HyperSearch: up to 7 signals,
76ms warm, 1ms cached

HyperSearch ("mode": "fusion" on POST /v1/search) combines up to 7 search signals in one API call — with confidence scoring that tells your LLM when to say "I don't know." The default mode, hybrid, combines vector and keyword (BM25) search.

7
Max signals (fusion mode)
~76ms
Fusion mode (warm, April 2026 run)
1ms
Cached repeat query
$29
Per month for 1M vectors

How HyperSearch Works

In fusion mode ("mode": "fusion" on POST /v1/search; /v1/answer uses it too), FluxVector scores up to 6 retrieval signals in parallel, fuses them, re-ranks the top candidates and scores confidence. The default mode is hybrid: vector + keyword.

Semantic
Deep meaning understanding. Finds relevant results even when words don't match.
Multilingual
100+ languages. Search in English, find results in Spanish or Chinese.
Fast Pre-filter
Small-model candidate screening. Optional: off by default, switched on per server.
Keyword
Exact term matching. Never misses results containing your exact words.
Expansion
Finds synonyms and related terms. "car" matches "automobile", "vehicle".
Fine-grained
Word-level precision. Understands which specific words matter most.
Re-ranking
Neural re-scoring of top candidates. Reads query and document together.
Confidence
Anti-hallucination scoring. Tells your LLM when to say "I don't know."
Pipeline StageTimeWhat It Does
Cache check<1msReturn cached result if repeat query
Multi-model embedding~45msParallel inference across multiple models
Multi-signal retrieval~25msUp to 6 retrieval signals scored in parallel
Score fusion<1msProprietary score-aware fusion
Re-ranking~3msNeural reranking on top candidates
Confidence scoring<1msAnti-hallucination confidence (0.0 – 1.0)

Total warm: ~76ms. Batched (10+ concurrent): ~10ms/query. Cached: <1ms.

Search Latency: HyperSearch vs Competitors

FluxVector's ~76ms (fusion mode) includes embedding + multi-signal search + reranking + confidence. Competitors require a separate OpenAI call and offer 1–2 signals.

Server Processing Time (all signals included)

FV cached
1ms
FV batched
~10ms/query
FluxVector
~76ms (fusion mode)
Pinecone
50–100ms (1 signal) *
Qdrant
30–80ms (1–2 signals) *
Weaviate
40–100ms (1–2 signals) *

* Competitors don't embed text — add ~100–200ms for OpenAI embedding call. And they use 1–2 signals, not 8.

Adaptive Batching: More Users = Faster

Concurrent UsersTotal TimePer-User Latency
1~76ms~76ms
10~100ms~10ms
100~120ms~1.2ms

Batch embedding is 5x more efficient. The more concurrent users, the faster each query gets. Inverse scaling.

Signal Count Comparison

SignalsBuilt-in EmbedRerankerFine-grainedFusionConfidenceAnti-Hallucination
FluxVector 8 Yes Yes (2.9ms) Yes Yes Yes Yes
Pinecone 1 No No No No No No
Qdrant 1–2 No No No No No No
Weaviate 1–2 Partial No No No No No

Anti-Hallucination: Confidence Scoring

Every HyperSearch response includes a confidence score (0.0 – 1.0). When confidence drops below 0.2, the response includes "warning": "low_relevance".

ConfidenceMeaningRAG Action
0.8 – 1.0 Strong match across multiple signals Use results with high trust
0.3 – 0.8 Moderate match, some signals weak Use results, note uncertainty
0.0 – 0.3 Poor match — warning: "low_relevance" Tell user "I don't have reliable info"

No other vector DB tells you when your retrieval failed. This is the missing piece for reliable RAG.

Monthly Cost (1M vectors, ~10K queries/day)

FluxVectorPineconeQdrant Cloud
Vector DB $29/mo $70/mo $65/mo
Embedding API (OpenAI) $0 ~$15/mo ~$15/mo
Reranker API (Cohere) $0 (built-in) ~$10/mo ~$10/mo
Total $29/mo ~$95/mo ~$90/mo

FluxVector embeds, reranks, and scores confidence on its own hardware. No per-token charges.

Methodology

ServerHetzner Dedicated (AMD dedicated CPU, 64GB RAM, NVMe SSD)
Dataset100 documents, varied topics
ModelsMultiple proprietary models, optimized for quality and speed
Search modeFusion (HyperSearch, up to 7 signals, score-aware fusion)
CacheRedis (60s TTL). Cold = first query. Cached = repeat within 60s.
LocationClient in Europe → Server in Germany
Competitor dataPublished docs + community benchmarks (single-signal query only)
DateApril 2026

Run your own benchmarks: pip install fluxvector

Try HyperSearch yourself

Free tier: 10K vectors, no credit card required. Hybrid search by default; "mode": "fusion" for HyperSearch.

Get API Key