nocobase-data-analysis（nocobase 数据分析）

Goal

Use NocoBase MCP tools to locate the right collection, query business data safely, and produce reliable summaries or grouped analysis.

Prerequisite

• NocoBase MCP must already be reachable and authenticated.
• If MCP returns authentication errors such as Auth required, stop and ask the user to refresh MCP authentication before continuing.

Default strategy

1. Inspect the data source path first.
2. Prefer the main data source.
3. If the target collection is not in main, inspect other enabled data sources.
4. Once the collection is located, use that dataSource explicitly in every subsequent resource_* call.
5. Prefer resource_query for counts and grouped analysis only after confirming the query parameter contract.
6. Fall back to resource_list plus manual counting when query results are suspicious or need cross-checking.

Useful references:

• references/analysis-patterns.md for common business analysis shapes
• references/metric-checklist.md for metric definition and scope checks
• references/entity-mapping.md for mapping business terms to collections and fields

Data source discovery

• If the user explicitly names a data source, use it directly.
• Otherwise start with main.
• When the target collection is not found in main, call data_sources:list_enabled and inspect other enabled data sources one by one.
• If multiple data sources contain the same collection name:

- default to main when main is one of them; - otherwise present the candidates and explain which one you are using.

• In the final answer, state which data source was used.

Collection discovery

• If the user gives a collection name, verify it exists before querying.
• If the user uses business terms such as "leads", "users", "orders", or "opportunities", inspect collection metadata to map the business term to the actual collection name.
• Prefer collections:listMeta for a fast overview.
• Then use collections:get with appends: ["fields"] when you need exact field names, relation targets, or enum options.
• Use references/entity-mapping.md as a reusable heuristic for common business nouns and likely field categories.

Query contract checks

Before using resource_query, verify the request shape matches the real backend contract:

• measures[].aggregation, not aggregate
• orders[].order, not direction
• field should usually be passed as a field path array such as ["id"] or ["owner", "nickname"]
• Use alias whenever the result will be referenced in output or having

Correct examples:

{
  "resource": "lead",
  "dataSource": "main",
  "measures": [
    { "aggregation": "count", "field": ["id"], "alias": "lead_count" }
  ]
}

{
  "resource": "lead",
  "dataSource": "main",
  "dimensions": [
    { "field": ["status"], "alias": "status" }
  ],
  "measures": [
    { "aggregation": "count", "field": ["id"], "alias": "lead_count" }
  ],
  "orders": [
    { "field": ["status"], "alias": "status", "order": "asc" }
  ]
}

Recommended workflow

1. Confirm reachability

• Use auth:check.
• If authentication fails, stop.

2. Find the collection

• First inspect main.
• If not found, inspect other enabled data sources.
• Read fields before querying if field names or relations are uncertain.

3. Start with simple counts

• Use resource_query with a single count(id) measure.
• Keep the first query minimal so you can validate the result shape quickly.

4. Add grouped analysis

Common grouped views:

• by status
• by owner
• by source
• by department
• by created date or month

For relation labels, use field paths such as:

• ["owner", "nickname"]
• ["mainDepartment", "title"]

5. Cross-check when needed

Re-check with resource_list when:

• the grouped rows look duplicated unexpectedly;
• the numeric result looks like record IDs instead of counts;
• totals do not match between summary and grouped output;
• the collection may be affected by ACL scope or hidden filters.

When cross-checking:

• fetch enough rows to cover the visible dataset or use pagination;
• count and group manually from the returned records;
• compare the manual result with resource_query.

6. Present the result

Report:

• the collection used;
• the data source used;
• the total count;
• the key grouped breakdowns;
• any caveat such as ACL scope, null values, or fallback to manual verification.

Analysis entry points

Classify the user request before querying:

• overview for current totals and main distributions
• distribution for grouped counts by status, owner, source, team, or department
• funnel for stage-based business progression
• trend for date or month-based change over time
• ranking for top owners, sources, accounts, or products
• quality-check for missing values, null-heavy fields, suspicious statuses, or orphaned relations

Use references/analysis-patterns.md for the recommended query shapes for each pattern.

Metric definition checks

Before returning an answer, verify the metric scope:

• what time range is included
• which time field drives the range
• whether the metric is total count, distinct count, sum, or average
• whether archived, inactive, null, or other terminal states should be included
• whether grouped totals reconcile with the grand total

Use references/metric-checklist.md when the user request is ambiguous or the metric may be interpreted in more than one way.

Common pitfalls

• Do not assume the collection is in main; check main first, then search other enabled data sources.
• Do not omit dataSource after the collection has been located.
• Do not use aggregate in query measures; the backend expects aggregation.
• Do not use direction in query orders; the backend expects order.
• Do not assume suspicious aggregate output is correct.
• If a "count" result looks like 36, 54, 80, or another plausible record ID, verify whether aggregation was actually applied.
• Relation label grouping requires the real relation path and target field, not guessed labels.

Verification checklist

• MCP is authenticated.
• The collection exists in the chosen data source.
• The fields used in dimensions, measures, orders, and filter actually exist.
• resource_query uses aggregation and order.
• Summary totals match grouped totals, or any mismatch is explained.
• The final answer states the data source and any verification fallback used.