OptionaladditionalInstructions?: string | nullOptionalauditQuestionId?: AuditQuestionIdOptionalenableClassifyRetrievalMode?: booleanOptionalenableDocumentCatalogue?: booleanOptionalenableMetadataAwareGate?: booleanWhen set, enrich the document listing the LLM picks from with each
chart's circle_metadata (level-of-care / program-stay window) so the
model selects with clinical context in view. ENRICHMENT only — never a
hard narrow: the candidate set is unchanged, so it can't starve
retrieval of the document holding the answer.
OptionalenableTimeFilteringInChartReview?: booleanOptionalexpandedSteps?: stringOptionalpatientCharts?: {Optional pre-fetched patient charts. When provided, skip the per-question DB fetch and filter in memory by the question's time filter. Caller is responsible for passing the full unfiltered set (or a superset covering the widest expected time window).
OptionalquestionType?: "pass_fail" | "summary"OptionalreferenceInstant?: DateThe run's shared reference instant; anchors the time filter.
OptionalretrievalFloor?: DateEarliest day any criterion of a compound question needs. Only ever widens the cutoff.
OptionalreviewCriteriaInstructions?: string | nullTemplate review criteria — appended to the prompt so the LLM selects documents that matter for the rubric.
OptionalsearchHints?: string[] | nullOptional title/type synonyms from question decomposition.
OptionaltimeZone?: string | nullFacility IANA zone for rendering document dates; resolved (with fallback) internally.
OptionaltreatmentEpisodeId?: TreatmentEpisodeIdWhen set, the chart fetch is scoped to this treatment episode — the
candidate set the LLM picks from contains ONLY that episode's charts, so
cross-episode documents can never be considered. Re-applied in memory to a
provided patientCharts set as well.
Render the candidates as a CATALOGUE — grouped by type, then by name, each name's documents indexed by week with its most recent named outright — instead of one line per document.
A flat listing repeats every document's type and tags on every row, so a few hundred documents is mostly duplicated strings and the model still does its own counting. Off by default so the two can be compared on the same questions.