TaskTalk Logo TaskTalk
DOCUMENT INTELLIGENCE PIPELINE · STEP 03

Sort, split, and route any document in seconds. Zero manual triage.

From messy multi-page PDF bundles to mixed claim dossiers. TaskTalk automatically identifies document types, detects page boundaries to split bundled files, and routes every item to its specialized extraction model. No manual sorting. No rule configuration.

Explore Classification Models → View the full pipeline ↓
98.4%
Classification Accuracy
Multimodal vision + semantic Vietnamese text engine.
< 3.0s
Multi-Page Split
Unbundles 10-page compound PDF packets into distinct items.
50+
Vietnamese Archetypes
Pre-trained insurance, garage, medical, and banking templates.
≥ 85%
STP Threshold
Straight-through automated schema routing with zero human pause.
THE TRIAGE BOTTLENECK

Before data can be extracted, you must know what you are looking at.

Every day, enterprise back-offices receive thousands of unstructured packets: 12-page claims scans containing mixed hospital discharge notes, pharmacy invoices, handwritten prescriptions, and laminated citizen IDs. Manual triage forces knowledge workers to act as human file sorters. Misclassification halts automated pipelines.

Conventional Manual & Rule Triage
Legacy Stack
  • Paging through multi-page PDFs Operators manually inspect 10–50 page bundles to determine where one document ends and the next begins.
  • Brittle keyword/OCR regex heuristics Fails when scans are skewed, titles are abbreviated, or official Vietnamese headers vary between municipal issuers.
  • Downstream extraction failures If a hospital bill is sent into a prescription schema, the extraction fails completely, triggering manual reviews.
  • Linear latency and operational cost Triage consumes 40% of total document handling time before verification even begins.
Average intake turnaround: 18 - 45 minutes per dossier
TaskTalk Multimodal Splitting & Classification
Automated
  • Automated visual page boundary detection Identifies header transitions, page numbering sequences, and physical formatting breaks with zero separator sheets.
  • Dual-engine visual geometry + NLP semantics Combines layout structure, stamps, tables, and Vietnamese lexical semantics to accurately verify the archetype.
  • Instant deterministic schema routing Sends invoices to invoice parsers, IDs to OCR identity engines, and quotes to table extractors concurrently.
  • Sub-3 second processing per compound packet Straight-through processing allows 85%+ of volume to proceed downstream with zero human intervention.
Intake turnaround: < 3.0 seconds per dossier STP Rate: 88.4%
HIGH-FIDELITY CONSOLE DEMONSTRATION

Interactive Dossier Splitting & Routing Inspector

Simulated Run: Motor Vehicle Damage Claim (Xe cơ giới)
DOSSIER-VN-2026-CLAIM-8921 | 6-Page Compound PDF Bundle | Payload: 8.42 MB
Split Complete: 4 Child Documents Created Execution: 2.18s
Compound PDF Pages & Split Boundaries
Cut lines detected: 3 boundaries
PAGE 01 NEW DOC
CCCD Citizen ID
Mặt trước / Sau
Conf: 99.4%
→ ID Verify
PAGE 02 CUT #1
Vehicle Reg.
Đăng kiểm xe cơ giới
Conf: 98.7%
→ Policy Check
PAGE 03 CUT #2
Garage Quote
Báo giá · Page 1/3
MULTI-PAGE LINK
PAGE 04 LINKED
Garage Quote
Báo giá · Page 2/3
Continued schema
PAGE 05 LINKED
SEAL
Garage Quote
Báo giá · Page 3/3
Stitched: 3 Pgs
→ Parts & Labor
PAGE 06 CUT #3
Front Bumper
Damage Photo
Hiện trường tai nạn
Conf: 94.1%
→ Visual Damage
Document Boundary Detected
Multi-Page Continuity Stitched
STP Pass Rate: 100% on this bundle
Classification Inspector Selected: Child Item 03
Detected Archetype
Báo giá sửa chữa ô tô (Garage Repair Quotation)
Span Range
Pages 3 – 5 (3 consecutive sheets stitched)
Multimodal Confidence Breakdown
Visual Layout & Geometry (Table Structure) 98.2%
Vietnamese Semantic NLP (Keywords: Báo giá, Phụ tùng) 96.8%
Seal & Signature Marker Verification 96.5%
Automated Dispatch Target
extraction-schema: auto-parts-labor-v2
Status: Dispatched to Step 04 Extraction Queue
ARCHITECTURE PIPELINE

