charts_by_report
ChatGPTGet all charts from a specific BCA report, ordered by chart index. Use ONLY when the user explicitly named a report_stem (filename stem like "USBS_SI_2026_04_16") AND asked for that report's charts/figures/visuals. Do NOT use to find or summarize a report itself - that needs a report tool, not this one. This returns a known report's charts, nothing else. Input JSON: {"report_stem": "USBS_SI_2026_04_16"} The report_stem is the filename stem (without extension) of the report. Returns chart cards in presentation order.
charts_by_series
ChatGPTFind all BCA charts that use a specific Macrobond series code. Use ONLY when the user explicitly named a Macrobond series code (e.g. "wti", "usunrate"). Do NOT use to look up reports or topics - those need a report tool, not this one. Input JSON: {"series_code": "wti", "limit": 20} Returns chart cards with thumbnails.
charts_get
ChatGPTGet a single BCA chart by its MongoDB ObjectId. Use ONLY when the user explicitly provided a chart_id (ObjectId string). Do NOT use to discover charts by topic - for that, use charts_search and only if the user explicitly asked for a chart/graph/visualization. Input JSON: {"chart_id": "<ObjectId string>"} Returns a single chart card with thumbnail and full metadata.
charts_search
ChatGPTONLY use when the user explicitly asks about charts, graphs, visualizations, Macrobond data series, or a specific chart. Do NOT use for general report queries - even when the topic is quantitative (FX rates, savings rates, yields, inflation, GDP). "Find a report on X", "which reports discuss Y" and "what did X write/expect/argue about Y" are written-report requests, not chart requests - route them to a report tool if you have one. Quantitative vocabulary alone is not a chart signal. Search BCA's Macrobond chart catalog by topic, series label, AI description, or folder path. Returns chart cards with thumbnails, metadata badges, and series labels (UI version). Input JSON: query (required), limit (1-5, default 5; a larger value is reduced to 5), offset (0 or more, default 0 - raise it to page through further matches), and filters: - publication: list of publication names (e.g. ["GeoMacro", "US Investment Strategy"]) - subcategory: list of subcategory folder names - year: {gte, lte} integer year range - created_by: list of analyst emails (e.g. ["katia.shtrevensky@bcaresearch.com"]) - ai_chart_type: list of chart types (line, bar, area, scatter, multi-panel) - series_code: single Macrobond series code for exact match (e.g. "wti") Filters are optional. A query alone is sufficient to search the full catalog. QUOTE A PHRASE TO REQUIRE IT. Double quotes in query turn that phrase into a filter, not a preference: only charts whose title, on-chart text or source note carry those exact words in that order come back, and two quoted phrases must both appear. Use it when the user asks for the words that are literally printed on a chart ('charts that say "own nationality"', 'charts sourced from "Eurobarometer"'). A matched result quotes the line it was found on. Quoting returns nothing when nothing carries the phrase, and that empty answer is correct - say so rather than re-running the query unquoted as if it had failed. Leave the query unquoted for ordinary topic search, which finds charts about a subject whether or not they use your words. ONE SEARCH PER REQUEST. Each call draws its own grid of charts, so a second search in the same turn shows the user charts, then more searching, then other charts. Put every part of the ask into ONE query ("data center margins and rents", not two searches) - the search matches several ideas at once. Then answer with what came back. Do not re-run a reworded or narrower query in the same turn because the page holds only 5 charts or a result looks off-topic: show what you have and offer to look further (the next page, or another angle) on the user's reply. Several searches in one turn are right only when the user asked for several DIFFERENT topics ("charts on oil, and separately on gold"). The response text lists each result with its chart_id and image_url - copy those verbatim when you cite a chart or embed its image. Getting back fewer than limit results is normal: charts whose image could not be rendered are left out, and the response says how many. Do not treat a short list as a broken search. Trigger words: "show me charts", "find charts", "get charts", "canadian housing charts", "charts on X", "macrobond charts", "display charts", "pull up charts". CRITICAL: a chart request must be served by a chart tool. Never answer "show me X charts" with a report search - a report that happens to contain images is not a chart, and filtering reports by image presence is not a substitute for searching the chart catalog. Sibling chart tools, when the user was more specific: charts_by_series (named a Macrobond series code), charts_by_report (named a report stem), charts_get (gave a chart_id).
list_authors
ChatGPTReturns all known BCA report author names, sorted alphabetically. Use this to resolve a partial or ambiguous author name (first name, last name, or nickname) to the exact full name string needed for filters.authors before calling reports_search or reports_analyze. Example: user says "Peter" -> find "Peter Berezin" -> use in filters.authors. Do not display this list to the user unless they explicitly ask for it.
