Outperformed Claude Code, Codex, and Mini-SWE-Agent on repository search·See what better retrieval means

▸ Parieto // Repository search   Research file · 2026

PARIETO

Repository intelligence, benchmarked.

848Repository-search cases
0.954HitFile
15.0kTokens to answer

Benchmark · Parieto Research

A Benchmark for Repository Intelligence

Repository search only matters when it gives the answer model the right code, with enough coverage to answer and little enough noise to reason clearly.

Across 848 repository-search cases, Parieto outperformed Claude Code, Codex, and Mini-SWE-Agent on the retrieval measures that most directly shape an AI product: reaching the right file, returning answer-bearing lines, balancing signal and coverage, and keeping the context budget compact.

The practical result is simple. The model starts closer to the answer, has more room to reason, and spends fewer tokens getting there.

What better retrieval changes

HitFile · 0.954

The right file, more often

A relevant file appears in the returned regions in more than 95% of evaluated cases, so the answer model begins in the right part of the repository.

F1 · 0.378

Better context for the answer model

Leading F1 combines useful precision with stronger line coverage, giving the model more of the code it needs without burying it in unrelated context.

Tokens · 15.0k

Far fewer tokens per answer

A compact evidence set leaves more context for reasoning and lowers the cost of repeated searches across large codebases.

The comparison

Parieto leads the supplied comparison on HitFile, precision, LineRecall, F1, and tokens to answer.

SystemHitFilePrecisionLineRecallnDCG@500F1Tokens to Answer
Parieto0.9540.4860.3090.4550.37815.0k
Claude Code0.6670.2930.1540.9380.202~630k
Codex0.6490.2620.1940.9010.223~500k
Mini-SWE-Agent0.6400.2560.1510.8850.190~200k

How we built the benchmark

Each measure tests a different part of the path from a repository question to answer-ready code.

The evaluation covers 848 repository-search cases. For each case, a system retrieves regions from a codebase and those regions are checked against the files and lines needed to answer the question.

Together, the measures capture whether search reaches the right file, how much useful code it returns, how much answer-bearing code it covers, how the result is ordered, and how much context the answer model must read.

MeasureHow it is calculatedWhy it matters
HitFileShare of cases where a returned region names a gold file.Shows whether retrieval reached the right place.
PrecisionGold lines returned divided by total lines returned.Measures the signal delivered to the answer model.
LineRecallGold lines returned divided by total gold lines.Measures how much answer-bearing code was recovered.
nDCG@500Ranking quality of the returned order, discounted by rank.Shows how useful evidence is ordered in the result.
F1The harmonic mean of precision and LineRecall.Captures the balance between signal and coverage.
Tokens to AnswerTokens needed to place the located code in the model context.Measures how much context is spent before reasoning begins.

Retrieval quality, made visible

Parieto reaches answer-bearing code more consistently while keeping the resulting context compact.

HitFile

Share of cases that reach a gold file
ParietoExternal
Parieto
0.954
Claude Code
0.667
Codex
0.649
Mini-SWE-Agent
0.640

Figure 01 · Reaching the right file gives the answer model a dependable starting point.

F1

Combined precision and LineRecall
ParietoExternal
Parieto
0.378
Codex
0.223
Claude Code
0.202
Mini-SWE-Agent
0.190
Context required15.0k tokens

Compared with approximately 200k to 630k tokens across the external systems in the supplied evaluation.

Figure 02 · Stronger retrieval gives the model more answer-ready context with far fewer tokens.

What each metric changes in the product

The benchmark becomes useful when each number maps to a better answer experience.

01

HitFile

How often search reaches a file that contains the answer.

Find the right place first.
02

Precision

How much returned code directly contributes to the answer.

Give the model less noise.
03

LineRecall

How much answer-bearing code reaches the model.

Return enough to finish the task.
04

nDCG@500

How relevant code is ordered within the result.

Put useful evidence near the top.
05

F1

How well retrieval balances signal quality and coverage.

Improve the working context.
06

Tokens to Answer

How much context is spent before answering begins.

Leave more room to reason.

Published August 27, 2026 by Parieto Research.

Bring us one repository question ↗