Truespar Torque
A high-performance search engine for instant, typo-tolerant search, with hybrid vector search and optional GPU acceleration. Typesense-compatible API and sub-millisecond filtering and sorting - built in Rust for predictable performance.
In-Memory Speed, Disk Durability
Indexes live in memory for instant access. Persistence saves your data to disk automatically - restart the server and pick up where you left off. No re-ingestion required.
GPU Acceleration
CUDA compute kernels offload scoring, filtering, vector search, and facet counting to NVIDIA GPUs. Full CPU fallback when no GPU is available - same API, same results.
Typesense API Compatible
Drop-in Typesense v30.1 REST API. Your existing clients and integrations work out of the box - switch engines, not code.
A Drop-In Typesense Alternative
Torque serves the Typesense v30.1 API, so switching is a config change: point your existing clients at Torque and they work. On 309,000 production records it measured 9.5x Typesense's throughput, winning all 15 query patterns, CPU-only. And if you pay per search for a hosted service such as Algolia, Torque runs on your own hardware with no per-search pricing. See the benchmarks.
Search Capabilities
Hybrid Search
Combine BM25F text scoring with HNSW vector search using reciprocal rank fusion. Get keyword precision and semantic recall in a single query.
Advanced Filtering
Roaring bitmap intersections for filter_by, facet_by, sort_by, group_by, and geo search. Element-level AND filtering ensures conditions match within the same array element.
Multi-Collection
Independent schemas per collection with aliases for zero-downtime schema migrations. Queries are never interrupted during index rebuilds.
Ingestion
Binary TCP Ingestion
Custom high-performance binary protocol for streaming documents. Batch mode for bulk loads, realtime mode for instant upserts and deletes. 7.8x faster than Typesense ingestion on the same hardware.
Typesense REST API
Full v30.1 HTTP API. Collections, documents, search, import/export, synonyms, presets, stopwords, and API key management - all compatible with existing Typesense clients.
Official SDKs
Client libraries for .NET, Python, Go, Java, and Node.js. Each SDK provides full HTTP API coverage, TCP binary ingestion, and TQBF file writing with typed models.
Language & Intelligence
Multi-Language
Snowball stemming for 17 languages including English, German, French, Spanish, Portuguese, Italian, Dutch, Russian, and more. Language is set per document - search across multiple languages in the same collection.
Vector Search
HNSW graph index with RaBitQ binary quantization for high-accuracy approximate nearest neighbor search. GPU-accelerated distance calculations via CUDA and cuBLAS.
Typo Tolerance
SymSpell fuzzy correction handles misspellings automatically. FST-backed prefix search for instant type-ahead suggestions as users type.
Frequently Asked Questions
Is Torque a drop-in replacement for Typesense?
Yes. Torque serves the Typesense v30.1 REST API, so existing Typesense clients, SDKs, and integrations connect by changing the endpoint. Collections, documents, search, synonyms, presets, stopwords, and API keys all work.
How fast is Torque compared to Typesense?
On 309,000 production records, Torque measured a 9.5x geometric mean speedup over Typesense v30.1 across 15 query patterns, winning every pattern, CPU-only. Ingestion measured 7.8x faster on the same hardware.
Does Torque support vector and semantic search?
Yes. An HNSW index with RaBitQ binary quantization serves approximate nearest neighbor queries, and hybrid search fuses BM25F text scoring with vector results using reciprocal rank fusion.
Does Torque need a GPU?
No. Torque runs CPU-only by default, and the published benchmark numbers are CPU-only. An optional CUDA path offloads scoring, filtering, vector search, and facet counting to NVIDIA GPUs with the same API and results.
Is Torque self-hosted?
Yes. Torque is one binary that runs on your own hardware, on Windows and Linux. There is a free tier for individuals and companies; see the pricing page.
Does Torque handle typos and type-ahead search?
Yes. SymSpell fuzzy correction handles misspellings automatically, and FST-backed prefix search returns instant type-ahead suggestions as users type.
See the numbers
Benchmarked on 309K production records against Typesense v30.1. 9.5x geometric mean improvement across 15 query patterns - text search, faceted search, filtering, geo, and grouping - all CPU-only, winning every pattern.