Amgix Features

A scannable overview of what Amgix supports. For details on any item, see the linked guides and pages.

Search

• Hybrid search (keyword + semantic, fused in a single query) — Search Query Guide
• Multiple fusion algorithms: RRF (default), Linear — Search Query Guide
• Per-query, per-vector-field tunable weights — Search Query Guide
• WMTR: typo-tolerant, identifier-aware keyword tokenizer, no model/training/GPU required — WMTR Benchmarks
• Multiple lexical tokenizer options: WMTR (aliased keyword), Full Text, Trigrams, Whitespace — Collections Guide
• Per-query WMTR trigram weight tuning — Search Query Guide
• Dense (model-based) vector search, auto-detected dimensions — Collections Guide
• Sparse (model-based) vector search (SPLADE-style) — Collections Guide
• Custom vector support: bring your own dense or sparse embeddings — Search Query Guide
• Asymmetric query/document embedding models — Collections Guide
• Group by: cap results per distinct metadata field value, with multi-fetch backfill — Search Query Guide
• Faceting: counts over a configurable candidate window for indexed metadata fields — Search Query Guide
• Metadata filtering: structural (nested and/or/not) and string-expression syntax, with null checks — Search Query Guide
• Tag filtering, always indexed, with ALL/ANY match modes — Search Query Guide
• Collection joins: query-time, expression-based, filterable, left join semantics — Search Query Guide
• Field exclusion from results — Search Query Guide
• Score thresholds — Search Query Guide
• Raw per-vector scores on demand — Search Query Guide
• Multi-language support: stopwords and stemming, with auto-detect option — Collections Guide
• Configurable distance metric for dense vectors: cosine, dot, euclid — Collections Guide

Documents & Collections

• Fixed, opinionated document schema for predictable indexing and performance — Documents Guide
• Free-form metadata with opt-in indexing per field — Documents Guide
• Tags: always-indexed, filtering shortcut, distinct from metadata — Documents Guide
• store_content option to omit stored text after indexing, reducing payload and storage size — Collections Guide
• Supply pre-computed vectors at upload time to skip re-embedding — Documents Guide
• Export collections with vectors included for zero-re-embedding migration/backup/distribution — Documents Guide

Ingestion

• Synchronous upsert endpoint for immediate search consistency — Ingestion Guide
• Asynchronous upsert endpoints (single-document and bulk) for high-throughput, non-blocking writes — Ingestion Guide
• Bulk upload — Ingestion Guide
• Automatic timestamp-based deduplication on both upserts and deletes, guarding against out-of-order writes and race conditions — Ingestion Guide
• Content-hash change detection: skips re-embedding unchanged documents — Ingestion Guide
• Separate internal queues for bulk vs. single-document writes so bulk jobs can't starve interactive writes — Ingestion Guide
• Differentiated retry policy: short retries for CPU-bound embedding failures, extended retries for I/O-bound backend/broker failures — Ingestion Guide
• Distributed locking for atomic writes across a cluster — Ingestion Guide
• Synchronous and asynchronous delete endpoints, timestamp-deduplicated — Ingestion Guide
• Per-document status inspection — Ingestion Guide
• Collection-level queue statistics and doc counts — Ingestion Guide
• Queue purge endpoint — Ingestion Guide

Deployment

• Single container to fully distributed cluster, same API and storage format — Amgix Building Blocks
• Multiple backend options: Qdrant, PostgreSQL, MariaDB, selected via connection URL — Database Backends
• Independently scalable API, encoder, and storage layers — Under the Hood
• Encoder node specialization: dedicate nodes to ingestion-only or query-only workloads — Under the Hood
• Adaptive model orchestration: automatic model-to-node assignment based on demand, no manual config — Under the Hood
• Message-queue-based embedding pipeline: encoder failures don't affect the API, no HTTP timeout buildup on bulk ingestion — Under the Hood
• Backend-native ingestion queue: no separate Redis/Kafka needed for async processing — Under the Hood
• CPU or GPU embedding — Getting Started

Security

• API key authentication: Admin, Read, Search tiers — Getting Started
• Zero-downtime key rotation — Getting Started

Client Libraries

• Python — PyPI
• TypeScript — npm
• Rust — crates.io
• C# — NuGet

Observability

• Metric definitions endpoint — Metrics Guide
• Prometheus-compatible metrics export — Metrics Guide
• Live rolling-window metrics per node, including per-node resource state — Metrics Guide
• Historical metrics with configurable resolution — Metrics Guide
• Built-in web dashboard: cluster overview, API/indexing/embedding metrics, cluster topology map, live query explorer with per-vector raw scores — Release v1.2.0