Menu

Use case

Which examples deserve space in my context window?

Context windows are a budget, not a box. Kno ranks candidate examples and policies by improvement per token, then selects the portfolio that fits.

The problem

Every few-shot example and policy in your context window costs tokens on every single call. Teams add them one at a time and never measure whether they pay rent.

The workflow

  1. Baseline

    Measure the agent without the candidates.

  2. Value

    Inject each candidate and record its delta with a confidence interval , plus delta-per-cost, so a tiny example beating a giant document is visible.

  3. Select

    Build a portfolio under a context budget (tokens) and a cost budget, greedy on delta-per-cost, with every decision at a Bonferroni-corrected interval.

  4. Export

    Render the selected examples as a context pack.

What it looks like

kno
budget    context ≤ 10000 tokens; cost ≤ $5.00

RANK  ASSET          DESTINATION  DELTA (95% CI)
1     refund-policy  context      +0.4000 [-0.1000, 0.9000]
2     pricing-tier   context      +0.3000 [-0.0500, 0.6500]

The recipe

Put every candidate example, instruction, and policy in a pool. Kno routes each one to the cases it could plausibly affect, re-measures against fresh controls, and reports the delta with its interval.

What you get

A portfolio under your token budget, ranked by delta-per-cost, with a rejection log explaining why the rest did not make it.

Stages involved

Try it →