RepoDNA v1.4.5: Building Smarter Offline Repository Intelligence

Understanding a software repository is more than reading its README or scanning its directory structure.

Developers need to understand architecture, dependencies, historical changes, code quality signals, and the evidence behind technical conclusions.

That's why I've been building RepoDNA, an open-source, local-first repository intelligence and code archaeology platform.

Today, I'm sharing two stable releases:

RepoDNA v1.4.4 — October 28, 2026

RepoDNA v1.4.5 — October 30, 2026

Together, these releases represent another step toward more reliable, locally evaluated, evidence-driven repository analysis.

What is RepoDNA?

RepoDNA helps developers analyze software repositories and understand how their codebases work.

Its existing capabilities include:

• Repository architecture and dependency analysis
• Git history and code archaeology
• Code quality and complexity signals
• Security, testing, and documentation signals
• Interactive analysis and Project DNA reports
• CLI, desktop, browser, and CI workflows
• Optional AI-assisted explanations

The central principle is simple: analysis should be supported by inspectable evidence.

v1.4.4: Introducing Native Evidence Intelligence

One of the most important additions in v1.4.4 is the optional native explanation provider.

Instead of requiring an external language-model API for every explanation, RepoDNA now supports an offline, keyless, built-in evidence digest.

This capability is extractive, meaning it works with evidence obtained from repository analysis rather than generating unrestricted natural-language responses.

The release also introduces a small CPU-only evidence-relevance ranker.

Developers can prepare labeled examples and train a local ranking checkpoint without downloading third-party model weights.

The trained ranker can optionally reorder the evidence selected for native explanations.

Better evaluation

v1.4.4 also introduced retrieval-quality measurements, including:

• Mean Reciprocal Rank (MRR)
• Precision@k
• Recall@k
• NDCG@k

The system compares trained ranking results against the original evidence ordering using repository-disjoint holdout data.

This gives developers a more useful way to evaluate whether local ranking is improving evidence retrieval.

Better findings review

The Findings interface now supports confidence-based filtering and expanded export capabilities.

Filtered findings can be exported in CSV, JSON, or Markdown formats.

These outputs help maintainers inspect findings, review supporting evidence, and incorporate the results into their own development workflows.

Additional improvements strengthen browser import safety, local API validation, package-release integrity, and reliability.

v1.4.5: Reproducibility and Evaluation

v1.4.5 builds on the native evidence-ranking foundation established in v1.4.4.

The focus of this release is not simply adding more ranking features.

It is making ranking behavior easier to reproduce, evaluate, and trust.

1. Deterministic local training

Training is designed to produce consistent outputs even when input dataset rows are shuffled.

The release strengthens validation of evidence identities and repository boundaries.

It also rejects unsafe or ambiguous training data, helping prevent misleading evaluation results.

2. Calibration diagnostics

v1.4.5 adds evaluation diagnostics including:

Brier score: Measures error in relevance probability estimates.

Binary log loss: Evaluates probability predictions against labeled relevance outcomes.

Calibration bands: Help examine whether estimated scores align with observed relevance frequencies.

These evaluate evidence-label relevance, not the factual accuracy of a generated answer.

3. Answerability threshold evaluation

The release introduces offline analysis of hypothetical answerability thresholds.

This helps explore how different thresholds would affect incorrect acceptance and missed answerable questions.

Importantly, these measurements are diagnostic. They do not automatically enable runtime abstention.

4. Stronger regression coverage

Additional tests cover training CLI behavior, checkpoint generation, deterministic output, file overwrite protection, and repository leakage prevention.

The objective is to make the native evidence-ranking system more predictable across supported development environments.

Why Local-First Matters

RepoDNA is designed around a local-first workflow.

Core repository analysis can be performed without uploading source code to a hosted analysis service.

The native extractive explanation capability does not require an external AI API key.

For developers working with internal projects, large codebases, or offline environments, those properties are particularly useful.

And because RepoDNA is open-source, developers can inspect the implementation and evaluate the results themselves.

How to Try RepoDNA

Option 1: Browser

Try the web application:

https://sanskarin.github.io/RepoDNA/

Option 2: Install the CLI

npm install --global @sanskarin/repodna

Then analyze a repository:

repodna analyze .

Option 3: Desktop

Download the latest release for Windows, Linux, or macOS from GitHub.

https://github.com/sanskarIN/RepoDNA/releases/tag/v1.4.5

What's Next?

These releases strengthen the groundwork for continued development of native repository intelligence.

The focus remains on reproducible analysis, evidence quality, useful developer workflows, and transparent limitations.

RepoDNA is not trying to replace software engineers' judgment.

It is being built to help developers make better-informed decisions about their codebases.

Explore and Contribute

GitHub: https://github.com/sanskarIN/RepoDNA

Latest stable release: https://github.com/sanskarIN/RepoDNA/releases/tag/v1.4.5

Web application: https://sanskarin.github.io/RepoDNA/

More projects: https://sanskarin.github.io/

If you're interested in developer tooling, repository analysis, Rust, or local-first software, I'd welcome your feedback and contributions.

RepoDNA — Understand your codebase. See its DNA.