TaskTalk Logo TaskTalk
Deterministic & Rule-Based Extraction Engine v4.2 · Production Live
STEP 01
Document Input Completed
STEP 02
OCR Recognition Completed
STEP 03
Classification Completed
STEP 04 CURRENT
Extraction & Valid. YOU ARE HERE
STEP 05 NEXT
AI Intelligence Pipeline Pending
STEP 06
Review & Approval HITL Stage
STEP 07
Analytics Telemetry
STEP 08
Export & Sync ERP/Core Integration
DOCUMENT INTELLIGENCE PIPELINE · STEP 04

Turn raw text into structured data. Validated against your business rules.

tasktalk extracts key-value pairs, multi-page tables, and structured data from Vietnamese business documents. Built-in validation cross-checks totals, dates, and reference databases instantly. The output is clean, reliable data ready for integration.

99.5%
Line-Item Accuracy

Deterministic multi-page table stitching & column boundary alignment across long records.

< 850ms
Validation Run

Sub-second cross-reference against core database and mathematical logic engines.

100%
Vietnamese Formats

Native parsing for 1.000.000 VNĐ, DD/MM/YYYY, and 4-tier administrative addresses.

0%
Silent Calculation Errors

Instant arithmetic check across line items, tax rates, and stated grand totals.

THE EXTRACTION BOTTLENECK

Reading a document is only half the battle. The real value is in understanding what that text means.

Generic OCR tools dump continuous blocks of text. Without contextual structural anchors, operations teams still spend thousands of hours manually locating subtotals, calculating line item sums, and manually cross-checking with enterprise master databases.

X

Conventional OCR & Manual Keying

Traditional Template Matching & Heuristics

  • Brittle Key-Value Heuristics

    Fails when invoice layouts deviate by even a few pixels; requires constant regex re-templating.

  • Collapsed Multi-Page Tables

    Cannot reconcile continuing tables across pages, causing dropped items and orphaned tax rows.

  • Undetected Math Discrepancies

    Reads corrupted or handwritten totals without performing line item addition verification.

  • High Operational Overhead

    Skilled underwriters and accountants spend 70% of their workday doing rote data cross-checking.

TASKTALK PARADIGM
✓

tasktalk Extraction & Automated Validation

Layout-Aware Transformers & Deterministic Rules

  • Layout-Aware Semantic Field Binding

    Spatial attention models bind keys to values accurately regardless of shifted invoice geometries.

  • Seamless Multi-Page Table Reconstruction

    Intelligently stitches item rows, strips repeated header boundaries, and links subtotal rows.

  • Automated Mathematical Verification

    Recomputes (Sum of Items) + (Tax Rate × Subtotal) against Stated Grand Total in <10ms.

  • Straight-Through Processing (STP)

    Automated lookup against internal ERP, CRM, and policyholder master databases with zero human touch.

ENGINEERING CONSOLE SIMULATION

Interactive Extraction & Arithmetic Verification Engine

Inspect real-time spatial boundary binding alongside the validated schema and automated zero-error arithmetic checks.

