arizona_docket_search
ChatGPTSearch the Arizona Corporation Commission (ACC) eDocket. limit caps rows (default 50, max 500). search_type selects what to return: "search" (default) — the docket register. Filters combine with AND; dates (YYYY-MM-DD) filter the filed date. `description_contains` matches the docket description keyword-wise: ANY word matches (OR), so multi-word terms are broad/noisy — set `search_as_string` true for exact-phrase topical searches like "data center". `company_name` matches company/DBA/respondent; `docket_number` e.g. 'E-01345A-22-0144'; `year_matter` the 'YY-NNNN' part; `case_type_id` (17 = Rates). Each row has a `docket_link`. "docket" — open one docket by docket_number and return its filings, one row per document with a direct pdf_url.
arkansas_docket_search
ChatGPTSearch the Arkansas PSC (APSC) docket system. limit caps rows (default 50, max 500). search_type selects what to return: "filings" (default) — filings in a date range (needs `start_date` and `end_date`; optional `docket_type` / `company_name` / `description_contains` / `docket_number`, each substring, `|`-alternatives). Each row has a `document_url`. APSC has no server-side pagination: it builds the entire window as one page before responding, at roughly 36 filings/sec (7 days ≈ 2.6s, 30 ≈ 11s, 60 ≈ 22s), so a narrower window is faster. Ranges longer than 31 days are therefore served in 31-day slices: the response carries `window_clamped`, `returned_through` and `next_start`, and you continue by calling again from `next_start`. The optional filters cost nothing extra, since everything is already in the page. "keyword" — full-text search over filed document CONTENTS (needs query); each hit is a file tagged with docket/document number + a direct URL. * "docket" — one docket's filings by docket_number (e.g. "05-081-U"); the docket's parties are attached to the first row's parties field.
california_docket_search
ChatGPTSearch California PUC (CPUC) proceedings. limit caps rows returned (default 50, max 500). All filters optional, combined with AND. proceeding_number may be partial (e.g. "A24"). filer is an applicant-name substring. Dates are YYYY-MM-DD or MM/DD/YYYY. description matches a NARROW category field (few hits) — prefer filer + dates to scope a search. assignment matches ALJ/Commissioner text. Provide at least one filter. Each row carries a detail_url to the CPUC proceeding page.
colorado_docket_search
ChatGPTSearch the Colorado PUC E-Filings system. limit caps rows (default 50, max 500). search_type selects what to return: "proceedings" (default) — dockets (uses proceeding_number, title, industry, docket_status, dates); each row has a `detail_url`. "documents" — filed documents with direct document_url links (adds document_type). * "efilings" — e-filings, searchable by entity_name (filer) or filing_number. Dates are YYYY-MM-DD or MM/DD/YYYY. industry/docket_status/document_type take the site's dropdown labels (e.g. "Electric", "Application"). Text fields do substring matching (auto-wildcarded). Provide at least one filter; wide date ranges are heavy (no server pagination).
connecticut_docket_search
ChatGPTSearch the Connecticut PURA docket system. limit caps rows (default 50, max 500). search_type selects the modality: "recent" (default) — recent filings across dockets in a date window (`start_date`/`end_date`, YYYY-MM-DD; default = the last 7 days). CT has no native date-range search, so this scans the `scan_dockets` (default 50, max 300) most-recently-opened active dockets and filters their filings by date (a filing in an older docket outside that set is not captured). Newest-first. "docket" — look up docket_number (e.g. "26-11-02") and return its filings. Each era database is tried in turn (unless database pins one); the response notes which database held the docket. Each filing has a document_url. * "keyword" — full-text search across filed documents for query; returns matching documents (title, docket_number, filed_date, document_url).
current_weather
ChatGPTCurrent weather for a location. Requires the workspace's OpenWeather key.
dc_docket_search
ChatGPTSearch DC PSC eDocket filings by date range and optional filters. limit caps rows returned (default 50, max 500).
delaware_docket_search
ChatGPTSearch the Delaware PSC (DelaFile) dockets. limit caps rows (default 50, max 500). search_type selects the modality: "search" (default) — grid search. Filters (all optional, AND): `date_from`/`date_to` (filing date, YYYY-MM-DD), `status` (Assigned/Closed/ Reopen/Unassigned), `utility_type` (Electric/Gas/Water/Renewable/Telecom/ Cable/…), `keyword`, `company`. Each row has a `detail_url`. "docket" — look up one docket by docket_number (e.g. "26-0416") and return its filed documents (item/type/date/filer/description).
eia_browse
ChatGPTNavigate the EIA open-data catalogue. With no route, returns the top-level categories; a category returns its child routes; a dataset (leaf) returns a one-row summary (frequencies, facets, data columns, date range). Use it to find a dataset route, then eia_describe(route) and eia_get_data(route, ...). Requires the workspace's free EIA API key.
