VYASA LABS · RESONANT · SEMANTIC RETRIEVAL WITH TONE

Meaning and tone, finally on separate axes.

Every embedding model crushes semantics, style, and intent into one vector. Resonant learns them apart so retrieval can rank by what was meant, how it was said, or both.

Request a demo See the benchmarks

INPUT TEXT →

MEANING
vector

TONE
vector

RESONANT ⟂

99.4%

TONE MATCH

Target register ranked first

0.89

nDCG@10

Retrieval quality preserved

98.5%

PROBE ACC

<2%

COST

Retrieval quality traded

HOW IT WORKS

One pass. Two vectors. You choose how to rank.

Resonant encodes each document once, then projects it into two independent heads. At query time, a single weight α dials how much tone influences the ranking.

QUERY

“calm refund policy for a late delivery”

Resonant encoder

shared transformer

MEANING HEAD

Semantic vector: what is being asked

TONE HEAD

JOINT SCORER

score = meaning
+ α·tone

Set α = 0 for pure semantic search. Raise it to make tone matter. No re-indexing required.

WHY RESONANT

Retrieval that understands communication style.

Every retrieval system today finds documents about the right topic. None of them can tell you whether the document is written in the right tone for your user.

Resonant adds an orthogonal tone axis to a high-performance retriever without degrading retrieval quality. One model, two independent signals, no tradeoff.

WHAT YOU CAN BUILD

→Find technically correct answers written for beginners.

→Retrieve prior customer emails in the same tone as the current conversation.

→Build RAG that respects both meaning and communication style.

→Surface formal documentation, not the casual Slack thread about the same topic.

→Match empathetic responses to frustrated customers.

BUILD WITH RESONANT

A tone axis changes what you can ship.

Five things that are awkward with one vector and natural with two.

resonant

⌕refund policy for a late delivery

tone: reassuring · α 0.3

3 on-topic matches · re-ranked by register

#1Late-delivery refunds — customer FAQreassuring

"If your order arrives late, you're fully covered — here's how to get your money back in a few clicks."

MEANING

TONE

#2Refund processing SLA (internal ops)procedural

"Refunds are issued within 5-7 business days per policy §4.2 pending finance approval."

MEANING

TONE

#3Chargeback escalation runbookterse

"Escalate to tier-3 immediately if the customer initiates a chargeback. Do not refund."

MEANING

TONE

All three mean the same thing. Resonant puts the calm, customer-ready answer first.

BENCHMARKS · MEASURED, NOT CLAIMED

Turn on tone. Lose almost nothing.

The same model, our own eval. As the tone weight α rises, documents start matching the target register while retrieval quality barely moves.

TONE WEIGHTα = 0.3 · tone-forward

0.00.10.20.3

99.4%

REGISTER MATCH

0.890

nDCG@10

— nDCG@10— register match

BEFORE / AFTER · same system

α = 0α = 0.3

Finds the right topic✓✓

Finds the right tone8.8%99.4%

nDCG@100.9040.890

TONE SEPARATION · length-controlled

98.5%

PROBE ACCURACY

vs 76.8% length-only baseline

Tone gap holds at ~0.45 whether passages are short or long. Register is a real axis, not a length artifact. (32 items × 4 length-controlled variants)

TRY IT LIVE

See tone-aware search in action.

Type a query, pick a tone, and watch the results re-rank. Same documents, different priority based on communication style.

TONE:

α=0.3

Hit search to see tone-aware results.

QUICKSTART

One call. Semantic search with tone.

Pass a query and a target tone. Get back ranked documents that match both meaning and communication style.

01Specify tone as a string or a weight.

02Results include relevance and tone scores.

03No reranker, no classifier, no second call.

from resonant import Resonant

client = Resonant(api_key="rsn-...")

# Search with tone preference
results = client.search(
    query="How do I reset my password?",
    tone="formal",
    collection="help_docs",
    limit=5,
)

for doc in results:
    print(doc.text, doc.relevance, doc.tone_score)

Build systems that understand more than words.

30-minute technical walkthrough. No sales deck. Bring your retrieval problem.

Request a demo