How it thinks, stage by stage.

Every question is cleaned of typographic characters that can break downstream processing — smart quotes, em dashes, ellipsis characters. This stage exists because a real crew member's question once crashed the system with a single special character.

"Fire drill procedures while transiting a TSS" becomes two separate searches — one for fire drills, one for traffic separation — then results are merged. Compound maritime questions are the norm, not the exception.

"Night transit" expands to include "hours of darkness, sunset, sunrise, navigation lights." Over twenty curated maritime vocabulary mappings ensure the right documents are found regardless of how the question is phrased.

Two independent search systems run in parallel. Semantic search catches conceptual matches. Keyword search catches exact terms. Reciprocal Rank Fusion automatically promotes documents that both systems agree on.

The system examines its own top results and uses their mathematical representations to expand the search — catching documents too different in vocabulary for the initial search. A purely mathematical operation completed in milliseconds.

Twenty candidates pass through a specialised neural model that examines the question and each document together — not independently. This is the single most important quality step, producing calibrated confidence scores for every candidate.

Every document carries its authority tier — vessel-specific, company SMS, fleet-type, regulation, or reference. A tier multiplier nudges rerank scores so your procedures lead, regulations corroborate. Near-ties break in favour of higher authority.

Scores are normalised to a 0-to-1 scale and filtered. Below the hard threshold: discarded. Below the soft threshold: the language model is instructed to flag uncertainty. This prevents overconfident answers from marginally-relevant documents.

When retrieval returns insufficient results, the system reformulates the question into formal regulatory vocabulary and retries. Results are re-scored against the original question to preserve intent. When this happens, the crew is told — the confidence badge reflects that we had to work harder.

When a retrieved chunk is part of a longer procedure, the parent section is pulled in automatically — step 7 of a 9-step man-overboard procedure never arrives in the model's context without the other eight. Source diversity is enforced.

Research shows language models pay most attention to the start and end of their input. The system reorders passages so the most relevant content occupies these high-attention positions, with weaker material in the middle.

A compact language model generates answers with inline numeric citations, resolved after generation to document name, section, and page number. Domain completeness rules enforce that safety-critical procedures always return every mandatory step — no summaries.

Every bracketed source number in the answer is checked against the chunks that were actually retrieved. If the model invents "[47]" when only [1]–[5] exist, we catch it — the crew sees a warning instead of a confident fabrication.

A second, short model call checks whether the retrieved context actually contained enough information to answer the question — independent of the answer the model wrote. Sufficient, partial, or insufficient. The verdict corrects the keyword-based refusal detector in both directions.

Every answer gets a composite confidence tier derived from reranker scores, source diversity, citation verification rate, and whether the pipeline had to self-correct. Refusal detection is length-aware. Verified answers are cached semantically with re-verification on every cache hit.