It inflates repeated carry-forward content into extra records, returning 55 rows for a statement that actually contains 52 distinct transactions because the BALANCE FORWARD line is duplicated across page breaks.
What was measured
Table & Record Completeness
Are all tabular rows and line items extracted without merging, duplication, omission, or phantom records?
decisive for this rankingtransformation
Missing, duplicated, or merged rows and line items directly corrupt the structured dataset and break query results. (3 of 3 judges)
What was given, what came back
Test input: Bank Statement PDF · pdf · group: business-document-extraction
Input — what we sent
A four-page bank statement PDF with dense transaction tables, balance-forward bridges, account metadata, rewards data, and disclaimer text. It was used to stress schema-driven extraction, multi-page continuity, row completeness, and financial numerical accuracy.
Why this input is hard
- · Table extraction across 50+ transaction rows
- · Multi-page continuity with balance-forward bridges
- · Parsing structured account metadata alongside unstructured transaction descriptions
- · Numerical accuracy for deposits, withdrawals, running balances, and summaries
- · Extraction of nested rewards and disclaimer sections
Output — unretouched
Also checked on this input — same tool, 3 other criteria
Extraction Accuracy◐ MixedIn the rewards scheme table, zero-value cells render as blank rather than as 0, so the extractor preserves the row structure but loses explicit zero values.Extraction Accuracy✗ FailedThe bank-statement extractor leaves multiple straightforward metadata fields empty: 1 Bank Name field, both Statement Period dates, all 4 Balances fields, and the Summary Total Transactions value were not populated despite the information being present in the document.Extraction Accuracy✗ FailedAcross the transaction grid, the tool fails to populate the Value Date and Transaction Type columns for the rows shown, leaving those fields empty instead of extracting per-row values.
Provenance
- Observation
- b18f6412-ef8f-41f1-ab39-e2924602dfaa
- Evidence run
- a061b9e7-a9c5-443d-a171-b296aaf51b8c
- Study
- Extract and query structured data from documents using natural language
- Research task
- 86b9y25e5
- Tested at
- not recorded
- Source
- first-party
- Evidence state
- verified
- Proof shown
- input + output shown
- Cost / latency
- not captured
- Repeat run
- not captured
- Tester
- not captured
The last three rows are honest blanks, not placeholders — our capture has no field for them yet.
Query this
get_evidence({
tool: "docsumo",
scenario: "business-document-extraction"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 7 other tools
measured on Table & Record Completeness
Datalab⚠ StruggledMerges adjacent bank-statement rows into one record; the 21 Jun crop shows two separate transactions, but the extracted description collapses them into one long UPI/merchant string.Landing AI✓ WorkedReturns the full transaction table as separate records; the report says all 51 statement entries were extracted without collapsing rows or dropping records.LlamaParse◐ MixedThe tool preserves transaction rows as separate records and maintains order, but the extracted transaction list is inconsistent in size: the report says the array contains 54 records even though the source statement has 51, and the summary shows 44 total_transactions.Nanonets✗ FailedLeaves the statement summary incomplete: total_transactions is null, and the report says the run extracted 47 transactions when the source contained 51.Reducto✗ FailedOver-produces the repeating transaction table, returning 53 transaction records when the bank statement contains 51, indicating row-count drift in line-item extraction.Retab◐ MixedOvercounts the bank-statement transaction table in the summary, reporting total_transactions 43 when the researcher says 40 should be counted after excluding balance-forward, tax, and charge entries.Unstract✓ WorkedKeeps the bank transaction table complete across all 4 pages, with the report stating the transactions array contains all 51 entries and no rows were dropped or merged.
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com