Embedding Model

QANATIX uses BGE-M3 as its embedding engine. It's the only model that produces all three vector types — dense, sparse, and ColBERT — in a single pass, powering the hybrid search pipeline.

Why BGE-M3

Attribute	Value
Model	`BAAI/bge-m3`
Dimensions	1024
Cost	Free — runs locally, no API key
Vectors	Dense + Sparse + ColBERT
Languages	100+ (multilingual natively)
Self-hosted	Yes — works fully offline

Three vector types

Dense vectors — semantic understanding ("luxury with pool" matches high-end hotels with pools)
Sparse vectors — keyword precision (learned BM25-style, "Vienna" boosts Vienna results)
ColBERT vectors — token-level late interaction for fine-grained reranking

All three are fused via DBSF for maximum retrieval quality, then optionally reranked by a cross-encoder.

Configuration

EMBEDDING_MODEL=BAAI/bge-m3
EMBEDDING_DIMENSIONS=1024

No API key needed. The model downloads automatically on first run (~2 GB, cached at ~/.cache/huggingface/).

Hardware guidance

Dataset size	CPU	GPU
< 10K entities	Fine, ~1-5s per batch of 20	Not needed
10K – 100K entities	Works, slower pipeline	Recommended (NVIDIA, 8GB+ VRAM)
> 100K entities	Slow	Strongly recommended

GPU recommendations

GPU	VRAM	Throughput	Cost tier
NVIDIA T4	16 GB	~100 embeddings/sec	Budget
NVIDIA L4	24 GB	~300 embeddings/sec	Best value
NVIDIA A10G	24 GB	~300 embeddings/sec	Cloud standard
NVIDIA A100	40/80 GB	~800 embeddings/sec	High throughput

How it works

Entity ingested
    ↓
Worker picks up entity (SAQ background job)
    ↓
description_llm text → BGE-M3 → dense + sparse + ColBERT vectors
    ↓
Vectors indexed in Qdrant (all three types)
    ↓
Entity marked as "indexed" — now searchable

Caching

Cache	TTL	Key
Query embedding	1 hour	model + dimensions + query hash
Entity embedding	7 days	model + dimensions + text hash

Embeddings are cached in Redis. Re-ingesting identical text skips encoding entirely.

Embedding Model

On this page