Evidence · first-party tested/Best AI Tools for Extracting Structured Data from PDFs and Business Documents
Does not preserve the repeating transaction table faithfully, returning 53 transaction records for a statement that contains 51 transactions.
What was measured
Table & Record Completeness
Are all tabular rows (transactions, line items) extracted without merging, duplication, omission, or phantom records?
decisive for this rankingtransformation
For document data extraction, missing, merged, duplicated, or phantom rows directly corrupt the structured dataset and break downstream querying. (3 of 3 judges)
What was given, what came back
Test input: Bank Statement PDF · pdf · group: financial-document-extraction
Input — what we sent
A 4-page bank statement PDF with 51 transactions, balances, rewards, and disclaimer text, used to test schema-driven extraction of dense financial tables and multi-page continuity.
Why this input is hard
- · Table extraction across 50+ transaction rows
- · Multi-page continuity with BALANCE FORWARD bridges
- · Structured metadata vs. free-text transaction descriptions
- · Numerical accuracy for balances, deposits, withdrawals, and summaries
- · Nested schema population for account, branch, balances, rewards, and disclaimers
Also checked on this input — same tool, 5 other criteria
Extraction Accuracy◐ MixedGets some bank values right but misstates a core aggregate: summary.total_transactions is reported as 70 even though the report says the correct count is 40.Extraction Accuracy✓ WorkedThe tool correctly extracts the bank statement currency as INR and includes citation metadata for that field.Schema Adherence✓ WorkedReconstructs the bank statement into a nested JSON structure aligned to the requested schema, with document metadata, account, branch, statement period, transactions, summary, and rewards-style sections instead of flat OCR text.Semantic Field Enrichment✗ FailedLeaves derived transaction metadata incomplete, with transaction_type and transaction_id missing across transaction rows.Structural Clean Output◐ MixedExports JSON directly, but the bank workflow is not cleanly consumable end-to-end because the report says the result still needs reconciliation and correction before use.
Provenance
- Observation
- d7084667-c138-440e-ae8a-85055f890e01
- Evidence run
- ec4d736d-95f9-4c88-884c-e280435f7b7b
- 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: "reducto",
scenario: "financial-document-extraction"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 7 other tools
measured on Table & Record Completeness
Datalab✗ FailedCan attach a transaction to the wrong date when rows are merged, with an 18 Jun 19 withdrawal appearing under a 19 Jun 19 record in the extracted output.Extend AI✓ WorkedThe extractor keeps transaction rows as separate records, and the report says it captured all 51 transactions without merging adjacent rows.Landing AI✓ WorkedPreserves the full transaction table as 51 separate statement records, without merging rows or dropping entries.LlamaParse✗ FailedOver-segments the bank-statement transaction table, outputting 54 transaction rows for a statement the report says contains 51 transactions and leaving the summary at 44 total_transactions, so row counts are not self-consistent.Nanonets✗ FailedIt undercounts the transaction table, extracting 47 transactions when the statement actually contains 51, so 4 records are missing.Retab◐ MixedOvercounts the transaction table summary: total_transactions is 43, while the report says the expected count is 40 after excluding Balance Forward, tax, and charge entries.Unstract✓ WorkedKeeps the transaction table intact across all four pages, with the full 51-row record set present and no merged or dropped rows.
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com