Amgix - Open-Source Hybrid Search Engine

Amgix (pronounced a-MAG-ix) - short for Amalgam Index
amalgam: a mixture or blend of different elements

Amgix is a hybrid search engine for applications. It provides a unified stack for ingestion, embedding, and search — combining keyword and semantic retrieval in an architecture that can scale from a single container to distributed services.

Amgix lets you add fast, relevant, modern search to your application without stitching together separate systems for indexing, embeddings, and retrieval.

How It Works

0. Run Amgix One:
Amgix One packs the API, encoder, RabbitMQ, and Qdrant into one container — the easiest way to try Amgix or run it with modest requirements.

docker run -d -p 8234:8234 -v /path/on/host:/data amgixio/amgix-one:1

This persists data and caches Hugging Face models under /data.

Use the short tag 1 for the latest 1.x release. For GPU support use amgixio/amgix-one:1-gpu (requires NVIDIA Container Toolkit).

1. Define your collection:

POST /v1/collections/products
{
  "vectors": [
    {"name": "keyword", "type": "keyword", "index_fields": ["name", "content"]},
    {"name": "semantic", "type": "dense_model", "model": "sentence-transformers/all-MiniLM-L6-v2", "index_fields": ["content"]}
  ]
}

2. Upload your data:

POST /v1/collections/products/documents/bulk
{
  "documents": [
    {
      "id": "part-001",
      "timestamp": "2026-03-15T00:00:00Z",
      "name": "Roller 12LP'-x03/5-XL",
      "content": "Precision pinch roller assembly for manufacturing."
    },
    {
      "id": "part-002",
      "timestamp": "2026-03-15T00:00:00Z",
      "name": "Bearing 12LP'-y03/5-XL",
      "content": "Deep groove ball bearing, double shielded."
    },
    {
      "id": "part-003",
      "timestamp": "2026-03-15T00:00:00Z",
      "name": "Belt 12LP'-x03/8-MD",
      "content": "Synchronous timing belt for power transmission."
    }
  ]
}

3. Search:

POST /v1/collections/products/search
{
  "query": "12lpy03"
}

Amgix's built-in keyword tokenizer handles missing punctuation to correctly return the Bearing (12LP'-y03/5-XL).

POST /v1/collections/products/search
{
  "query": "motor energy transfer loop"
}

Even without keyword overlap, Amgix's semantic vector understands the concept and returns the Belt (Synchronous timing belt for power transmission).

Feature Highlights

True Hybrid Search: Combine dense semantic models, sparse models (SPLADE), and built-in tokenizers (WMTR, full-text, whitespace).
Server-side Fusion: Send one query, let Amgix execute all vector searches and fuse the results using weighted Reciprocal Rank Fusion (RRF).
Production-Ready Ingestion: Built-in asynchronous processing queue with automatic retries, timestamp-based deduplication, and distributed locking.
Adaptive Model Orchestration: Models are automatically loaded, unloaded, and balanced across nodes based on traffic and available RAM/VRAM.
Backend Agnostic: Run on Qdrant, PostgreSQL, or MariaDB without changing a single line of your application code.
Developer Friendly: Clean REST API, boolean metadata filtering, tag filtering, score thresholds, and automatic model dimension discovery.

Why Amgix

Integrating modern search into applications is hard. You usually have to stitch together multiple systems for retrieval, embeddings, and indexing, build custom re-indexing pipelines, handle deduplication and failures, and maintain ML embedding workflows.

Amgix was built to solve these headaches. Instead of building your own infrastructure, you get a single unified stack with:

Built-in ingestion queues: Async processing with automatic retries and deduplication.
Adaptive model orchestration: Models load and rebalance across nodes automatically based on demand.
Server-side fusion: One API call handles query vectorization, semantic search, keyword search (including our custom tokenizer for messy identifier data), and ranking.
Backend Agnostic: Run on Postgres, MariaDB, or Qdrant without changing your code.

Get Started

Quickstart Guide