AI in Aliquora — Technical Overview

AI surfaces in the product

Feature	Who uses it	What it does
AI Lab Assistant	Lab staff (plan-gated)	One chat that answers QC/LIMS questions, searches your lab data with citations, and — for administrators — proposes lab actions for approval.
Configure by Chat	Administrators	Describe a configuration change in plain English; the assistant proposes a structured before/after change set that an admin approves before anything is applied.
Client portal Q&A	Your clients (per-link)	Sample-status questions answered strictly within the disclosure level of the shared portal link — never internal notes, specs, or data beyond what the link already shows.

Architecture of the AI Lab Assistant

The assistant is a single chat backed by a large language model with tool calling. For each message, the model decides — within a bounded number of rounds — whether to simply answer, to search your lab data, or to propose actions. Which tools the model is even offered is decided server-side per request:

• The lab-data search tool is only offered when your organization has enabled AI data access (see below). Even if the model attempts a search, the permission is re-checked at execution time.
• Action tools are only offered to administrator roles, and the role is re-checked server-side when an action is proposed.
• Responses stream to the browser over server-sent events, including status updates and source citations.

Lab-data access is opt-in

By default the assistant answers from general laboratory and QC knowledge only. Access to your lab data is a per-organization switch in Settings → AI & Privacy, controlled by your administrators. Until it is enabled, no lab records are included in AI requests. The switch is enforced on the server on every request — not in the browser.

How lab-data search works (retrieval with citations)

When data access is enabled and a question calls for it, the assistant combines two org-scoped sources and cites what it used:

• A live snapshot of current lab state — sample counts by status, tests awaiting verification, out-of-spec flags — computed at question time from organization-filtered queries.
• A semantic search index of your lab content. Content is split into overlapping passages, embedded, and stored per organization; the most relevant passages (top 8) are retrieved for the question and shown as citations under the answer.

What is indexed

Source	Indexed content
Documents	Title, description, and extracted text of uploaded documents (SOPs, methods, certificates).
Notebook entries	Entry title and content.
Samples	Identifiers, client/project, material, status, and test summary.
Test definitions	Test names, codes, methods, units, and specification limits.
Material types	Names and descriptions from your sample catalog.

Embeddings are computed on Aliquora's own infrastructure by a local embedding model — indexed lab content is never sent to a third-party service to build or maintain the search index. Administrators can rebuild their organization's index at any time from the assistant page; rebuilding deletes and re-creates only their own organization's index.

Tenant isolation

Aliquora is multi-tenant, and the AI features enforce the same hard organization boundary as the rest of the LIMS — one tenant's data can never appear in another tenant's answers, citations, or proposals:

Layer	Control
Identity	The organization is always derived from the authenticated server-side session — never from anything the client sends. Requests without an organization are rejected.
Search index	Every indexed passage is stored with its organization ID, and every retrieval query is hard-filtered to the requesting organization. There is no unscoped query path.
Live lab context	Sample counts, pending verifications, and out-of-spec summaries are built exclusively from organization-filtered queries through the same data-access layer as the rest of the LIMS.
Action proposals	Proposals are stored with the proposing organization and re-checked at apply time — a proposal created in one tenant returns "not found" in any other.
Rebuild index	Re-indexing deletes and rebuilds only the requesting organization's rows. It cannot read or affect another tenant's index.

Actions are proposed, never executed by the AI

When an administrator asks the assistant to do something — create samples, enter results, verify tests, generate a COA — the AI cannot touch your records directly. It works through a fixed registry of structured operations (it can never emit free-form code or SQL), and every request goes through a two-phase flow:

1. Propose. The assistant's suggested operations are validated and stored as a pending proposal. Nothing changes in the LIMS. The UI shows a preview of exactly what would happen.
2. Approve & apply. A human administrator approves the proposal. The server then reloads it, re-validates every operation against current data, re-checks plan features and permissions, and atomically claims the proposal so it can only ever be applied once — then executes.

• The client never submits the change itself — only a reference to the stored proposal. The AI's output is always treated as untrusted and re-checked server-side.
• All business rules still apply: QC release gates, verification locking, e-signature requirements, and plan limits are enforced at apply time exactly as if a person performed the action in the UI.
• Every applied change is written to the audit trail, attributed to the approving user.

What leaves your environment

• Chat requests — your message, recent conversation turns, and (only when data access is enabled and the question requires it) the retrieved lab-data passages are processed by a large-language-model provider to generate the response.
• Nothing else. The search index, embeddings, live snapshots, proposals, and audit records all stay inside Aliquora's infrastructure. Disabling AI data access stops lab records from being included in any AI request.
• Client-portal Q&A is additionally capped to the disclosure level of the specific shared link — it can never reveal result values, specifications, or internal notes that the link itself does not already show.

Availability, limits, and abuse protection

• The AI Assistant is a plan feature; availability depends on your subscription tier.
• All AI endpoints sit behind authentication, role checks, and per-organization usage guards that bound spend and prevent abuse.
• AI features degrade gracefully: if the assistant is unavailable, the rest of the LIMS is unaffected — no lab workflow depends on AI.

Related documents

• Security & compliance overview — the broader security posture this document sits inside.
• 21 CFR Part 11 & ALCOA+ capability matrix — audit trail, e-signature, and data-integrity controls.
• API guide & reference — programmatic access with scoped API keys.