Data catalogue
Find datasets, series, publications and extracts across the marketplace.
- Warehouse rows
- 6.17B
- Structured DB
- 1,711.35 GB
- Live data store
- 189.73 GB
- Data tables
- 1,03,091
- Publications & data products
- 1,749
About these figures
BigQuery reports 6,167,191,751 physical rows across 89 native base tables in 9 catalogue warehouses, measured on . Structured DB measures logical bytes across the same tables. The headline excludes 22 verified staging and backup tables with 365,935,268 rows and 57.49 GB of logical storage. Separate publisher editions can contain overlapping observations.
The earlier 40-MCP inventory reported 1,115,017,091 records across warehouses, files and upstream sources, including 811,051,329 upstream survey records from MoSPI NADA.
Live data store measures 189.73 GB of live catalogue objects across 39 Cloud Storage buckets, measured on .
UNCTAD ingestion files account for 153.97 GB of additional live storage. Retained soft-deleted objects account for 109.84 GB. Raw publisher archives and their corresponding BigQuery tables are measured in their respective storage layers.
Data tables counts 1,03,091 publisher tables held across the marketplace MCPs, measured in each MCP's own store on 2026-11-10: one table, sheet, dataflow, report section or API dataset in one edition counts once, the way data.gov.in counts resources. Raw copies, backups and archive duplicates are not counted twice. They belong to 1,749 publications and data products — six Economic Survey reports are one publication — which hold 10,664 editions, 8,62,701 series and 19,717 documents. India Intelligence and Survey chapter packs are built from these MCPs and are not added. Geo-Spatial can also reach 2,36,742 data.gov.in resources live; they are not held and not counted.
The search index includes 18,933 datasets (tables, survey studies and map layers) and 493 individually indexed series. Separate live and archived access routes can have their own entries. These figures apply across the catalogue, regardless of filters.
All assets
74 results
build dataset
End to end in one call: spec → entities → plan → retrieve → reconcile → fill gaps → validate → workbook. Takes minutes; clients with short HTTP timeouts should use start_dataset_job + get_dataset_job. To review the spec first, call define_dataset_spec, edit, then start_dataset_job(dataset_id=...).
build workbook
Write the .xlsx: README, the spec's data tabs (derived rows as live formulas, sources in cell comments), Sources, Provenance, Conflicts, Gaps, Validation.
cancel article job
Cancel an active background job, retaining data/evidence and drafts for inspection. Repeated calls return its terminal status. Interrupted jobs can be resumed stepwise.
cancel dataset job
Cancel a local background build; keep recorded evidence for restart with start_dataset_job(dataset_id=...). Completed tasks are not retrieved again.
check data artifacts
Before researching causes, rule out artifacts: base effects, seasonal repeats, reverting spikes, unit/scale jumps, near-zero bases. Defaults to kept findings.
check dataset values
Dry-run the value gate. Return each rejection reason and normalized value without writing candidates or marking a task complete.
check research evidence
Dry-run the same evidence gate as submit_research_result without changing evidence, brief status or coverage. Returns each rejection reason for repair.
compute metrics
Period-on-period and YoY change (pp for rates, % for levels), 4-period moving average, CAGR and summary for one series. `filters` picks dimension levels, e.g. {"segment": "BFSI"}.
correlate
Correlation between two series at lags 0..max_lag (a leading b). Series are selectors {table, measure, filters}. on='changes' (default) or 'levels'. Stored as a citable fact (A-id).
coverage matrix
Entity × period grid of cell status (R reported, D derived, O override, · missing) per metric.
decompose change
What drove a change between two periods. Without `weight`: contribution of each dimension level. With `weight` (e.g. revenue when measure is a margin): mix vs rate effects. The result is stored as a citable fact (A-id).
define dataset spec
Create a dataset and its Spec: entities (rows), metrics (definition, unit, basis, domain, derive formula), periods (frequency, range) and tab layout. Pass `spec` to supply one directly (same shape as returned) instead of having it drafted from `query`. Review the result — and edit with update_dataset_spec — before retrieval: every later cost is driven by this grid.
define scope
Create the Research Charter every later stage is keyed to: core question, sub-questions SQ1..n, in/out of scope, completion rule. Pass `charter` to set or edit it directly (same shape as returned) instead of having Claude derive it from `query`. Works before any data is ingested (the Charter then comes from the question alone). create_if_missing=True creates the session under `session_id` if it do
detect patterns
Run pattern detection on some series without writing findings: unusual moves vs history, step changes beyond trend, sustained trends, inflections, outliers, mix shifts.
discover sources
Connect to registered sources, list their tools, infer capabilities; optionally list candidates from the MCP registry (INSIGHT_MCP_REGISTRY_URL).
discover sources
Connect to registered sources, list their tools and infer capabilities. With from_registry=True, also list candidate MCP servers from INSIGHT_MCP_REGISTRY_URL (candidates are not enabled until register_source).
download workbook
The latest workbook: a signed 24-hour download link that opens in a browser, and, with include_base64=True, the file itself (≤ 15 MB).
draft article
Write the article from the outline, using only fact-sheet numbers and cited evidence.
estimate cost
Cells, tasks, expected tool calls and minutes for the current Spec.
fill gaps
Plan one narrower follow-up round for missing cells (must-have metrics by default) and, with run=True, execute it and re-reconcile. Cells still missing afterwards stay blank and are listed on the Gaps tab.
Inventory snapshot 15 Nov 2026. Publisher dates and individual source checks can differ.