Features

Everything Torque offers - from search, AI, and GPU acceleration to ingestion, vector search, and developer tools.

Search Engine

BM25F Text Search

Field-weighted BM25F scoring across multiple fields. Configure query_by, query_by_weights, and search across any combination of text fields in your schema.

Hybrid Search

Combine BM25F text ranking with HNSW vector search using reciprocal rank fusion. Get keyword precision and semantic recall in one query.

Vector Search

HNSW graph index with RaBitQ binary quantization for high-accuracy approximate nearest neighbors. Supports cosine, dot product, and Euclidean distance metrics.

Prefix Search

FST-backed prefix index for instant type-ahead suggestions. Results appear as users type, with scored ranking across configured fields.

Typo Tolerance

SymSpell fuzzy correction with configurable edit distance. Handles misspellings automatically so users find results even with typing errors.

Hit Highlighting

Matched terms are highlighted in search results with configurable start and end tags. See exactly which words matched and where.

AI-Powered Search

Conversational Search

Let users ask questions and get answers grounded in your data. Torque retrieves the most relevant documents and generates a streamed answer, remembering context across follow-up questions. Works with any OpenAI-compatible model - hosted or self-hosted.

Natural-Language Queries

Turn "cheap laptops under 500 dollars sorted by rating" into a precise, filtered, sorted search - automatically. An LLM translates plain language into structured queries using your actual schema and data values, so filters always hit real values.

MCP Server for AI Agents

Connect Claude and other AI assistants directly to your search data. The built-in Model Context Protocol server lets agents search, explore schemas, and manage documents - with the same API key security as everything else.

Automatic Embeddings

Semantic search without an embedding pipeline. Declare which fields to embed and Torque generates vectors automatically at import time - and embeds query text at search time. Zero client-side model code.

Curation

Put the right result first. Pin featured documents, hide others, and schedule promotions with time windows - per query, per filter, or per campaign tag.

Search Analytics

Know what your users are looking for. Capture popular queries, spot searches that return nothing, and count clicks and conversions - all searchable like any other collection.

Filtering & Faceting

Roaring Bitmap Filtering

Compressed roaring bitmaps for all filter operations. Numeric ranges, exact matches, not-equals, and boolean AND/OR/NOT - computed in microseconds even on millions of documents.

Faceted Search

String and numeric facets with stats (min, max, sum, average). Control max facet values per field. Over 14x faster than Typesense v30.1 on real-world data.

Nested Field Filtering

Filter on nested object and array fields with element-level AND semantics. Conditions match within the same array element, ensuring correct results for structured data with multi-field arrays.

GPU Acceleration

CUDA Compute Kernels

Seven dedicated CUDA kernels accelerate bitmap operations, BM25F scoring, range filtering, geo radius search, vector distance, and facet counting on NVIDIA GPUs.

cuBLAS Vector Search

GPU-accelerated vector distance calculations via cuBLAS for batch dot products. Works with any CUDA-capable NVIDIA GPU - from RTX consumer cards to A100 data center and Blackwell workstation hardware.

Automatic CPU Fallback

Full CPU-only mode when GPU hardware is unavailable. Same API, same results, same correctness - the GPU just makes it faster. No code changes needed.

Ingestion & Data

Binary TCP Protocol

Custom binary protocol for streaming documents over persistent TCP connections. 7.8x faster than Typesense ingestion on the same hardware. Supports all field types including geo points, arrays, and vectors.

TQBF Binary File Upload

Pre-encoded binary files with LZ4 or ZLib compression. A single HTTP POST imports the entire dataset with no per-document JSON parsing overhead. Ideal for ETL pipelines, data distribution, and initial loads.

Batch & Realtime Modes

Batch mode for bulk loads with full index rebuild on commit. Realtime mode for instant upserts and deletes with backpressure handling. Choose per collection.

Persistence

Indexes are persisted to disk by default. On restart, Torque reloads collections, schemas, aliases, synonyms, presets, stopwords, and API keys automatically.

