Keyword search that ranks like the best
Multi-field full-text search with best-in-class relevance and a fast path that skips any document that cannot reach the top results.
‹ qOS
Agent Search· Torque
Search that thinks in keywords and meaning, at once.
Torque is a search engine rebuilt from scratch in Rust that serves every query from memory, so most lookups finish in well under a millisecond. It speaks the same API as Typesense, so existing apps switch with a config change, and underneath it adds true semantic search, hybrid ranking, and a built-in interface agents can query directly.
Sub-ms typical query, from memory
6 to 19x faster than Typesense, our benchmarks
4.6M documents searched in 3ms
Why Torque
Search forced a choice. We refused it.
Every search stack makes you pick a side. Bolt vectors onto a keyword engine and semantic search feels like an afterthought. Run a vector database and exact phrases, filters, and facets fall apart. And almost all of them lean on disk, adding latency to every single query.
Torque keeps the entire index in memory and does keyword and meaning-based search in one engine, while staying a drop-in replacement for the tools teams already run.
Capabilities
What it can do.
Semantic and hybrid search
Meaning-based vector search blended with keyword results, so a single query gets both exact-match precision and conceptual recall.
Typo tolerance and instant autocomplete
Fuzzy typo correction, prefix autocomplete, synonyms, and per-language stopwords, tuned across seventeen languages.
Filter, facet, sort, group
Rich filtering by number, geo-radius, nested field, or array, faceted counts with statistics, and sort by relevance, distance, or recency.
Live updates, zero downtime
Upsert and delete in sub-milliseconds, or swap an entire freshly-built index atomically without ever blocking a query.
A search tool agents can hold
A native Model Context Protocol endpoint exposes search as a self-describing tool, so an agent queries and manages the engine with no glue code.
Curation and federated search
Pin or hide specific results, run up to fifty searches in one parallel request, and save reusable query presets.
Cross-collection joins
Reference fields let a search pull in related documents from another collection and filter across the link.
Built different
Why it is not just another search box.
-
Rebuilt from the ground up, in memory
Not a fork. The index is stored as zero-copy bitmaps mapped straight from disk and served entirely from RAM, so it loads instantly and answers fast.
-
State-of-the-art vector compression
Uses RaBitQ binary quantization from recent database research for vectors up to thirty-two times smaller, with better accuracy than the usual approach and no retraining as data drifts.
-
GPU-ready data structures
Pointer-free layouts run on CPU today and accelerate on NVIDIA GPUs, with automatic fallback when no GPU is present.
-
Agent-first by design
The agent tool interface is part of the engine itself, a feature of the database, not a sidecar you bolt on.
-
Drop-in compatible
Speaks the Typesense API, so existing clients and tools connect unchanged while you gain semantic search underneath.
By the numbers
Fast where it counts.
Measured on real data, CPU only, in our own benchmarks.
Sub-msp50 on most query types
70K/sdocuments ingested per second
17languages stemmed
5zero-dependency client SDKs
Agent-first
A search tool an agent can hold.
Point an agent at Torque and it immediately has a self-describing search tool: discover collections, read schemas, run keyword or semantic queries with filters and facets, and read or write documents, with zero integration code. Sub-millisecond latency means an agent can fire many searches inside a single reasoning loop without stalling.
- A native agent interface with nine tools, including parallel multi-search.
- One query fuses exact keyword hits and semantic similarity for precise and related context.
- Write a document and search it back in sub-milliseconds: read-after-write memory.
Run it yourself
Torque is also released as a standalone product by Truespar, our research arm. Run it on your own infrastructure, the getting-started guide takes it from there.
Put Agent Search to work.
See HQ running in your own Slack or Teams, on the operating system we built for agents.