eia_describe
ChatGPTThe full query surface for an EIA dataset: data columns (measures + units), facets (filter dimensions), frequencies, and the available period range. Call this before eia_get_data to learn the valid data_columns / facets / frequency. Requires the workspace's free EIA API key.
eia_facet_values
ChatGPTList the valid values for a facet on an EIA dataset — the exact strings to pass to eia_get_data(facets={facet_id: value}). Discover facet ids with eia_describe. Requires the workspace's free EIA API key. Filter with `search (every term must appear in the id or name) and limit. Some facets are large: petroleum/pri/gnd series` is ~52KB.
eia_get_data
ChatGPTFetch data points from an EIA dataset route (e.g. "electricity/retail-sales"). Returns a preview plus a signed CSV download URL for the full result. Find the route with eia_browse and the valid data_columns / facets / frequency with eia_describe. facets is {facet_id: value or [values]}; start/end are inclusive period bounds in the dataset's format; results sort by period (newest first unless sort_desc=false); length caps rows (default 500). Omit data_columns and every measure the dataset publishes is fetched. EIA returns only facet/metadata columns when no measure is requested, so the server fills them in. Pass it explicitly to narrow the result. Requires the workspace's free EIA API key.
eia_search
ChatGPTFind EIA dataset routes by keyword. Matches every term against the route, name and description, so "henry hub" or "retail gasoline" lands on the right route in one call. Follow with eia_describe(route) for facets, then eia_get_data(route, ...). Falls back to eia_browse when the catalogue snapshot is unavailable.
ferc_docket_search
ChatGPTSearch FERC eLibrary docket filings using the AdvancedSearch API. Dates are YYYY-MM-DD. keyword is optional — omit it to return every filing in the date range. limit caps rows returned (default 50, max 500).
get_schema
ChatGPTSchema for one table, or a market's table list. Call it before writing SQL against a table you have not queried yet. To answer "what datasets are available" or a dataset count, use list_markets instead. market: lowercase market id, e.g. 'ercot'. table: optional table name. When omitted, returns datasets (user-facing, with dataset_count) and join_tables. join_tables are internal lookups (zones, counties, legends) — queryable for joins but NOT datasets; never list or count them as available datasets. When given, returns columns, types, indexes, row estimate, and the dataset card's instruction body (valid filter values, query rules).
gridstatus_list_datasets
ChatGPTList available GridStatus datasets (id + name). Requires the workspace's GridStatus API key. Use inspect_gridstatus_dataset for a dataset's details. Filter to what the question needs: `iso matches the market prefix (ercot, caiso, pjm, miso, isone, nyiso, spp), search matches words in the id or name ("lmp", "fuel mix"), and limit caps the rows. The full catalog is 534 datasets, so an unfiltered call returns a very large result. A single iso` call returns every dataset for that market (129 for ERCOT), which is enough to group, count and summarize from.
gridstatus_query
ChatGPTQuery a GridStatus dataset by ID with optional filters, time range, and resampling. Returns a preview and a CSV download URL. Requires the workspace's GridStatus API key. Use start_time+end_time OR time, never both. Rows count against the workspace's export quota, which is finite and shared — exhausting it returns 403 for everyone until it resets. Ask for the rows the question needs, not the whole range: - Aggregates (highest, average, total) — set `resample_frequency (e.g. "1 hour") with resample_function. A day of 5-minute ERCOT LMP across all settlement points is ~40,000 rows; hourly is a fraction of it, and the peak hour is still the peak. - One location — set filter_column/filter_value rather than pulling every settlement point and filtering afterwards. - A sample or a recent slice — set limit`. Pull the raw series only when the question genuinely needs every interval.
historical_weather
ChatGPTHistorical weather for a location. start/end are Unix timestamps (UTC). Requires the workspace's OpenWeather key.
inspect_gridstatus_dataset
ChatGPTFull metadata for GridStatus datasets. Requires the workspace's key. Pass several ids at once, comma-separated, to compare candidates in one call: `"ercot_lmp_by_bus,ercot_spp_real_time_15_min"`. Deciding which dataset carries the column you need means inspecting each candidate, and passing them together answers it in one call.
inspect_query
ChatGPTReturn the PostgreSQL execution plan for a query (no rows fetched). Look for 'Seq Scan' — that means no index is being used and the query will scan the whole table. The query goes through the same SQL guard as run_market_sql. market: lowercase market id from list_markets (e.g. 'ercot', 'us').
