Skip to content

docs: add the Alice guardrail page - #1056

Open
seanyasno-af wants to merge 1 commit into
Portkey-AI:mainfrom
seanyasno-af:feat/alice-guardrail
Open

docs: add the Alice guardrail page#1056
seanyasno-af wants to merge 1 commit into
Portkey-AI:mainfrom
seanyasno-af:feat/alice-guardrail

Conversation

@seanyasno-af

Copy link
Copy Markdown

Documents alice.evaluate, added in Portkey-AI/gateway#1791. Companion to that PR — happy to hold this until the plugin's shape is agreed.

What this adds

  • integrations/guardrails/alice.mdx
  • the page's entry in docs.json (Integrations → Guardrails, after akto)
  • a card in the partner list in product/guardrails/list-of-guardrail-checks.mdx

The page follows the shape of the recent provider pages (CrowdStrike AIDR, F5): credentials, adding the check, the parameter table, supported hooks, and the config + request samples.

Three sections go beyond that template

They are the parts an operator is most likely to get wrong, so they get their own headings rather than a table row.

Naming the application. Alice configures policies per application, and a project usually holds several — one API key fronts all of them, so the application has to be a parameter on the check rather than part of the credential. A guardrail naming none fails its check rather than being measured against a guess, which would otherwise be a confusing first-run failure. The section covers where the id comes from, what to do when different traffic needs different policies, and why it is deliberately not read from x-portkey-metadata (the caller writes that, so it would let a caller point their own traffic at laxer policies).

What is sent. The plugin forwards the hook context unchanged, so what is worth evaluating stays Alice's decision rather than something users upgrade their Gateway for. Credentials are the single exception: headers, credentials, api_key, apiKey, authorization and cookie are dropped at any nesting depth and never leave the Gateway.

When Alice cannot be reached. Every failure — a 4xx included — is reported as a check error rather than as a pass, which makes failOnError the single place that decides whether unscreened traffic proceeds. Both settings are spelled out, with a note on when to pick which.

The streaming caveat is stated honestly too: blocking works on a streamed response, masking does not.

Checks run

  • npx mint broken-links — the new page and the edited list report none. (The run does surface pre-existing broken links elsewhere, mostly under virtual_key_old/; none are touched here.)
  • docs.json parses.

One choice worth flagging

The card omits img, following the existing Akto card, rather than adding a logo asset to the repo. Glad to add one if you'd prefer the grid stay uniform — just say where the file should live.

@seanyasno-af seanyasno-af changed the title docs: add the Alice by ActiveFence guardrail page docs: add the Alice guardrail page Sep 3, 2026
Documents `alice.evaluate`, added in Portkey-AI/gateway#1791.

Adds `integrations/guardrails/alice.mdx`, registers it in `docs.json`, and adds a
card to the partner list in `product/guardrails/list-of-guardrail-checks.mdx`.

The page follows the shape of the recent provider pages (CrowdStrike AIDR, F5):
credentials, adding the check, parameters, supported hooks, config and request
samples.

Three sections go beyond that template, because they are the parts an operator is
most likely to get wrong:

**Naming the application.** Alice configures policies per application and a
project usually holds several, so one API key fronts all of them and the
application has to be a parameter on the check rather than a credential. A
guardrail naming none fails its check rather than being measured against a guess,
which would otherwise be a confusing first-run failure. The section covers where
the id comes from, what to do when different traffic needs different policies,
and why it is deliberately not read from request metadata.

**What is sent.** The plugin forwards the hook context unchanged so that what is
worth evaluating stays Alice's decision, with credentials as the single
exception: `headers`, `credentials`, `api_key`, `apiKey`, `authorization` and
`cookie` are dropped at any nesting depth and never leave the Gateway.

**When Alice cannot be reached.** Every failure, a 4xx included, is reported as a
check error rather than as a pass, which makes `failOnError` the single place
that decides whether unscreened traffic proceeds. Both settings are spelled out.

The streaming caveat is stated honestly too: blocking works on a streamed
response, masking does not.

Verified with `npx mint broken-links` — the new page and the edited list report
none. The card omits `img`, following the Akto card, rather than adding an asset.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant