Structured data out.
Never free-form text.

Documents are parsed, OCR'd and cleaned before the AI ever sees them. The output is strict structured JSON — every field with a confidence score and its exact source.

A real document pipeline

Before any AI runs, each attachment goes through file-type detection and the right parser: native text extraction for PDFs (with OCR fallback), worksheet and table detection for Excel with sheet and cell references preserved, and OCR for images and scans. Page numbers are kept throughout.

Crucially, the pipeline filters first: a 100-page PDF is classified down to its relevant pages and tables before anything reaches the model. Less cost, less latency, less noise, fewer hallucinations.

Pipeline

Attachment
  → File-type detection
  → ┌──────┬──────┬──────┬──────┐
    │ PDF  │ XLSX │ DOCX │ IMG  │
    │parser│parser│parser│ OCR  │
    └──────┴──────┴──────┴──────┘
  → Normalized content
  → AI extraction → structured JSON

Every field carries its evidence

{
  "weight_kg": {
    "value": 12480,
    "unit": "kg",
    "confidence": 0.98,
    "source": {
      "type": "attachment",
      "document": "Packing_List.pdf",
      "page": 2,
      "text": "Gross Weight: 12,480 KGS"
    }
  }
}

In the UI this renders as: Weight: 12,480 kg — Source: Packing_List.pdf, page 2.

Confidence + source on every field

Enterprise teams don't trust a black box. So Quotiveo never stores a bare value: each field carries a confidence score and a source pointer — the document, page, sheet, or email message and exact text it came from.

System-level confidence goes further, combining AI confidence with source quality, validation results and cross-document consistency. When sources disagree, the field is marked for review instead of silently picked.

Freight terminology, normalized

Customers write "40' HC", "40 HIGH CUBE" and "40FT HIGH CUBE" for the same box. A normalization layer maps variants to canonical forms through maintained dictionaries — equipment, ports, weights, units — rather than relying on the model to guess consistently.

Normalization examples

Written as	Normalized to
40' HC · 40 HIGH CUBE · 40FT HIGH CUBE	`40HC`
12.48 MT · 12480 KG · 12,480 kgs	`12,480 KG`
Shanghai port · SHA · Shanghai, China	`Shanghai · CNSHA`

Provider-independent AI

// The extraction service receives:
Email + document text
  + document metadata
  + RFQ type + schema
// …and returns structured data.
// Swap models on accuracy, cost,
// latency, privacy — no rewrite.

Not locked to one model

The AI layer sits behind a provider-independent gateway. The extraction service takes email, document text, metadata, RFQ type and schema — and returns structured data. Which model does the work can change over time based on accuracy, cost, latency, privacy or customer requirements, without rebuilding the product around it.

Next stage

Extracted. Now prove it.

Structured data means nothing without checks. The validation engine verifies required fields, hunts for missing info and contradictions — then a human approves.

Continue to validation & review →

⚠ Why this matters

A forwarder quoting on the wrong container count loses money or credibility. Extraction you can't audit is extraction you can't trust — which is why every value ships with its receipt.

Test extraction on your messiest RFQs.

Hand us the PDFs and spreadsheets your team dreads. We'll show you structured, cited output — and exactly where it needs a human.

Book a pilot

Structured data out.Never free-form text.