browse_portfolio
ChatGPTFilter, total and group the partner portfolio's operations with the parameters of the site's project browser: partner, agency, kind, instrument, stage, sector, dept, start_from, start_to, close_from, close_to, min, max, budget, type, extended, restructured, cerc, q, fy, family, group, sort, page, per, year_basis. Totals count funding partners and state how many operations report each figure.
fetch
ChatGPTRetrieve a resource by id from search. Returns title, human-readable text, url, and metadata (source attribution, license, freshness, citation) for display or full read. For a document, pass 'query' to get the passages of the body most relevant to your question; 'max_chars' bounds the returned text. For an indicator, 'component', 'source_id', and 'measure' return a specific slice of the series (e.g. one component, or one source) instead of the headline. The source attribution and citation in the response can be used to ground a brief or note.
fetch_many
ChatGPTRetrieve several resources at once by passing a list of ids (bounded). Prefer this over repeated single fetch calls when you need more than one resource. Returns the same shape as fetch for each resolved id.
fetch_map_layer
ChatGPTReturn the GeoJSON CONTENT of a map layer for rendering or spatial analysis. Use this when you need to display or analyse the actual layer data and cannot fetch the URL directly (e.g. in a Python sandbox without external DNS access). Large layers are automatically sampled and coordinate-rounded to a safe payload size; the response includes a 'note' describing any sampling and the full-fidelity 'download_url' for the complete layer. Optional 'bbox' ([min_lon, min_lat, max_lon, max_lat]) pre-filters features to an area of interest before the feature cap is applied, giving higher spatial detail for a sub-region. Optional 'max_features' (default 6000) sets the cap. This tool complements 'fetch', which returns metadata and a URL without content.
fiscal_catalog
ChatGPTList the structured fiscal datasets: budget and execution lines extracted from the budget books and unified across years, with their fiscal years, stages, versions, components, currency, quality notes and a summary of the reconciliation checks. Call this first, then fiscal_query.
fiscal_checks
ChatGPTReconciliation checks that came with a fiscal dataset: a printed total against the sum of its parts, with the difference and a note where the source book is known to disagree with itself. Filter by dataset, status (pass, fail, not_testable), scope (a department name or code) and version.
fiscal_query
ChatGPTQuery budget and execution lines. Filters: dataset, entity_code, entity_contains, programme_code, project_code, object_code, object_contains, economic_class, sector_code, component, grain, stage, measure, version_id, fiscal_year, fy_end, fy_from, fy_to, in_total (true, false or any), label_contains, and dims.<name> for source columns with no core field. 'group_by' sums by up to five fields (for example entity_code and fy_end). Every row carries its source page, table and row, and amounts are in the dataset's currency as printed. Add amounts only inside one component, grain, stage, measure and version: the response lists any cut a sum mixes in 'warnings'. 'limit' is at most 200.
get_passages
ChatGPTRetrieve the document passages most relevant to a query, drawn from the full text across the public library (or within one 'doc_id'). Each passage carries its source citation and link. Use this as the primary tool to gather sourced excerpts to ground or draft a brief or note: the returned passages are cited and ready to weave into prose. CITE BY NUMBER, ONE NUMBER PER DOCUMENT: mark each claim with a bracketed number, reuse that same number for every claim drawn from the same document, and list the sources once at the end of the answer, numbered, each with its 'citation_key' and url. Do not repeat a document's full title inline.
list_documents
ChatGPTEnumerate the public document library with paging and facet counts (by year and document type). Optional 'filters' (organization, doc_type, year_from, year_to, topic, language, query). Use it to sweep every source on a theme.
search
ChatGPTSearch Haiti's public knowledge and data library: curated economic indicators, map layers, documents, World Bank projects, and news. Returns a ranked list with id, title, and url; pass the id to fetch or get_passages. Use this to scan what exists on a topic before grounding with get_passages or opening a full resource with fetch. Prefer a broad query with no filters first. Optional 'filters': 'types' (any of: doc, indicator, map_layer, project, development), organization, doc_type, year_from, year_to, topic, language; plus 'offset' (paging) narrow and page the results.