list_markets
ChatGPTThe catalogue of datasets the caller can query, by market. Use this to answer "what datasets are available" and for dataset counts: each market has a dataset_count and the response has a total_datasets. It returns only user-facing datasets — never lookup/join tables — so these counts are the ones to report. A market is a label and may span several physical databases; a table is listed only when its database has a configured connection. Tables are further filtered by data_type, so a user with data_types=[price] never sees a _load table. Markets left with no permitted tables are hidden.
maryland_docket_search
ChatGPTSearch Maryland PSC official filings by date range. limit caps rows returned (default 50, max 500). Each row carries document_url, the maillog viewer page. Set resolve_documents to also fetch that page's actual PDF links into documents — one request per row, so it is capped at 25.
ny_docket_search
ChatGPTSearch New York DPS (PSC) cases by date range and optional keyword. limit caps rows returned (default 50, max 500).
rise_gas_forwards
ChatGPTForward gas and power curves published in Rise Energy Services' monthly market update reports, from April 2025 onward. Each report prices the same hubs out several years, and a new report re-prices the whole curve, so a row is identified by BOTH the report it came from (`report_month, YYYY-MM) and the delivery period (tenor). Comparing the same tenor across two report_month values is how you see the curve move. NOTHING HERE IS A CHANGE — every row is a level. A month-over-month move is the same tenor in two ADJACENT report_month values, subtracted. It is never two tenors inside one report. Getting this wrong does not fail, it returns a confident wrong number: asked how the Feb-26 contract moved MoM, an agent differencing neighbouring tenors within the February 2026 report answered "down $0.14" and, on a rerun, "down $0.47", where the report prints +$3.47. Fetch both reports, subtract, and say which two you used. tenor is YYYY-MM for a monthly contract, cal-YYYY for a calendar strip, and strip-12m for the rolling 12-month strip. hub matches on a substring because the publisher's spelling drifts between editions, so "transco" finds every Transco zone. commodity is "gas" or "power"; market narrows power rows to PJM, NYISO or ISO-NE. quote_type on each row says whether the value is an outright price or a basis differential to Henry Hub. The regional charts changed convention between editions, so never compare an absolute row against a basis one. basis_to_absolute=True converts where it can: Henry Hub is plotted on the same charts as an outright price and printed in the contract tables, so a basis row gains absolute_value = basis + Henry Hub for the same report and tenor, plus absolute_anchor and absolute_anchor_precision naming what it was added to. It covers 517 of the 591 basis rows; the rest sit at tenors where Henry Hub was dropped as occluded, and those get absolute_value: null rather than an interpolated guess. READ THE ERROR NOTE BELOW before quoting a converted figure. These are published research figures as of the report's settlement date, not live quotes. For current prices use the EIA or GridStatus tools instead. PRECISION: every row carries a precision, and it is the first thing to look at. text is exact, parsed from the PDF's text layer. table_image was read off a printed table. chart_estimate was traced off a curve and then verified against exact values it had not seen, with the measured error in tolerance. chart_unvalidated was traced but never confirmed — some charts have no anchor to check against, and a few were checked and failed. Most of this dataset is estimated, because most of these reports publish their curves only as pictures. That is a deliberate trade for coverage: use the rows for level, direction and how a curve moved between reports, and filter with precision="text" when an answer needs an exact number. Never present a chart-derived figure as a settlement price, and when you quote a chart_unvalidated row, say it is unverified. THERE IS NO CALENDAR STRIP FOR POWER. Gas carries cal-YYYY` rows because the contract tables print them; the power charts plot months and nothing else. So a question about "PJM DOM Cal 2027" has to be answered by averaging that hub's twelve 2027 monthly rows, and returning a single month instead is wrong by a wide margin: DOM's January 2027 is $124.5 where the calendar average is $104.6, and the September 2026 report prints $104.97. Check how many months you actually got before averaging — a point dropped as occluded means an eleven-month mean, and if the missing one is the winter peak the answer comes out low. Chart granularity follows the report. Most editions plot quarterly points, so their chart rows land on quarters; the reports from August 2026 onward and the April 2025 edition plot monthly. Henry Hub has monthly resolution throughout, from the contract tables. How far the chart rows were checked differs by region, and it is worth knowing which you ar…
rise_market_takeaways
ChatGPTThe top takeaways from Rise Energy Services' monthly market update reports, each with the verbatim sentence from the report that supports it. Use this for quick market context before answering a natural gas or power question: what moved this month, which way, and why the report says so. Five are recorded per report, ranked most decision-relevant first. `report_month is YYYY-MM; omit it to search across every edition. commodity is "gas", "power" or "both". query matches all of its terms against the takeaway and its evidence. Every evidence` span was checked to appear verbatim in the source PDF, so quote it directly and cite the report month. These are one publisher's views as of that month, not Tume's, and not a live market read.
