Platform
Everything a reply needs before it goes out with your name on it
The landing page shows four things. This is all of it — the retrieval layer, the guardrails that decide what may be said, the analytics that explain a bad answer, and the controls a security team will ask about before any of it touches a customer.
01
Retrieval
Where the answer comes from, and what happens when it is not in there.
Permission-aware indexing
Halyard inherits the access rules each connected system already enforces. A passage nobody in the requesting queue is allowed to read is never retrieved, so a citation can never leak a document by quoting it.
Hybrid retrieval
Semantic search over passages, combined with exact matching on the identifiers customers actually quote — order numbers, plan names, error codes. Embeddings alone reliably miss the one string that mattered.
Freshness that follows the source
Edits to a help article or a release note re-index within minutes, and every answer records the revision it read. An answer that was right in March is legible as an answer that was right in March.
Conflict surfacing
When two of your own documents disagree, Halyard says so instead of picking one. The contradiction goes to a person, which is usually the fastest way anyone finds out their documentation drifted.
02
Guardrails
The part that decides whether a draft is allowed to send at all.
Grounding floor
Every factual sentence scores against the passages retrieved for it. Set the floor per queue; anything under it never sends on its own, however fluent it reads.
Claims you never make
Write the rules in plain language — no commitments on delivery dates, no legal interpretation, no pricing outside the published table — and they compile into checks that run on every draft before a human reads it.
Review without the fatigue
Only drafts that breach a rule or fall below the floor enter the queue. Reviewers see the failing sentence and its sources, not the whole conversation to re-read from the top.
Refusal logging
Everything Halyard declined to answer is recorded next to what it did answer. This is where your documentation has a hole, and it is a better backlog than any content audit.
03
Answer analytics
Explaining a wrong answer, not just counting the right ones.
Sentence-level citations
Each claim links to the passage that supports it rather than to the document that contains it. This is the artefact that gets an assistant past a legal review.
Full context capture
The retrieved passages, the resolved prompt, the model and the reviewer's edit, stored together per answer. The only record that makes a bad reply reproducible.
Cost per resolution
Token spend attributed to the conversation it closed rather than to the month it happened in, so an expensive queue is traceable to the retrieval step making it expensive.
Replay against today's docs
Re-run any past question against the current index and diff what would be said now. The fastest check that a documentation change actually fixed the answer.
04
Security & compliance
What the review board asks for, answered before they ask.
No training on your content
Your documents and your conversations are never used to train a model, by us or by any provider we route to. Contractual, not just configured.
SSO, SCIM and RBAC
SAML and OIDC single sign-on, directory-synced provisioning, and role-based access down to individual queues and channels.
Data residency
Pin indexing, inference and storage to a region, or route generation to a model running inside your own VPC while the workspace stays hosted.
Configurable retention
Retention windows per workspace with a hard floor, explicit legal hold that suspends deletion, and a surviving record of anything deleted on schedule.
Point it at what you already wrote.
Connect one source, run a week in draft mode, and read the diff against your team's replies. If it isn't better, you've lost an afternoon.
Free for 14 days · No card required