SCHEMA: vn_medical_v4.json
DOC_PREVIEW: BVBM_INV_0012345.PDF PAGE 1 OF 3
BỆNH VIỆN BẠCH MAI
78 Giải Phóng, Phương Mai, Đống Đa, Hà Nội
PHIẾU THU TIỀN VIỆN PHÍ KIÊM HÓA ĐƠN
Số hóa đơn / No: 0012345 BBOX_01 · 99.8%
Ngày khám / Date: 15/10/2026 BBOX_02 · 99.4%
Họ và tên bệnh nhân: Nguyễn Văn A BBOX_03 · 99.9%
Mã hồ sơ bệnh án: HS-2026-99120
TABLE_STITCH_3_PAGES
DANH MỤC SL THÀNH TIỀN
01. Xét nghiệm máu tổng quát 1 450.000
02. Chụp X-Quang phổi 1 350.000
03. Siêu âm ổ bụng 1 300.000
+ 17 mục tiếp theo trên Trang 2 & 3... -- 400.000
TỔNG CỘNG THANH TOÁN: 1.650.000 VNĐ
OCR Boundary Engine: Active Optical Spatial Confidence 99.82%
VALIDATED SCHEMA & ARITHMETIC AUDIT STATUS: PASSED (0 ERRORS)
Latency: 412ms
PRIMARY FIELD EXTRACTION
INVOICE_NUMBER 0012345
Format Check: Valid 7-digit series VALID
DATE_OF_SERVICE 15/10/2026
Range Check: Valid DD/MM/YYYY VALID
PATIENT_NAME Nguyễn Văn A
DB Match: Policyholder #VN-90821 MATCHED
ITEMIZED TABLE RECONSTRUCTION (20 ITEMS ACROSS 3 PAGES) 3 Pages Stitched
# Line Item Description Page Amount (VNĐ) Status
01 Xét nghiệm huyết học tổng quát P1 450.000 ✓
02 Chụp X-Quang tim phổi thẳng P1 350.000 ✓
03 Siêu âm Doppler màu ổ bụng P1 300.000 ✓
04 Khám chuyên khoa Tim mạch P2 150.000 ✓
05 Điện tim đồ thường 6 cần P2 100.000 ✓
06 Đơn thuốc theo phác đồ (15 khoản mục) P3 150.000 ✓
DETERMINISTIC ARITHMETIC RECONCILIATION ZERO DISCREPANCY
CALCULATED SUBTOTAL 1.500.000 VNĐ
STATED TAX (10%) 150.000 VNĐ
STATED GRAND TOTAL 1.650.000 VNĐ
VERIFICATION EQUATION: 1.500.000 + 150.000 == 1.650.000 VNĐ
STATUS: PASSED (Delta = 0 VNĐ)
Record 0012345 · Ready for AI Claim Intelligence (Step 05)
EXECUTION ARCHITECTURE

How Data Extraction & Validation Works

A three-phase deterministic process ensures data integrity before records move to machine learning decisioning engines.

STAGE 4.1

Extract

"Pull the data that matters."

Isolates critical fields from dense text blocks with layout-aware spatial geometry.

  • Key-value pairs: Names, identification numbers, invoice codes, tax IDs.
  • Multi-page tables: Itemized lines stitched across page boundaries seamlessly.
  • Visual indicators: Checkboxes, stamps, handwritten marks, and signature lines.
Output: Normalized Key-Value Map
STAGE 4.2

Validate

"Ensure data accuracy and consistency."

Executes rigorous mathematical and programmatic rules to verify every extracted datapoint.

  • Format checks: DD/MM/YYYY dates, 12-digit CCCD identifiers, phone syntax.
  • Math checks: Calculated line items + declared tax = stated grand total.
  • Database lookups: Cross-references against internal customer and policy catalogs.
Output: Certified Arithmetic Audit
STAGE 4.3

Flag for Review

"Highlight exceptions, not the rule."

Isolates discrepancies and routes low-confidence fields directly to operators for rapid human review.

  • Straight-Through Processing: 100% compliant data flows automatically to Step 5.
  • Targeted highlights: Low-confidence tokens flagged without halting the pipeline.
  • Fast remediation: Operators review only the exact discrepancy with 1-click accept.
Output: HITL Exception Queue
DIFFERENTIATION

What makes tasktalk extraction different

Engineered specifically for complex Vietnamese operational realities rather than generic English templates.

01

Purpose-Built for Vietnamese Documents

Pre-trained on Vietnamese date formats (DD/MM/YYYY), currency notation (1.000.000 VNĐ with period separators), administrative structures (Phường/Xã, Quận/Huyện, Tỉnh/Thành phố), and standard hospital and garage estimate templates.

Support: 63 Provinces, 705 Districts, all Ministry of Health invoice formats.
02

Multi-Page Table Extraction

Tables that span multiple pages are intelligently stitched together. Headers repeated on continuation pages are recognized and reconciled, and running subtotals are validated against the final invoice amount without manual intervention.

Feature: Automatic column alignment across page breaks up to 50+ pages.
03

Continuous Learning (HITL Feedback)

When human reviewers correct an extracted field, that feedback trains the extraction model. Over time, straight-through processing rates improve for recurring document formats, progressively lowering cost-per-claim.

Metric: STP rates typically surge from 74% to 94%+ within 90 operating days.
REAL-WORLD VERIFICATION

Extraction in Practice — Medical Invoice

Scenario: Ingesting and reconciling an unstructured 3-page hospital invoice from Bệnh viện Bạch Mai with 20 itemized service lines.

