Typesense

Implements fuzzy text search using Adaptive Radix Tree (ART) data structure for memory-efficient prefix and fuzzy matching, enabling instant search-as-you-type with automatic handling of typographical errors. The ART index maintains a compressed trie structure that supports both exact and approximate string matching through edit-distance calculations, allowing users to find results even with misspellings without explicit configuration.

Supports semantic search by indexing and querying dense vector embeddings alongside traditional text indexes. Documents can include vector fields (e.g., from embedding models like OpenAI, Sentence Transformers), and queries can specify a vector to find semantically similar documents using distance metrics. The vector search integrates with the same filtering and faceting pipeline as text search, enabling hybrid queries that combine semantic relevance with structured filters.

Enables result ranking and sorting by combining text relevance scores with custom field values. Results can be sorted by any indexed field (numeric, text, or date) in ascending/descending order, or by relevance (BM25-like scoring on text fields). Multi-field sorting is supported, allowing complex ranking strategies (e.g., 'sort by relevance, then by rating, then by date'). Sorting is applied after filtering but before pagination.

Supports pagination through offset and limit parameters, allowing clients to retrieve result sets in chunks. The page parameter is a convenience wrapper around offset (page N = offset N*limit). Results are returned with metadata including total hit count, search time, and facet information. Pagination is applied after filtering and sorting, enabling efficient result navigation without re-executing the full query.

Enables multi-dimensional filtering through faceted search, allowing queries to specify boolean conditions across multiple fields (AND, OR, NOT operators) and retrieve aggregation counts for each facet value. The filtering layer operates on top of the inverted index and numeric indexes, composing posting lists to efficiently narrow result sets before ranking. Facet counts are computed during query execution, reflecting the current filtered result set.

Indexes numeric fields (integers, floats) in specialized numeric index structures enabling efficient range queries (e.g., 'price between 100 and 500') and geo-spatial queries (latitude/longitude proximity). Numeric indexes use B-tree or similar structures for fast range lookups, while geo queries compute haversine distance to find documents within a radius. Both integrate with the filtering pipeline for combined queries.

Exposes a clean HTTP REST API for document ingestion, schema management, and search queries. Documents are indexed as JSON objects validated against a collection schema that defines field types, searchability, and faceting behavior. The API uses standard HTTP verbs (POST for indexing, GET for search) and returns JSON responses, enabling direct consumption by web applications without query language learning curve. Authentication is handled via API keys managed by AuthManager.

Maintains primary index structures (ART trees, posting lists, numeric indexes) in memory for fast query execution while persisting all data to RocksDB (embedded key-value store) for durability. The Store abstraction layer mediates between in-memory indexes and RocksDB, ensuring that all mutations are written to disk before acknowledging to clients. This architecture enables sub-50ms query latency while guaranteeing data persistence across restarts.

Provides optional distributed deployment via RaftServer, implementing master-replica replication with Raft consensus protocol for cluster coordination. A leader node accepts writes and replicates changes asynchronously to replica nodes, which serve read-only queries. Raft ensures eventual consistency and automatic leader election if the master fails. This enables high-availability deployments without requiring external consensus systems.

Optimizes for instant search-as-you-type experiences by returning results with minimal latency (<50ms target) as users type each character. The system processes prefix queries efficiently using the ART index, returning partial results that are progressively refined as more characters are typed. Results are ranked by relevance and can be sorted by custom fields, enabling responsive autocomplete and search interfaces without debouncing.

Organizes data into named collections, each with its own schema defining field types, searchability, faceting, and sorting behavior. CollectionManager coordinates access to collections, enabling multi-tenant deployments where different data types or customers have separate indexes. Schema changes are applied at collection creation time; fields cannot be added/removed after creation without re-indexing. This design enforces schema discipline and prevents accidental field indexing.

Supports both batch and real-time document indexing via REST API. Batch indexing accepts JSONL (JSON Lines) format for efficient bulk loading, while individual document updates use standard JSON POST/PUT operations. The indexing pipeline validates documents against the collection schema, updates all index structures (ART, posting lists, numeric indexes) atomically, and persists changes to RocksDB. Batch operations are optimized for throughput; real-time updates prioritize latency.