list_content_categories
ChatGPTReturns all BCA content categories with a combined code and display name. The code is the combined form (platform + category, e.g. "GEOM-TRO"); some categories have no platform prefix (e.g. "EQA"). Use this to resolve a content-category name or abbreviation to its combined code before filtering. Example: user names a category -> find its combined code -> use in filters.content_category_codes (the combined form, e.g. "GEOM-TRO"). Do not display this list to the user unless they explicitly ask for it.
list_platforms
ChatGPTReturns all BCA platforms with code and display name (e.g. ESSN, CORE, PCON, EQTY, FICC, GEOM). Use this to resolve a platform name or abbreviation to its code before filtering. Example: user names a platform -> find its code -> use in filters.platform_codes. Do not display this list to the user unless they explicitly ask for it.
list_products
ChatGPTReturns all BCA product lines with code and display name. Use this to resolve a product reference or abbreviation to its code before filtering. Example: "GIS" or "Global Investment Strategy" -> code "GIS" -> use in filters.product_codes. Do not display this list to the user unless they explicitly ask for it.
list_publications
ChatGPTReturns all BCA publications with code and display name. Use this to resolve a publication name or abbreviation to its code before filtering. Example: "Global Investment Strategy" -> code -> use in filters.publication_codes. Do not display this list to the user unless they explicitly ask for it.
reports_analyze
ChatGPTUse this for ANALYTICAL questions that require synthesizing evidence across multiple reports. This is the primary tool for questions about views, stances, outlooks, and concerns. It finds the globally best matching passages across all reports in one call using rank fusion over lexical and vector search on report chunks. Input JSON: query (the analytical question), filters (same as reports_search: authors, product_codes, publication_codes, year, publication_date), top_k_docs (default 10, max 20), top_k_chunks (default 5, max 10), snippet_max_chars (default 800), sort ("relevance" default, or "date_desc"). When the user wants the LATEST reports to read and summarise ("what did <author> publish this month", "summarise the newest GeoMacro reports"), set sort="date_desc": you get the newest reports among the best matches, not the best-matching passages from any year. Leave query empty only for a pure listing of what a filter covers (platform_codes, authors, publication_date), with top_k_docs as the count. An analytical question about a view or outlook keeps sort unset. Use filters to ground the search: - Question about a specific author -> filters.authors (resolve via list_authors first) - Question about a product line -> filters.product_codes (resolve via list_products first) - Question within a time period -> filters.year or filters.publication_date After calling this tool, synthesize a comprehensive answer from the returned excerpts. Combine evidence across ALL returned reports. Do not stop at presenting a list. Follow the synthesis answer format from the server instructions. ATTRIBUTION (read before generating): attribute every view to the report's author(s), platform name (e.g. "Ryan Swift, Fixed Income (Oct 7)") and publication_date. Never write a product name, product code or team to the user, and repeat nothing from the Product code line. If a report has no Platform line, use author + date + title. Do NOT write "BCA's view" or "BCA team(s)" - not in the body and especially not in the lead sentence. "BCA" is ambiguous (firm vs. Bank Credit Analyst publication vs. Strategist Views consensus). See VIEW ATTRIBUTION in server instructions. If the user typed a product name or code, follow USER TYPES A PRODUCT in the server instructions: translate it to its platform and ask before searching, unless the period is entirely before May 2026. When the USER typed the product, tell them in one sentence that it is now part of the [Platform] platform and ask whether they want that platform's reports or the older product's reports from before May 2026 (and for which period). Do not analyze until they answer.
reports_get
ChatGPTUse this when the user provides or confirms a report_id (or document_id) and wants report metadata or a link to the report. This tool is not for content questions or excerpts; use reports_retrieve only if the user explicitly asks for evidence or what the report says. Input JSON: {"report_id": "...", "include_content": false}. report_id is required. include_content is optional and should be used only when the user explicitly requests report text and the environment allows it. Response format: After calling this tool, respond in Markdown with headings in this order: Report details, Summary. List title, report_id, year, source, authors/publications/products when present, and report_url (fallback to pdf_url). Summarize in 2-4 bullets. Do not include raw JSON; do not include excerpts unless the user asked for content and include_content was true.
reports_retrieve
ChatGPTUse this for SINGLE-REPORT deep dives only: when the user asks what a specific known report says. For analytical questions across multiple reports, use reports_analyze instead - it is faster and finds evidence from all matching reports in one call. Requires report_id and a focused query. Input JSON: {"report_id": "...", "query": "...", "top_k_chunks": 5}. Optional: candidates_per_pipeline, vector_num_candidates, fusion_limit, weights, include_score_details, snippet_max_chars. Response format: Answer the question directly, then include "Relevant excerpts" only if requested. Each excerpt: report title, report_id, chunk_id, short snippet. Do not include raw JSON. ATTRIBUTION (read before generating): attribute the view to the report's author(s), its platform name (never a product name or code, and never a team) and publication_date. Do NOT say "BCA's view" - "BCA" is ambiguous (firm vs. Bank Credit Analyst publication vs. Strategist Views consensus). See VIEW ATTRIBUTION in server instructions.