FIELD / COMPONENT EXTRACTED VALUE VALIDATION RULE & EXECUTION RESULT
Invoice No. 0012345 FORMAT MATCH Checked for expected 7-digit series length and hospital code convention.
Date 15/10/2026 RANGE VERIFIED Checked against acceptable policy date range and admission timeframe.
Patient Name Nguyễn Văn A MASTER DB MATCH Matched against insurance core database (Policy #VN-90821).
Line Items (20 items across 3 pages) Blood test, X-Ray, Ultrasound, Consultation, Medications... TABLE STITCHED Extracted as a clean, structured list. Column headers aligned across 3 pages.
Totals & Tax Calculation Subtotal: 1.500.000 VNĐ
Tax (10%): 150.000 VNĐ
Total: 1.650.000 VNĐ
ARITHMETIC PASSED Math check: 1.500.000 + 150.000 = 1.650.000 VNĐ. Verified zero deviation.
OPERATIONAL RESULT
Processed 3 pages with 20 items and validated all totals in 412ms.

Zero human intervention required. Ready for Step 5: AI Intelligence & Claim Decisioning.

END-TO-END AUTOMATION MAP

The 8-Step Document AI Pipeline

Data Extraction & Validation (Step 4) bridges sensory capture with downstream cognitive reasoning and core system integration.

STEP 01

Document Input

Multi-channel ingestion (API, email, upload) supporting PDF, TIFF, JPG, and raw mobile scans.

STEP 02

OCR & Text Recognition

Specialized Vietnamese OCR engine with accent normalization and diacritic precision.

STEP 03

Document Classification

Automated separation into invoices, receipts, diagnostic reports, and IDs.

STEP 04 · CURRENT ACTIVE

Data Extraction & Validation

Layout-aware key-value extraction, multi-page tables, and strict arithmetic validation.

STEP 05 NEXT

AI Intelligence

Fraud detection, coverage checking, medical necessity review, and auto-recommendation.

STEP 06

Review & Approval

Human-in-the-loop workspace highlighting only edge cases and arithmetic exceptions.

STEP 07

Processing Analytics

Real-time dashboards for turnaround time, STP rates, error trends, and audit logs.

STEP 08

Data Export & Integration

Push clean validated data into core insurance, ERP, CRM, or accounting systems via REST/webhooks.

TECHNICAL SPECIFICATIONS

Frequently Asked Questions

Everything enterprise architects and operations heads need to know about tasktalk data extraction.

What types of data can tasktalk extract?

tasktalk extracts key-value pairs (names, dates, IDs, totals), itemized tables (multi-page line items with quantities, unit prices, and amounts), checkboxes and radio buttons (selection states), signatures and stamps (presence and boundary coordinates), and unstructured text blocks (diagnoses, remarks, notes).

How does tasktalk handle math errors on invoices?

tasktalk automatically recomputes all line items, subtotals, tax calculations, and grand totals. If the numbers don't match, the document is flagged with the exact discrepancy highlighted (e.g., "Calculated total: 1.650.000 VNĐ vs. Stated total: 1.600.000 VNĐ — difference: 50.000 VNĐ"). The human reviewer can see both values and decide how to resolve it.

Can tasktalk validate data against our internal databases?

Yes. tasktalk integrates with your core systems via REST APIs to cross-reference extracted data in real time. For example, it can verify that a policy number exists, that the claimant's name matches the policyholder, or that a hospital is within your approved network.

Does tasktalk support custom extraction fields?

Yes. You can define custom extraction schemas for your specific document types through our schema builder or API. Specify the fields you need, their expected formats, validation rules, and confidence thresholds. No custom coding or regex hacking is required.

What happens if tasktalk extracts a value with low confidence?

Every extracted field includes an individual confidence score (0–100%). Fields that fall below your configured threshold are automatically flagged and routed to the Human-in-the-Loop review queue (Step 6). High-confidence fields on the same document are still processed straight through.

Is the extraction process secure?

Yes. All data is encrypted in transit (TLS 1.3) and at rest (AES-256). tasktalk is SOC 2 Type II certified and complies with Vietnamese data protection regulations (Decree 13/2023/NĐ-CP). We support on-premise and private cloud deployments for clients with strict data residency requirements.

DEPLOY STEP 04 IN PRODUCTION

Stop keying in data manually. Let tasktalk extract and validate it for you.

See how tasktalk handles multi-page tables, Vietnamese address formats, and automated math checks on real documents.

Enterprise inquiries: hello@tasktalk.com · Response time: < 2 business hours