Stop ranking. Start selecting evidence.
What if the evidence you need is ranked 49th?
Most retrieval pipelines rely on rerankers and fixed top-k cutoffs. Results are scored, the top few survive, and everything else is discarded.
But rank is not evidence.
Tarka is yukti's System 1 model for evidence selection. Give it 50 or more search results, and it identifies the evidence that matters, producing a structured evidence set for downstream reasoning.
No prose generation. No lengthy explanations. Just the selection your system needs.
Search results50 candidates
Evidence set0 selected
- 01···rank ·
- 02···rank ·
- 03···rank ·
- 04···rank ·
- 05···rank ·
Search results in. Evidence set out.
01 · Selection
Beyond the top-k cutoff.
A lower-ranked result might contain the missing fact, resolve a contradiction, or complete the evidence needed to answer a question. Tarka is designed to look beyond a fixed ranking cutoff and select evidence based on the task at hand.
Large candidate sets.
Process 50 or more retrieved results.
Evidence-first selection.
Don't let rank alone determine what survives.
Structured output.
Return an evidence set ready for downstream models and agents.
02 · Pipeline
The evidence layer.
SearchTarkaReason
Search
Anvesh or another retrieval system finds candidate information.
Select
Tarka determines which results belong in the evidence set.
Reason
A downstream model uses that evidence to produce an answer or complete a task.
Tarka handles the selection. Other systems handle the reasoning that follows.
Benchmarks
Benchmarks coming soon.
We're preparing evaluations to test evidence selection across large candidate sets, including cases where critical information appears far down the ranking.
Our goal is to demonstrate when Tarka can preserve evidence that conventional reranking and fixed top-k pipelines might miss.
Resultsnot yet published
- Large candidate setspending
- Far down the rankingpending
- Versus fixed top-kpending
03 · Pricing
Pricing built for scale.
- Input tokens$0.042/ 1 million tokens
- Output tokensFree
$42per billion input tokens
That's $42 per billion input tokens. Tarka produces structured decisions rather than generated prose, so output tokens are free.
Estimate your month
Input tokens$0
Output tokens$0
Estimated usage cost$0
Early access
Get early access.
Tarka is coming soon. Join the waitlist for benchmark results, launch updates and early access.
Tarka by yukti
The right evidence changes everything.