This page is written for the person who has to approve the deployment: an IT head, a security architect, a QA systems owner. It describes what gets installed, where data sits, and what happens when a user asks a question.
Documents are collected from your existing stores — file shares, SharePoint, DMS, object storage, or direct upload. Text is extracted, scanned pages are OCR'd, and structure (headings, tables, document metadata, version and effective date) is preserved. Nothing is copied outside your environment at any point.
Extracted content is chunked and embedded into a vector index that lives on your infrastructure alongside a metadata store. Document-level permissions are attached at index time, not applied as a filter afterwards — so a user's access boundary is enforced at retrieval, not at display.
Open-weight language models run on your own GPU. Queries are processed locally. No API call leaves the network. In an air-gapped installation there is no network route out at all.
A query is resolved against only the documents that user is entitled to see. Retrieval is multi-step where the question requires it — comparing across documents, following references, filtering by version or date — but every step draws on your indexed content, never on the model's own recall.
The answer is assembled from retrieved passages with a citation to each source document, version, and page. If the corpus does not support an answer, the system says so rather than generating one. Every query is written to the audit log.
Deliberate limits.
FortiqAI does not take actions in your systems, does not approve or close records, and does not make decisions. It retrieves, compares, extracts, and drafts — for human review. This is a design constraint, not a missing feature. In regulated work, unsupervised action is a liability.
Containerised services (Docker/Helm) or a VM appliance: ingestion workers, vector store, metadata database, inference service, application backend, and web UI. Deployed within your network boundary and administered by your team.
Hardware sizing, deployment modes, and the honest answer to "what will this actually take?"