reports_search
ChatGPTUse this for DISCOVERY: when the user wants to find or list reports on a topic. This is the DEFAULT for any "find a report on X", "which reports discuss Y", "list reports about Z" request - including phrasings that name a verb like "talks", "says", "said", "expects", "argues", "wrote", "thinks", or "discusses". Verbs in casual phrasing do NOT imply audio/video - written reports are the default. For analytical questions (views, stance, outlook, what someone thought), use reports_analyze instead. A quantitative topic (FX, rates, yields) is not on its own a request for charts, and casual "talks/says" phrasing is not on its own a request for webcast clips - both default to written reports and belong here. This tool is NOT a fallback. An explicit request for charts, graphs, data series or webcast clips is not a report request, and must not be answered with a report search. If no tool available to you covers what was asked, say so and stop. NEWEST FIRST. When the user asks for the "latest", "newest" or "most recent" reports, "the last N reports", "everything published in <period>" or "what did <author> publish", set sort="date_desc" and limit=N. For a pure listing ("latest GeoMacro reports", "the last 5 reports by Matt Gertken", "everything published in Q1 2026") leave query EMPTY and put the scope in filters (platform_codes, authors, publication_date). For a topic plus recency ("the latest reports on inflation") send the topic as query with sort="date_desc": you get the newest of the best-matching reports. Without sort, results are ranked by relevance and the newest report is not necessarily first. An empty query needs at least one filter, and does not combine with page_token. Known limit: with a query, the author and publication_date filters are applied after the passage search, so for a pure "by author" listing use an empty query. Input JSON fields: query (required unless sort="date_desc" with a filter), sort ("relevance" default, or "date_desc"), limit (default 10; use 12-15 for broader coverage), page_token, intent ("discovery" or "analytical"), and filters: - authors: exact author names (resolve partial names via list_authors first) - product_codes: product codes (resolve abbreviations via list_products first) - publication_codes: publication codes (resolve via list_publications first) - year: {gte, lte} integer year range - use for named years ("in 2024") - publication_date: {gte, lte} ISO 8601 strings - use for sub-year precision ("last 6 months") - sources: exact source names has_pdf and has_images are restrictive presence filters on the index - NOT relevance preferences. Leave them unset unless the user explicitly asked for a PDF / downloadable document or for reports with images. Setting has_pdf=true in particular will silently exclude many of the most relevant reports (their pdf_key is empty in the index). Strip author/year/product words from the query - those belong in filters, not query text. Use double quotes in query for exact phrases (e.g. "leading indicators"). If the user typed a product name or code, follow USER TYPES A PRODUCT in the server instructions: translate it to its platform and ask before searching, unless the period is entirely before May 2026. The server enforces this: a search by product with no period that ends before May 2026 returns status needs_clarification instead of results. Relay its question to the user in one short sentence and do NOT search yet. Once the user picks, call again with filters.platform_codes, or with filters.product_codes and a period that ends before 2026-05-01. Response format: Markdown with headings: Summary, Top reports, Facets and filters, Warnings, Next steps. Each report: title, id, year, source, authors/publications/products, report_url (fallback pdf_url). Do not include raw JSON.
transcripts_get
ChatGPTGet a single BCA transcript clip by clip_id. Use ONLY when the user explicitly provided a clip_id AND asked for the full transcript / spoken content of that clip. Do NOT use to look up written reports - that needs a report tool, not this one. To discover clips by topic, use the clip SEARCH tool instead, and only when the user explicitly asked about webcasts/video. Returns full metadata, complete transcript_text, and individual speaker segments. Input JSON: {"clip_id": "clip_000001"}
transcripts_meeting
ChatGPTGet a WHOLE BCA Live & Unfiltered meeting - the daily call - rather than individual clips. Use this when the user asks what was discussed, for talking points, for a recap or catch-up, or about a meeting by day: "today's BCA Live", "what were the main talking points yesterday", "what did they cover on Tuesday", "catch me up on last week's calls". Returns the meeting headline, date, speakers, the Top Takeaways, each thematic section with its narrative, and the clips belonging to each section. Takeaways ARE the talking points - lead your answer with them. Use transcripts_search instead when the user wants clips ON A TOPIC across many meetings ("clips about the dollar"). Use this tool when the unit the user cares about is the MEETING. Input JSON: exactly one of date, date_from/date_to, or webcast_id. - date: "2026-09-15" (resolve "today"/"yesterday" to a date yourself) - date_from + date_to: a range, newest first, for recaps - webcast_id: exact id, e.g. taken from a transcripts_search result - limit: meetings to return for a range (default 1, max 14) A meeting with no summary yet still returns its speakers and clips in timeline order. If a section carries no clips the response says why - do not infer that the meeting had none. CLIPS IN YOUR ANSWER - do not reproduce every section's clip list. Link only the clips you refer to (at most about 5), next to the point they support. If more clips exist, say so in one line ("More clips from this meeting are available - ask if you want them.") instead of listing them.
