Skip to main content
Assistant is HASP’s chat surface: a person talks to a model directly, in the browser, at app.usehasp.com/assistant. It is one of the four consumption surfaces, and it rides exactly the same substrate as the others — the same AI Gateway, the same PHI Guard, the same audit chain, the same BAA gate. A chat turn and an API call are the same kind of event to the substrate; only the entry point differs. There is no separate Assistant plan. Every paid tier includes it.

​
Conversations

Work is organised into conversations. Each one holds a transcript, a PHI mode, and an optional persona, and can be pinned, renamed, archived, or exported. Conversations opened in the browser are the same server-owned conversations the API exposes — see Conversations for the programmatic surface. The transcript lives inside the compliance boundary rather than in the browser, which is what lets an org export or review it later without depending on any one person’s session.

​
Compaction

Long conversations can be compacted: earlier turns are summarised so the conversation stays within the model’s context window without starting over. Compaction is a recorded action — the original turns remain in the audit chain even though the active context no longer replays them in full.

​
PHI mode

Every conversation carries a PHI mode deciding whether protected health information may appear in it. The mode is resolved against the org’s policy, so a user cannot select a mode the organization has not enabled. Raising a conversation to allow PHI is one-way. Once a transcript may have contained PHI, HASP keeps treating it as PHI-bearing for retention and disclosure purposes even if the mode is lowered again — the safe assumption is the only correct one after the fact. PHI handling itself is a universal substrate capability gated on an active BAA, not a tier feature. See PHI Guard and Compliance posture.

​
Personas

A conversation can be bound to a persona — a named, versioned instruction layer. Users pick either a specific persona from the org’s library or no persona at all; the organization may also set a default that applies when no explicit choice has been made. Persona authoring, publishing, and promotion are documented under Personas.

​
Attachments

Files can be attached to a conversation and referenced in the discussion. Attachments are virus-scanned on upload and are subject to the same disclosure accounting as any other content read — listing attachment metadata is not a disclosure; reading the bytes is. The same objects are reachable programmatically through the Files API.

​
Model access and credits

Assistant draws on the org’s model allow-list and AI credit allotment like every other surface. A model the org has not enabled is not selectable, and a conversation cannot spend past the org’s budget cap. Members can raise a credit request to an admin when they hit a limit rather than being silently blocked. Relative model cost and which models are premium are covered in Model selection.

​
What is audited

Assistant turns produce the same audit events as any other gateway traffic: the caller, the model, the resolved persona, PHI decisions, and credit consumption. Conversation exports and transcript reads are recorded as disclosures. Everything lands in the tamper-evident chain described in Audit chain and is covered by the same independent verification story — including for Free Evaluation orgs.