FrontierTuner
Retrieval Pipeline Optimization

The right RAG pipeline for your data, found automatically.

FrontierTuner searches across every pipeline choice — chunking, embeddings, retrieval, reranking, and generation — and reasons about why each configuration falls short, converging on the best balance of accuracy and cost for your own corpus.

high low low cost high cost accuracy Recommended
Accuracy keeps rising with cost, but the gains shrink. FrontierTuner finds the knee of the curve — the configuration where extra spend stops buying meaningful accuracy.

How it works

Most optimizers reduce a trial to a number. FrontierTuner reads the failures.

Every configuration is scored on the same exam built from your documents, and every failure is traced to the stage that caused it.

01 / Exam

Score on your data

FrontierTuner builds an exam from your own corpus and freezes it, so every configuration is graded on identical questions and its results stay directly comparable.

02 / Diagnose

Find what broke

After each run it pins every failure to a stage — retrieval missed the context, the reranker buried it, or the model answered wrong — instead of collapsing the result to a single score.

03 / Propose

Make the next move

A reasoning agent, aware of how models rank and what they cost, reads the diagnosis and chooses the next configuration to try — skipping the ones already destined to be too weak or too expensive.

Why FrontierTuner

Stronger pipelines, found in fewer trials.

Better accuracy and retrieval

FrontierTuner finds configurations that answer more questions correctly and surface the right context — outperforming random and Bayesian search across our evaluations.

A fraction of the search cost

By reasoning about model strength and price before it spends a trial, FrontierTuner reaches a strong pipeline in a handful of runs instead of hundreds, using far less compute to get there.

Decisions you can audit

Each search returns a recommended pipeline, an accuracy-cost frontier, and a per-trial reasoning trace — so you can see exactly why each choice was made.

About

Pipeline tuning, turned into an automated decision.

Configuring a retrieval pipeline normally means an expensive, manual search across dozens of interacting choices. FrontierTuner automates it — and, unlike blind search, reasons about why each configuration succeeds or fails.

It began as research at the ETH Agentic Systems Lab and is now in early access. We work closely with early teams to tune retrieval on their own data — reach out if that's you.