File formats
What the agent reads, how each format is handled, and what to do about the awkward ones.
Ledgers and balances
| Format | Handling |
|---|---|
| XLSX / XLS | Each sheet read; header row detected (first row with ≥ 3 non-numeric cells); merged cells unmerged |
| CSV / TSV | Delimiter and encoding detected (UTF-8, GBK, Big5, Windows-1252) |
| Multi-currency columns | Local-currency and group-currency columns both captured when present |
| Sign conventions | Debit/credit columns, signed amounts, or CR/DR suffixes — normalised per entity mapping |
Documents
| Format | Handling |
|---|---|
| PDF with text layer | Text extracted per page; tables detected |
| Scanned PDF, JPG, PNG | OCR; confidence per region; low-confidence regions highlighted |
| DOCX | Converted to text with structure |
| Handwritten amendments | Detected and flagged; never read as the value |
Bank statements
PDF (text or scan), CSV, MT940, camt.053. Fee lines, FX lines and references are extracted as separate fields so netted fees can be matched to the payment they sit next to.
Correspondence
EML and MSG files, PDF of a thread, or pasted text. The agent extracts decisions, amounts, dates and who said what; quoted history is de-duplicated.
Agreements and policies
PDF or DOCX. Clauses are indexed by their numbering so explanations can cite §4.1 and link to the page.
Limits
- 200 MB per file, 2 GB per upload batch.
- 10,000 pages per month on Practice before soft-cap review; higher on Group and Enterprise.
- Password-protected PDFs must be unlocked before upload.
Awkward cases
- Exports with totals rows — detected and excluded; confirm in mapping review.
- Two entities in one file — split by the entity column if present, otherwise upload per entity.
- Statements in images only — OCR works; check the highlighted regions.