Search

QANATIX uses hybrid search combining dense vectors (semantic), sparse vectors (keyword), and exact-match identifier detection to return the most relevant results in < 200ms.

Query classification

Every query is automatically classified:

Examples:

• SS-M8-40-A2 → identifier — exact match, score = 1.0
• stainless bolt → keyword — BM25 weighted
• ISO 9001 certified M8 suppliers in Germany → semantic — hybrid fusion

Hybrid fusion

For semantic and keyword queries, QANATIX runs two searches in parallel:

1. Dense vector search — semantic similarity (OpenAI, Cohere, or BGE-M3)
2. Sparse vector search — keyword relevance (BM25 or model-learned sparse)

Results are fused using Distribution-Based Score Fusion (DBSF) — scores normalized by their statistical distribution, not min/max. Prefetch: 50 candidates per signal.

With BGE-M3 embeddings, a third signal is available:

3. ColBERT reranking — token-level late interaction for fine-grained matching (top 20 candidates)

Reranking

When enabled (default), results are re-scored with a cross-encoder model (BAAI/bge-reranker-v2-m3):

• Triggered when 10+ results are returned and rerank=true
• Over-fetches max(limit * 2, 20) candidates for the reranker pool
• Final score: 0.3 * retrieval_score + 0.7 * reranker_score

Zero-result fallback

If hybrid search returns no results, QANATIX automatically:

1. Relaxes filters (removes all filter constraints)
2. Falls back to dense-only search
3. Falls back to Postgres full-text search (degraded mode)

The response metadata.search_mode indicates which strategy was used.

Scoring

Results are scored 0-1 (DBSF normalized). Higher is better.

Results with score < 0.05 are filtered out.

Identifier matches always return score = 1.0.

Caching

Next

• REST API — full endpoint reference
• Filters — structured filtering with range operators
• Response Formats — compact, full, yaml