Auto Secure LoginPlatform
Live Shared infrastructure Bilingual EN / ES

Answers that cite their sources — and stay silent when the evidence isn't there.

ASL Knowledge Agent is a collection-independent, citation-first help and knowledge engine shared across ASL products. It retrieves the most relevant reviewed passages, returns an answer grounded in cited evidence, and abstains when authorized evidence is insufficient — rather than inventing one. Before it retrieves or generates anything, it enforces collection, tenant, language, audience, visibility, and lifecycle boundaries. Its public face is Policy Lens.

cite-or-silentabstention protected by CI
27/27unit + 32/32 bilingual golden
3governed collections
0identity in the request body

The problem

A confident wrong answer is worse than no answer.

In-product help that hallucinates is a liability — especially in regulated domains like therapy or compliance, where an invented "fact" can cause real harm. And a shared help engine has to respect that one tenant must never see another's content, or a restricted collection's answers.

Knowledge Agent is built to refuse. It answers only from cited, authorized evidence, abstains when there isn't enough, and enforces entitlement boundaries before it ever retrieves.

Five honest verdicts

  • direct — grounded, cited answer
  • conditional — answer with stated caveats
  • potential_conflict — sources disagree
  • supported_across_sources — corroborated
  • insufficient — abstains rather than invents

Capabilities

Grounded retrieval with real access control.

📑

Citation-first answers

Every answer is grounded in cited evidence, with explicit abstention when the evidence is insufficient.

🔎

Lexical + concept ranking

BM25-style ranking with concept coverage, phrase, heading, proximity, and identifier signals across 9 query-intent classes.

Fully bilingual

EN/ES inference with accent-normalization and stemming — not just a translated UI.

Conflict detection

Flags when sources disagree instead of silently picking one, so the reader knows the ground is contested.

🎯

Diverse evidence selection

Diversity caps and marginal-utility scoring pick a well-rounded evidence set, not five copies of the same passage.

🔐

Per-consumer signed requests

An HMAC signature binds client, subject, org, tenant, roles, collection, language, purpose, path, body-hash, timestamp, and a one-time nonce (60s window).

Two-sided entitlement

Access is the intersection of consumer ∩ collection — restricted collections fail closed.

🎟

Signed run ownership

A signed 24h run-ownership token binds feedback and escalation to the run that produced the answer.

📋

Versioned manifest

A machine-readable, versioned collection manifest with live source-health checking.

27/27 + 32/32unit + bilingual golden tests
adversarialforged sig, replay, wrong tenant — all rejected
no raw questionsnever written to audit
40-charexact commit provenance
The engine behind trustworthy help. Knowledge Agent powers three governed collections today — Policy Lens help (public), ShopFit help (18 articles), and Therapy help (82 articles / 278 sections, restricted) — with per-consumer signed assertions that replaced an older shared-token model. Abstention is protected by CI, so "cite or stay silent" can't quietly regress.

Help your users can trust.

Source-backed answers, honest abstention, and real entitlement boundaries — one shared engine every product can safely embed.