Notion import

/notion-import

Bulk-load rows into an existing Notion DB without manual one-at-a-time entry. Always non-destructive — never modifies existing rows, only creates new ones (unless explicitly told to upsert).

Inputs

• Target database (required). URL or ID.
• Source data (required). One of:
  • File path to a .csv / .tsv / .md / .json on disk
  • Google Sheets URL (read via Composio per memory composio_default_mcp.md)
  • Pasted block of text in the conversation (parse as CSV/TSV/markdown table)
• Mapping override (optional). If the user wants Vendor Name → Tool, they can specify upfront; otherwise infer.
• Upsert behaviour (optional). Default: create-only. If user passes "upsert by <property>", match existing rows on that property and update instead of duplicating.

Steps

1. Read target. notion-fetch the DB → get the schema (property names + types + select options).
2. Read source. Parse the CSV / pasted text / Sheet. Detect header row + types per column.
3. Auto-map columns. For each source column, find best-match target property by:
   • Exact case-insensitive name match (highest priority)
   • Substring / synonym match (e.g. "Cost" → "Monthly Cost $")
   • Type compatibility (don't map text-only column to a number property)
   • Leave unmapped columns as a "drop these" list — surface them.
4. Coerce values.
   • Numbers: strip "$", "%", "," — parse to float.
   • Select: match value to existing option (case-insensitive). If no match, add to the "new options to create" list — show user before creating.
   • Multi-select: split on , | or ;.
   • Dates: parse ISO 8601 / MM/DD/YYYY / YYYY-MM-DD / Mon DD YYYY. Surface ambiguous values.
   • Title: take the first non-empty mapped source column if title not explicit.
5. Dry-run report. Output BEFORE any write:
   • N rows to create (M rows skipped + reasons)
   • Column mapping table (source → target, type, sample value)
   • New select options that'll be created
   • Any rows that failed parsing (with reasons)
   • First 3 fully-mapped rows as a sample.
6. Confirm gate. Wait for user "go" / "yes" / explicit confirmation.
7. Write. Batch via notion-create-pages with data_source_id parent. Notion API caps at 100 pages/call — chunk if needed.
8. Confirm. Print: rows created (with URLs of first 3 + last 3), any failures.

Decision rules

• If the user pastes data without a clear header row, ask whether row 1 is headers or data before parsing.
• If a source column maps to a number property but contains "$1,234.56", strip + parse, do NOT skip the row.
• If a source column maps to a select property but the value isn't an existing option, surface in the dry-run as "new option to create" — do not silently create.
• If a source row has no value for the title property, skip it AND list it in the failures section. Don't make up a title.

Anti-patterns

• Don't auto-create new select options without surfacing them first.
• Don't upsert by default — duplicates are recoverable, overwrites aren't.
• Don't import without a dry-run, even if the source is small.
• Don't trust the source's column order — match by name.

Output

• Dry-run report (in conversation).
• After confirm: count of rows created + URLs of a few samples.
• If failures: a list with reasons.

Hard-won gotchas

• Google Sheets via Composio: the composio_sheets_gotchas.md memory has read+write gotchas. Use proxy_execute for anything advanced.
• Notion API pagination: chunk creates into batches of 100.
• Rate limits: Notion is generous but if writing 500+ rows, space the batches.