server_info
ChatGPTReport server build and data-access contract version, transport, and access model. Use as a health check.
browse_portfolio
ClaudeFilter, total and group the partner portfolio's operations with the parameters of the site's project browser: partner, agency, kind, instrument, stage, sector, dept, start_from, start_to, close_from, close_to, min, max, budget, type, extended, restructured, cerc, q, fy, family, group, sort, page, per, year_basis. Totals count funding partners and state how many operations report each figure.
fetch
ClaudeRetrieve a resource by id from search. Returns title, human-readable text, url, and metadata (source attribution, license, freshness, citation) for display or full read. For a document, pass 'query' to get the passages of the body most relevant to your question; 'max_chars' bounds the returned text. For an indicator, 'component', 'source_id', and 'measure' return a specific slice of the series (e.g. one component, or one source) instead of the headline. The source attribution and citation in the response can be used to ground a brief or note.
fetch_many
ClaudeRetrieve several resources at once by passing a list of ids (bounded). Prefer this over repeated single fetch calls when you need more than one resource. Returns the same shape as fetch for each resolved id.
fetch_map_layer
ClaudeReturn the GeoJSON CONTENT of a map layer for rendering or spatial analysis. Use this when you need to display or analyse the actual layer data and cannot fetch the URL directly (e.g. in a Python sandbox without external DNS access). Large layers are automatically sampled and coordinate-rounded to a safe payload size; the response includes a 'note' describing any sampling and the full-fidelity 'download_url' for the complete layer. Optional 'bbox' ([min_lon, min_lat, max_lon, max_lat]) pre-filters features to an area of interest before the feature cap is applied, giving higher spatial detail for a sub-region. Optional 'max_features' (default 6000) sets the cap. This tool complements 'fetch', which returns metadata and a URL without content.
fiscal_catalog
ClaudeList the structured fiscal datasets: budget and execution lines extracted from the budget books and unified across years, with their fiscal years, stages, versions, components, currency, quality notes and a summary of the reconciliation checks. Call this first, then fiscal_query.
fiscal_checks
ClaudeReconciliation checks that came with a fiscal dataset: a printed total against the sum of its parts, with the difference and a note where the source book is known to disagree with itself. Filter by dataset, status (pass, within_tolerance, fail, not_testable), scope (a department name or code) and version.
fiscal_query
ClaudeQuery budget and execution lines. Filters: dataset, entity_code, entity_contains, programme_code, project_code, object_code, object_contains, economic_class, sector_code, component, grain, stage, measure, version_id, fiscal_year, fy_end, fy_from, fy_to, in_total (true, false or any), label_contains, and dims.<name> for source columns with no core field. 'group_by' sums by up to five fields (for example entity_code and fy_end). Every row carries its source page, table and row, and amounts are in the dataset's currency as printed. Add amounts only inside one component, grain, stage, measure and version: the response lists any cut a sum mixes in 'warnings'. 'limit' is at most 200.
get_passages
ClaudeRetrieve the document passages most relevant to a query, drawn from the full text across the public library (or within one 'doc_id'). Each passage carries its source citation and link. Use this as the primary tool to gather sourced excerpts to ground or draft a brief or note: the returned passages are cited and ready to weave into prose.
list_documents
ClaudeEnumerate the public document library with paging and facet counts (by year and document type). Optional 'filters' (organization, doc_type, year_from, year_to, topic, language, query). Use it to sweep every source on a theme.
search
ClaudeSearch Haiti's public knowledge and data library: curated economic indicators, map layers, documents, World Bank projects, and news. Returns a ranked list with id, title, and url; pass the id to fetch or get_passages. Use this to scan what exists on a topic before grounding with get_passages or opening a full resource with fetch. Prefer a broad query with no filters first. Optional 'filters': 'types' (any of: doc, indicator, map_layer, project, development), organization, doc_type, year_from, year_to, topic, language; plus 'offset' (paging) narrow and page the results.
server_info
ClaudeReport server build and data-access contract version, transport, and access model. Use as a health check.