run_market_sql
ChatGPTValidate + execute SQL against a market. SELECT-only; one market per call. sql: SELECT statement; cross-market joins and DML are rejected. Use exact column names from the dataset card (see get_schema). market: lowercase market id from list_markets (e.g. 'ercot', 'us'). Must be in your permissions. A market may span several databases; all tables in one query must live in the same database. format: 'inline' returns up to INLINE_ROW_LIMIT (default 20) rows. 'csv' returns a single-use signed download URL (1h TTL). On a schema error the response includes actual_columns for each referenced table — rewrite the SQL using only those columns and retry.
sc_docket_search
ChatGPTSearch South Carolina PSC docket filings by date range and optional filters. limit caps rows returned (default 50, max 500).
solar_profile
ChatGPTAnnual normalized solar PV capacity-factor profile for a lat/lon (1980-2024). One call answers a site question: the default range is a full year of hourly values, which is what an annual capacity factor needs. Rows carry local_time (the site's clock) and utc_time. Use local_time for hour-of-day, daily shape and month; utc_time only to join UTC data. Costs to respect: the response is 8,760 rows, and Renewables Ninja meters both a 1-request-per-second burst limit and an hourly quota. Narrow date_from/date_to when you only need a season or a shape, and never repeat a call you have already made. The API rejects ranges over 1 year; for multi-year spans call once per year. Requires the workspace's Renewables Ninja API token.
texas_air_permit_detail
ChatGPTFull TCEQ record for one air authorisation, including its investigations. Pass the `permit_id` from a search result. This is the only call that carries compliance history: violations, enforcements, complaints and investigations are null on every search result and populate only here, so never state compliance from a search result. Investigations carry links to the agency's own investigation records.
texas_air_permits_search
ChatGPTSearch Texas air authorisations (TCEQ) by county, operator, site, industry, NAICS, permit program, status, issued-date window, or distance from a point. `county is matched exactly. operator and site match substrings. naics matches as a prefix, so "2211" is electric power generation and "221112" is fossil-fuel generation. program is "Air New Source" or "Air Operating"; status is "ISSUED PERMIT", "NEW APPLICATION" or "RENEWAL/AMENDMENT". Dates are YYYY-MM-DD and filter on the issued-to begin date, which is when the authorisation was attached to its current holder: a change of ownership re-dates every permit at a site, so a cluster of recent dates can be one transfer rather than new build. For a radius search pass latitude, longitude and radius_miles together. Results are newest first and carry no compliance history — use texas_air_permit_detail` for that.
texas_air_permits_summary
ChatGPTCount Texas air authorisations by county, operator, industry, naics, program, status or city, with the same filters as the search. Use it to rank counties, find an operator's largest footprints, or see which industries hold the permits in one county.
texas_docket_search
ChatGPTSearch Texas PUCT Interchange filings by date range. Returns the filings themselves — PUCT needs a control number per filings page, so this resolves the dockets for the range and fetches each one's filings server-side rather than making the caller loop. Set dockets_only for just the docket list. limit caps rows returned (default 50, max 500).
utility_tariffs
ChatGPTLook up utility electricity tariffs from the OpenEI rates database. Provide a full address (including state and zip) and sector (e.g. Residential, Commercial). `limit caps how many rates are returned (default 50); pass null for every match. Utilities that file per-zone variants of one schedule (Con Edison files Zone H/I/J for NYC and Westchester) can exhaust the default cap on near-duplicates alone, hiding distinct schedules. The response always reports total and truncated` so a capped result is visible.
virginia_docket_search
ChatGPTSearch Virginia SCC daily filings by date range (YYYY-MM-DD). limit caps rows returned (default 50, max 500). One call is enough. The SCC API returns every filing in the date range and the text filters are applied here, so a different phrasing returns the same underlying filings. docname_contains and case_contains accept several alternatives separated by |, matching a row if ANY of them appears. Pass every phrasing you would otherwise have guessed one at a time: docname_contains="rate|tariff|fuel factor|rates and charges" Leave them unset to get everything in the range and narrow it yourself.
weather_forecast
ChatGPT5-day weather forecast for a location. Requires the workspace's OpenWeather key.
wind_profile
ChatGPTAnnual normalized wind capacity-factor profile for a lat/lon (1980-2024). turbine takes a model name ("Vestas V90 2000"); height is the hub height in metres, and changing it materially changes the profile, so state the assumption in your answer. Rows carry local_time and utc_time, as in the solar profile: use local_time for hour-of-day, daily shape and month. Same costs as the solar profile: 8,760 rows per call, a 1-request-per-second burst limit, and an hourly quota. Narrow the date range when a season will do. The API rejects ranges over 1 year. Requires the workspace's Renewables Ninja API token.