transcripts_search
ChatGPTONLY use when the user explicitly asks about webcasts, video clips, transcripts, podcasts, recordings, or what someone said on air / in a recorded talk. Do NOT use for written-report queries - even if the user used the verb "talks", "says", "said", "discusses", "argues", or "expects" loosely. Those verbs in casual phrasing default to written reports, not webcasts. For "what did X write/argue/ expect about Y", "find a report where X discusses Y" and "report where X talks about Y" are written-report requests; route them to a report tool if you have one. Search BCA webcast transcript clips by topic, speaker, or spoken content. Each result is a focused clip (2-15 min) on one topic - clips contain verbatim spoken analysis and answer the question directly. Returns clip cards with an inline video player, title, speaker badges, duration, and a transcript excerpt. RECENCY - when the user asks for "the latest", "today's", "this week's", "the most recent" or "the last" clips, set sort="date_desc". Leave query EMPTY for a pure newest-first listing; there is no need to invent a topic. Searching a topic and relying on relevance alone does NOT work - relevance ranking only gently favours recent clips, so use sort="date_desc" for "the latest". BLU - when the user says BLU, BCA Live, BCA Live & Unfiltered, or "the daily meeting", set filters.content_type="blu". It is the only kind of video today, so omit it for a general search. Use filters.speaker_names to find clips by analyst (e.g. ["Peter Berezin"]). Use filters.publication_date to narrow by date range (ISO 8601). Input JSON: query (optional - empty means "latest"), limit, offset, sort, filters: - sort: "relevance" (default) or "date_desc" (newest first) - limit: leave it unset (default 5) unless the user names a quantity. Max 8. - speaker_names: list of analyst names (e.g. ["Dhaval Joshi"]) - publication_date: {gte, lte} ISO 8601 date strings - webcast_id: filter to clips from a specific webcast - content_type: "blu" for BCA Live & Unfiltered Clips answer the question. Synthesize an analysis paragraph first, then list clips below. Trigger words: "show me clips", "find clips", "webcast clips", "replay clips", "what did [speaker] say about", "find clips where X said Y", "show me clips on [topic]", "transcript clips", "video clips", "BCA webcast". CRITICAL: Each clip is self-contained - transcript_text contains verbatim spoken analysis. A clip or webcast request must be served by a transcript tool; never answer one with a report search. TRANSCRIPT SYNTHESIS FORMAT - use this structure for ALL transcripts_search responses: INLINE CITATIONS - MANDATORY Use numbered inline citations - (1), (2), (3) - in the analysis text to link every point or quote to a specific clip in the Clips list. Do NOT write analysis without inline citations. Number clips in the order you first cite them. Cite only clips that are on topic; skip old or off-topic ones. Cite at most about 5. Analysis One or two paragraphs synthesizing what the clips say on the topic. Lead with the key takeaway, then support with specific points drawn from clip summaries and transcript excerpts. Every claim must have an inline citation. Example: "Peter Berezin argued that US recession risk remains elevated (1)(2), pointing to deteriorating leading indicators. He noted that Fed rate cuts are constrained by sticky core inflation (1), but expected 50bp of easing by year-end (3)." Clips ONLY the clips cited in the Analysis, in citation order, numbered to match the citations. Never list a clip the Analysis does not cite: 1. [Clip Title](bca_webcast_url) - Speaker(s) - YYYY-MM-DD 2. [Clip Title](bca_webcast_url) - Speaker(s) - YYYY-MM-DD If other returned clips were left out, add one line instead of listing them: "More clips are available on this topic - ask if you want them." AI-generated summary - verify claims against the original clips linked above.
widget_list
ChatGPTList the BCA dashboards you can open, with their ids. Pass a few words of the title to narrow the list. Then call widget_open with the dashboard_id.
widget_open
ChatGPTOpen a BCA dashboard by its id and draw it in the chat. Its data is loaded now, with your own subscription: a part your subscription does not include is shown as not available. Pass params to change a filter. Don't know the id? Call widget_list first.