Typesense

Implements fuzzy matching and typo tolerance using an Adaptive Radix Tree (ART) data structure that enables memory-efficient prefix and fuzzy matching across indexed text fields. The ART index is maintained in-memory for fast reads while persisted to RocksDB for durability, allowing sub-50ms query latency even with spelling variations. Queries automatically expand to include typo variants without requiring explicit configuration.

Supports dense vector search by storing and indexing embedding vectors alongside document fields, enabling semantic similarity queries beyond keyword matching. Integrates with ONNX Runtime for optional on-device embedding generation, allowing documents and queries to be embedded without external API calls. Vector search results can be combined with keyword filters and facets in a single query.

Supports geopoint fields for storing latitude/longitude coordinates and enables distance-based filtering (e.g., find results within 10km of a location) and polygon-based filtering (e.g., find results within a geographic boundary). Geospatial queries are evaluated during search using spatial indexing, and results can be sorted by distance. Integrates with standard GeoJSON formats.

Enables sorting search results by one or more fields (text, numeric, date) in ascending or descending order, with support for relevance-based ranking (BM25 or vector similarity scores). Sorting is applied after filtering and faceting, and results are paginated using offset/limit parameters. Multi-field sorting allows complex ranking strategies (e.g., sort by relevance, then by date, then by rating).

Supports bulk indexing of multiple documents in a single API request, reducing HTTP overhead and improving throughput for large-scale data imports. Bulk operations are processed in batches and persisted to RocksDB atomically, ensuring consistency. Supports both insert and update operations in a single batch request.

Tracks search queries, user interactions, and system events through an Analytics component, enabling real-time insights into search behavior and system performance. Events are collected asynchronously and can be exported for analysis. Supports custom event tracking for application-specific metrics.

Enables drill-down filtering across multiple document fields with automatic aggregation of result counts per facet value. Facets are computed during search by maintaining inverted indexes per field, allowing fast computation of value distributions without post-processing. Supports hierarchical faceting and numeric range facets alongside categorical facets.

Enforces explicit schema definition for collections, where each field specifies type (string, int, float, bool, geopoint, object), indexing behavior (indexed, sortable, facetable), and optional parameters like tokenization strategy. Documents are validated against schema at index time, and fields are indexed according to their configuration using specialized index structures (ART for strings, NumericTrie for ranges, etc.). Schema changes require explicit migration.

Implements NumericTrie data structure for efficient range queries on numeric fields (integers, floats), enabling fast filtering by numeric ranges without scanning all documents. Range queries are evaluated during search using trie traversal, and results can be combined with text search and faceting. Supports both inclusive and exclusive range boundaries.

Exposes all search, indexing, and management operations through a clean RESTful API built on H2O HTTP server, accepting JSON payloads and returning JSON responses. API endpoints follow REST conventions (GET for retrieval, POST for creation, PUT for updates, DELETE for removal) and support query parameters for filtering, pagination, and sorting. Authentication is enforced via API keys managed by AuthManager.

Manages API key generation, validation, and scoping through AuthManager component. Each API key can be scoped to specific collections and operations (search, index, delete), enforced at request time before query execution. Keys are validated on every HTTP request via Authorization header, and invalid/expired keys are rejected with 401 responses.

Maintains primary index structures (ART trees, NumericTries, inverted indexes) in memory for fast query execution, while persisting all data to RocksDB key-value store for durability and recovery. The Store abstraction layer handles all persistence operations, decoupling in-memory indexes from disk storage. On server restart, indexes are reconstructed from RocksDB snapshots.

Implements optional Raft consensus for high-availability deployments, allowing multiple Typesense nodes to replicate data and elect a leader. Write operations are committed to the Raft log before being applied to local indexes, ensuring consistency across replicas. Read operations can be served from any replica, enabling load distribution. Raft handles leader election and automatic failover.

Supports conversational search workflows where user queries are processed through LLM-based intent understanding, and retrieved documents are fed back to LLMs for answer generation. Integrates with external LLM APIs (OpenAI, Anthropic) for query expansion and answer synthesis. Retrieved documents are ranked by relevance and passed to LLMs with context windows, enabling RAG (Retrieval-Augmented Generation) pipelines.