Multi-Collection

Independent schemas per collection. Collection aliases point to any collection and can be swapped atomically - useful for zero-downtime schema migrations.

Zero-Downtime Rebuilds

Indexes are rebuilt in the background and swapped in atomically. Queries are never blocked - the old index serves requests until the new one is ready.

Nested Fields

Enable nested fields on a collection to use dotted field names. HTTP import auto-flattens nested JSON objects. TCP protocol supports flat field names directly.

Query Features

Geo Search

Location-based filtering and sorting with R-tree spatial indexing. Search within a radius and sort results by distance from a reference point.

Sorting & Grouping

Multi-field sort_by with support for integers, floats, and strings. group_by for result clustering. Decay sorting for time-based relevance. 58x faster sorting and over 200x faster grouping than Typesense v30.1.

Synonyms, Presets & Stopwords

Define synonym rules so users find results regardless of terminology. Save search presets to reuse complex queries. Configure stopwords to ignore common terms.

Developer Tools

Torque Studio

Web-based search IDE with Monaco query editor, collection browser, schema inspector, document import, and saved queries. Embedded in the Torque binary - no separate install needed.

REST API

Full Typesense v30.1 REST API. Collections, documents, search, multi-search, import/export, synonyms, presets, stopwords, API keys, and health endpoints.

Official SDKs

Client libraries for .NET, Python, Go, Java, and Node.js. Each SDK provides HTTP API coverage, TCP binary ingestion, TQBF binary file writing, and typed models.

API Key Authentication

Scoped API keys with per-collection and per-action permissions. Bootstrap, admin, and search-only key levels. Key expiration and HMAC-based scoped key generation.

Multi-Language Stemming

Snowball stemmers for 17 languages including English, German, French, Spanish, Portuguese, Italian, Dutch, Russian, Arabic, and more. Set language per document - mix languages in the same collection without separate indexes.

torque-admin CLI

Command-line tool for collection management: list, export (JSONL and TQBF binary), import (batch and realtime modes via TCP), delete documents by ID, and binary file operations.

Feature Comparison

TorqueTypesense
Search
Text search scoringBM25F (field-weighted)Custom token overlap
Vector search (HNSW)Yes (RaBitQ quantization)Yes (float32 only)
Hybrid searchYes (RRF + alpha weighting)Yes (alpha-weighted)
Prefix searchYes (FST-backed)Yes
Typo toleranceYes (SymSpell)Yes (Levenshtein)
GPU-accelerated searchYes (CUDA)No
Conversational & natural-language searchYes (any OpenAI-compatible LLM)Yes
Built-in MCP server for AI agentsYesNo
Filtering & Faceting
Filter by numeric/string/boolYes (roaring bitmaps)Yes
Faceted search with statsYesYes
Nested object filteringYesYes
Element-level array ANDYesYes
Geo search (radius + sort)Yes (R-tree)Yes
Sorting & Grouping
Multi-field sortingYesYes
Group byYesYes
Decay sortingYesYes
Ingestion
HTTP JSONL importYesYes
Binary TCP protocolYesNo
Binary file upload (TQBF)Yes (fastest method)No
Batch ingestion modeYesYes
Realtime upsert/deleteYes (with backpressure)Yes
Infrastructure
In-memory engineYes (fully in-memory)Hybrid (index in-memory, docs on RocksDB)
Disk persistenceYesYes (RocksDB)
Collection aliasesYesYes
GPU accelerationCUDA (search, filter, facet, vector)Embedding generation only (Cloud)
Scoped API keysYesYes
Developer Tools
Visual IDE (bundled)Torque StudioNo
Official SDKs (TCP + REST).NET, Python, Go, Java, Node.jsREST only (community)
SynonymsYesYes
Search presetsYesYes
StopwordsYesYes
Platform & Runtime
LanguageRustC++
WindowsYes (+ Windows Service)Docker only
LinuxYesYes
macOSYesYes

See it in action

Get a walkthrough of Torque features with your own data and use case.