Release Notes

Version history and changelog for Torque.

0.6.1 - 2026-07-14

Code-signing release - functionally identical to 0.6.0.

• Windows binaries (torque-server.exe, torque-admin.exe) are Authenticode-signed by "The Intelligence Company AB (publ)" with an RFC 3161 timestamp
• The Linux tarball ships with a detached GPG signature (.asc) and SHA-256 checksum
• Client SDKs are unchanged and remain at 0.6.0

0.6.0 - 2026-07-14

Typesense v30.1 API coverage complete, LLM-powered search, an MCP server for AI agents, crash-safe persistence, and the fastest query performance of any Torque release.

Typesense v30.1 API complete

• Curation sets - pin, hide, filter, and re-rank results for specific queries via the /curation_sets API, with tag and time-window matching
• Analytics - popular-queries and no-hits rules with automatic search capture, counter rules, and aggregation into destination collections
• Bulk update by filter (PATCH /documents?filter_by=) with atomic validation; bulk delete requires filter_by or explicit truncate=true
• Document write actions create / upsert / update / emplace, plus return_doc on imports
• Custom stemming dictionaries (/stemming/dictionaries + per-field stem_dictionary)
• Operations endpoints: snapshot, compact, cache clear, schema changes, /metrics.json, runtime /config
• Multi-search union mode merging all searches into one ranked result set
• New field types: object / object[] / auto with import-time sub-field discovery, string*, bool[], geopoint[]; collection metadata and collection-bound synonym_sets
• Full search parameter coverage: advanced typo tolerance, ranking controls, highlight_full_fields / snippet_threshold, facet sampling, and more
• Typesense bare-colon filter syntax (field: value, field:[a, b])

LLM-powered search

• Conversational search (conversation=true) - grounded answers over your top hits with durable multi-turn history and SSE streaming (conversation_stream=true)
• Natural-language search (nl_query=true) - an LLM translates free-form questions into structured search parameters before the search runs
• Auto-embedding - embed config on vector fields fills embeddings at import time; vector_query embeds query text server-side
• All three work with any OpenAI-compatible endpoint, hosted or self-hosted

MCP server

• POST /api/mcp exposes nine tools (search, multi-search, schema, document and collection CRUD) to Claude Desktop, Claude Code, and any MCP client - enabled in all shipped binaries

Performance

• 9.5x faster than Typesense (geometric mean over 15 query patterns on a real 309K-document dataset, single client; wins every pattern) and 7.4x faster than Quickwit
• Infix search on any searchable field, TTL-bounded search result cache (use_cache / cache_ttl)
• Pure-Rust storage engine (roaring-rs) replacing the previous C bitmap library

Reliability

• Crash-safe generation-scoped index persistence - an interrupted save can no longer corrupt a collection
• Segment compaction handles multi-gigabyte text-heavy collections; large text stays memory-mapped through compaction
• Import memory admission is streaming-aware, enabling multi-GB binary imports on smaller hosts

SDKs (all at 0.6.0)

• Typed search parameters cover the complete v30.1 set in all five SDKs
• New union multi-search helpers in every SDK
• Python fix: boolean search parameters now serialize correctly (previously rejected by the server) - upgrade the SDK with the server
• Go SDK is distributed as its own download

0.5.0 - 2026-04-11

First public release of Torque - Truespar's high-performance search engine. Typesense v30.1 API compatible.

Search

• Full-text search with relevance ranking and multi-field queries
• Filters: exact match, numeric range, boolean, array, geo radius, and nested object conditions
• Sorting: field, text relevance, geo distance, and decay functions (Gaussian, linear, exponential)
• Faceted search with stats (min, max, sum, average)
• Grouping with configurable per-group limits
• Prefix / autocomplete search
• Typo tolerance with automatic correction
• Vector search (dense embeddings) with hybrid relevance blending
• Synonyms (bidirectional and root), presets, stopwords
• Highlighting with configurable tags
• Stemming for 17 languages: English, German, French, Spanish, Portuguese, Italian, Dutch, Swedish, Norwegian, Danish, Finnish, Russian, Arabic, Greek, Hungarian, Romanian, Turkish

Ingestion

• HTTP JSONL import - Typesense-compatible import endpoint with create, upsert, and update actions
• TCP binary ingest - custom high-throughput streaming protocol for live database sync (70K docs/s measured on a 309K-document benchmark, 7.5× faster than Typesense on the same hardware)
• TQBF binary file upload - pre-encoded compressed file format for bulk imports, ETL pipelines, and data distribution
• Batch mode for full index rebuilds on commit
• Realtime mode for instant upsert and delete with backpressure handling

API Compatibility

• Typesense v30.1 compatible HTTP REST API - drop-in replacement for existing Typesense clients
• Full collection CRUD, document CRUD, search, multi-search, aliases, API key management
• Collection aliases for zero-downtime schema migrations
• HMAC scoped search keys

Persistence & Durability

• Full in-memory serving with automatic disk persistence
• Realtime overlay persistence with sub-second flush
• Atomic writes with integrity checks (temp file + fsync + rename)
• Automatic collection, schema, alias, key, synonym, preset, and stopword restore on startup

Server & Operations

• TLS / HTTPS support
• Prometheus metrics endpoint at /metrics
• HTTP IP allowlist via --allowed-ips / TORQUE_ALLOWED_IPS (CIDR ranges)
• Configurable collection deletion guard (allow_drop_collection) for production hardening
• Configurable Studio gating (--no-studio) for API-only deployments
• Auto-generated bootstrap API key on first run (persistable via --api-key, env var, or TOML)
• Slow query logging with configurable threshold
• Multi-node failover and retry across nodes (configured client-side)
• SIGTERM graceful shutdown (Docker and systemd compatible)
• Daily rotating log files
• Windows Service support (--service-install / --service-uninstall)
• TOML configuration file, environment variables, and command-line arguments

GPU Acceleration

• NVIDIA CUDA acceleration for scoring, filtering, vector distance, and facet counting
• Shipped as a separate GPU-enabled build artifact - contact Truespar for availability
• Automatic CPU fallback when GPU hardware is unavailable

Torque Studio

• Embedded web UI served from the torque-server binary - no separate install
• Search editor with syntax highlighting and schema-aware completions
• Schema browser, collection management, and data import panels
• Benchmark comparison panel (Torque vs Typesense vs Quickwit)
• API key management and license / device management
• In-app update notifications via /api/update/download proxy
• Dark and light theme

Client SDKs

All five SDKs provide the full HTTP API, TCP binary ingest, and TQBF binary file writer. Each one targets zero external runtime dependencies.

• .NET 10+ - Torque.Http on NuGet. Source-generated JSON serialization.
• Python 3.10+ - torque-http on PyPI. Standard library only. Typed search parameters and response models.
• Node.js 20+ - @truespar/torque-http on npm. Native fetch, ESM, TypeScript declarations included.
• Go 1.23+ - torque-sdk-go downloadable from truespar.com. Context-aware API, go.mod replace directive.
• Java 21+ - com.truespar:torque-http on Maven Central. Built on Java's built-in HttpClient. Typed SearchParameters, SearchResult, and Hit models.

Admin Tooling

• torque-admin CLI for collection management, JSONL export/import, TQBF binary file export/import, and document deletion by ID

Distribution

• Windows x64 ZIP
• Linux x64 tarball
• Docker image: truespar/torque:0.5.0