When a document is uploaded into DNXT Publisher Suite, the platform's ingestion engine parses it at the structural level — extracting headings, section identifiers, table of contents entries, footnotes, and existing hyperlink targets. For Word documents, tracked changes are resolved and the clean version is indexed. This isn't a simple text dump — the parser understands eCTD structure (module hierarchy, section numbering, leaf-level relationships) so the AI has regulatory context from the first byte, not just raw text.

Each document section — not the entire document as a monolithic block — is encoded into a high-dimensional vector embedding using the configured AI provider's embedding model (e.g., text-embedding-ada-002 for OpenAI, or equivalent via Azure OpenAI or Google Vertex AI). These embeddings capture semantic meaning: a section discussing "pharmacokinetic bioavailability in fasted subjects" will be close in vector space to a clinical study reporting that data, even if the exact terminology differs. The vector index is stored per-dossier and updated incrementally as new documents are added — a full re-index is never required for minor additions.

When the Auto Linker or Bookmark Advisor is invoked, the system uses Retrieval Augmented Generation: it takes the target section's embedding and performs a cosine similarity search across the entire dossier's vector index to surface the most semantically relevant candidates. This is fundamentally different from keyword search — it finds relationships that a human might recognize immediately ("this clinical summary is clearly referencing that study report") but that rule-based systems miss entirely because the section doesn't literally quote the target document's title. The top-N candidates are retrieved with their similarity scores, filtered against eCTD linking conventions, and passed to the language model for final reasoning.

The retrieved candidates are passed to the configured language model (GPT-4, Claude 3, or equivalent) with a structured prompt that includes regulatory context: which module the source section belongs to, what type of cross-reference is appropriate per ICH M4 guidance, and any tenant-specific linking rules the system has been trained on. The model reasons over the candidates and returns a ranked list of suggested links with natural-language explanations. A confidence score is computed from a combination of vector similarity, model certainty, and a secondary validation pass that checks whether the suggested target actually exists as a resolvable document in the current dossier structure — preventing ghost links before they're ever accepted.

Suggestions appear in the AI Suggestions panel inside DNXT's publishing interface — visible alongside the document being worked on, not in a separate tool requiring context switching. Each suggestion shows the source section, target document or anchor, confidence score, and a one-line explanation of why the link was proposed. Publishers can accept a suggestion in a single click (the hyperlink is inserted automatically with the correct bookmark target resolved), dismiss it, or flag it for senior review. Bulk accept is available for suggestions above a configurable confidence threshold — for high-volume operations, this allows teams to process 80–90% of suggestions automatically while reviewing only the edge cases.

Whether a suggestion is accepted, dismissed, or modified before acceptance, the action is written to DNXT's immutable audit log in real time. The log entry captures: timestamp (UTC), user identity (authenticated via SSO/MFA), action taken, source document and section, target document and anchor, AI model used, model version, confidence score at time of decision, and a hash of the AI response payload. This means a regulatory inspector can trace every hyperlink in a submission back to the exact decision that placed it — who, when, what model, what confidence, what was accepted versus what was changed. For the first time, AI automation in regulatory publishing is fully defensible under GxP requirements.

Accepted and dismissed suggestions are fed back into DNXT's tenant-specific fine-tuning layer. When publishers consistently dismiss a certain type of suggestion (e.g., linking pharmacology summaries to toxicology studies they consider unrelated), the system learns that preference and adjusts future suggestions accordingly — without requiring any manual rule configuration. This learning is scoped per tenant so it never crosses organizational boundaries, and all feedback-derived model adjustments are themselves logged in the audit trail. Over a 6-month deployment, teams typically see suggestion relevance improve by 15–20% as the system adapts to their specific dossier conventions and therapeutic area.