at the storage layer, drastically reducing the candidate set for the subsequent approximate nearest neighbor search and enabling real-time alerting.

SaaS applications serving thousands of customers use ANN with filters for secure, isolated search. All user queries are automatically scoped by a mandatory tenant_id = 'xyz' filter. This is implemented via query rewriting at the API layer, ensuring hard isolation. The vector database's index, often a filtered HNSW variant, is optimized for this high-selectivity filter, allowing efficient traversal of only the graph nodes belonging to the specified tenant's data partition.

A modification of the Hierarchical Navigable Small World (HNSW) graph index that incorporates metadata filtering logic directly into the graph traversal algorithm. Instead of searching the entire vector space, the algorithm prunes paths that lead to nodes violating the filter constraints, enabling efficient filtered similarity search without a full pre-filter scan.

• Core Mechanism: During the greedy graph traversal, candidate nodes are checked against the provided Boolean filter. Only nodes that pass the filter are added to the candidate set for further exploration.
• Performance Benefit: This integrated approach avoids the performance cliff of naive post-filtering, where a broad vector search returns many irrelevant results that are subsequently discarded.

A critical database query optimization technique where filtering predicates are evaluated as early as possible in the query execution pipeline, ideally within the storage engine. For ANN with filters, this means applying metadata constraints before or during the vector index traversal to minimize data movement and computational overhead.

• Analogy: Similar to a SQL query using a WHERE clause on an indexed column; the database uses the index to skip irrelevant rows entirely.
• Implementation: In vector databases, this often involves integrating filter logic into the ANN index (like HNSW or IVF) or using auxiliary bitmap indexes to quickly identify eligible vector IDs before a similarity search begins.

The two fundamental strategies for applying metadata constraints in a vector search pipeline.

• Pre-Filtering: Metadata filters are applied first to create a reduced candidate set of document IDs. The expensive vector similarity search is then performed only on the embeddings corresponding to these IDs. Risk: overly restrictive filters may eliminate all true nearest neighbors, causing low recall.
• Post-Filtering: A broad ANN search is executed first, returning a set of candidate vectors. Metadata filters are then applied to this set, discarding non-compliant results. Risk: if the filter is highly selective, most of the initial vector search work is wasted, leading to high latency and empty results.

Modern systems use filter-aware indexes (like HNSW with filters) to blend these approaches optimally.

A specialized, low-level database index structure that enables extremely fast set operations for filtering. It represents the membership of records for specific attribute values using arrays of bits (bitmaps).

• How it works: For each unique value in a metadata field (e.g., category=news), a bitmap is stored where each bit corresponds to a row ID; a 1 indicates the row has that value.
• Use in ANN with Filters: To apply a filter like category IN ('news', 'sports'), the database performs a bitwise OR on the corresponding bitmaps, instantly producing a bitmap of all eligible row IDs. This bitmap can be used for efficient pre-filtering before vector search.
• Advantage: Provides O(1) complexity for set union/intersection operations, making it ideal for high-cardinality, concurrent filtering.

The process by which a vector database's optimizer analyzes an incoming query—including its vector, metadata filters, and Boolean logic—to generate the most efficient sequence of operations, known as an execution plan. This is crucial for ANN with filters to avoid performance pitfalls.

• Key Decision: The planner estimates the selectivity of each filter predicate (the fraction of total records it will match). A highly selective filter (e.g., user_id=123) suggests a pre-filtering strategy, while a broad filter (e.g., lang='en') may be better applied via filter-aware index traversal.
• Output: The plan dictates the order of operations: whether to use a bitmap index first, which ANN algorithm to invoke, and how to merge intermediate results.

A quantitative measure, expressed as a fraction or percentage, that estimates the proportion of records in a dataset that will satisfy a given metadata filter predicate. It is the primary heuristic used by a query planner to choose between pre-filtering and post-filtering strategies.

• High Selectivity (< 1% of records match): Favors a pre-filtering approach, as scanning a small candidate set for vectors is efficient. Example: customer_id='a1b2c3'.
• Low Selectivity (> 50% of records match): Favors integrating the filter into the ANN index traversal or using post-filtering, as pre-filtering would still leave a very large set to search. Example: created_year > 2020.
• Calculation: Selectivity is often estimated via pre-computed statistics (histograms) on metadata fields.