Invoice Processing Pipeline — How We Get to 100% Accuracy

Courier invoices arrive in many formats — from clean spreadsheets to dense, multi-page PDFs with merged cells and multi-row records. The pipeline tries the cheapest approach first and only escalates when accuracy is measurably insufficient.

End-to-End Flow

Extract

Spreadsheets (CSV, Excel) are parsed natively. Wide CSVs (250+ columns like UPS) have empty columns stripped automatically before being sent to the LLM.

PDFs go through Docling (IBM open-source) in 10-20 page chunks. Docling's memory accumulates per document (~100-600MB) with no way to reclaim, so the pipeline monitors RSS via a health endpoint and proactively restarts Docling when it exceeds 3GB — checked before every chunk and every page render.

Both produce identical output: a list of tables with column headers and rows.

Tier 1: Text Mapping

The LLM receives column headers, sample rows (8 for large tables, all rows for small tables ≤20 rows), and a column activity summary. It returns:

• Column mappings — which column maps to which schema field
• Classification rules — how to categorise services from text patterns
• Skip rules — which rows to ignore (totals only — never empty cells, which would break multi-row formats)
• Summary items — charges extracted directly from small summary tables
• Invoice total — the pre-VAT total found in summary tables (or from repeated-value columns in CSVs)

CSVs and PDFs get different prompts. Each courier has optional hints:

Self-Healing (85-115% threshold)

After applying rules, the pipeline checks: does the extracted total match the LLM-detected invoice total? If accuracy is outside 85-115%, it retries with feedback — which tables overcounted, which column it chose, and the per-table breakdown. Up to 3 attempts, each with more sample rows.

The 85-115% band is tight enough to catch real errors (the LLM mapping a wrong column at 77%) while loose enough to accept minor gaps from surcharges on summary pages (~1-5%).

Tier 2: Guided Text Extraction

Only triggers when two conditions are met:

1. Text mapping accuracy is below 95%
2. The extraction ratio is below 15% — fewer than 15% of input rows became line items

The extraction ratio gate prevents this tier from triggering on DPD invoices at 94% accuracy with thousands of items (format is understood, gap is from surcharges). It only activates for genuinely complex formats like DHL's multi-row shipment records (22 items from 600 rows = 3.6%).

Tier 3: Vision Fallback

Pages rendered as high-resolution PNG images on-demand (2 at a time, discarded after extraction) and read visually by AI models. This bypasses text extraction entirely — the last resort when the Docling text output is too damaged or the format too complex for even guided text extraction.

Tested Results

12 invoices from 4 couriers. UPS CSV was previously non-deterministic at 77% — fixed by tightening the self-healing threshold from 50-150% to 85-115%, allowing the pipeline to catch and correct wrong column mapping.

Average accuracy: 98.3%. Average cost: ~$0.05/invoice.

Summary