Skip to content
borderv0.1.0

Documentation

Two functions, and a file a reviewer can read

There is no client to construct and no gateway to run. You call two functions, decide what to do with what they return, and archive the record.

your model callmodelscan_inputscan_output

Generated from docs/ in the library repository, which is where every statement about how the API behaves is written. Each page is also readable as markdown.

Getting started

  • IntroductionWhat border is, what it returns, and what it deliberately does not do.
  • QuickstartInstall it, write a policy, scan a turn in both directions, and read the record.

Concepts

  • The two functionsWhy the API is two calls, what a Decision contains, and which text to use afterwards.
  • The policy fileHow a policy is written, how it resolves, and why it is data rather than code.
  • TiersWhat a tier decides, what it does not, and what a full scan costs.
  • The evidence recordWhat a record contains, what it never contains, and how to verify one.
  • Offline and determinismWhat "no network at scan time" means exactly, and what is asserted rather than promised.

Reference

  • TypesDecision, Finding, EvidenceRecord and the literals they use.
  • Policy schemaEvery key a policy document accepts, with its default and its constraint.
  • Language coverageThe 26 target languages, what is actually covered, and where it fails.
  • Measured performanceGenerated from the library's own benchmark run.

Guides

  • Detectors: what exists, what does notIf you are describing this library anywhere public, describe it from here. ## The numbers
  • Migrating from llm-guard`flowx_border.adapters.llm_guard_compat` keeps llm-guard's `scan_prompt` and `scan_output` signatures and its tuple return shape, so a migration starts as an import change. The tab
  • Open items before flowx-border goes publicEvery number here comes from the repository or from a measurement recorded in it, and the ones that are guesses say so.
  • Porting the Guardrails Hub validators[guardrails-ai/guardrails-hub-monorepo](https://github.com/guardrails-ai/guardrails-hub-monorepo) ships 65 validators. This is what happened to each of them. Every decision below i
  • Which detectors should complete the setDeliberately written after the Guardrails Hub port rather than folded into it, so that the question is "what is missing" rather than "what did we not get round to porting".