How Document Classification & Splitting Works

A three-stage deterministic process operating in sub-second intervals to split compound dossiers and route discrete items.

3.1 TRANSITION PASS

Detect & Split Compound Files

The engine evaluates visual transitions, header font weight alterations, stamp positions, and page counter indicators (e.g., "Trang 1/3" switching to "Trang 1/1") to segment monolithic PDFs into independent files.

• Zero blank sheets required
• Boundary accuracy: > 99.1%
• Handles rotation & skewed pages
3.2 MULTIMODAL EVAL

Multimodal Classification

Simultaneously combines spatial visual layout (2D geometry, bounding boxes, seals, photo regions) with dense semantic NLP (Vietnamese terminology, formal legal phrasing) to assign high-confidence archetype labels.

• 50+ out-of-the-box archetypes
• Joint Vision-Language Transformer
• Confidence score generated per item
3.3 ROUTING ENGINE

Automatic Schema Routing

High-certainty documents (confidence ≥ 85%) bypass human staging and immediately trigger their assigned schema model. Lower-confidence items are cleanly flagged for one-click human confirmation in the verification console.

• STP Threshold: configurable ≥ 85%
• Zero manual intervention for verified docs
• Webhook & REST API schema dispatch
CORE COMPETITIVE ADVANTAGES

What Makes TaskTalk Classification Different

Engineered specifically for high-variability Vietnamese document ecosystems where rigid English-first global OCR platforms continuously break.

Pre-Trained on Vietnamese Document Archetypes

Trained on millions of authentic Vietnamese administrative documents: hospital discharge summaries (Giấy ra viện), garage quotes (Báo giá sửa chữa), VAT e-invoices (Hóa đơn GTGT), citizen IDs (CCCD 12 số / gắn chip), and driving licenses (GPLX).

No cold-start template setup required.

Structured, Semi-Structured & Unstructured

Handles everything from fixed-grid government identity cards, semi-structured vendor invoices from over 5,000 corporate formats, to entirely unstructured adjuster narrative reports with freeform paragraphs.

Robust against mobile phone photos and skewed scans.

Self-Improving HITL Classification

Human-in-the-Loop continuous feedback mechanisms capture manual corrections when an edge case is reclassified. Corrections feed into isolated tenant re-training loops to systematically elevate STP accuracy over time.

Zero retraining downtime · Strict enterprise data isolation.
EMPIRICAL PRODUCTION BENCHMARK

Classification in Practice — Auto Insurance Claim Dossier

Real-world execution telemetry for a raw 6-page compound PDF scanned at a commercial claims triage desk.

Page Span Identified Document Class Confidence Splitting & Stitching Action Downstream Routing Model
Page 1 Citizen ID (CCCD Mặt trước/sau) 99.4% Split as independent 1-page document Routed to Identity Verification Model
Page 2 Vehicle Registration (Đăng kiểm xe) 98.7% Split as independent 1-page document Routed to Vehicle Policy Check
Pages 3–5 Garage Repair Quotation (Báo giá 3 trang) 97.2% Stitched as 1 continuous item (3 pages joined) Routed to Parts & Labor Extraction Model
Page 6 Photographic Damage Overview 94.1% Split as independent photographic payload Routed to Visual Damage Assessment Model
Outcome: 1 unorganized 6-page PDF unbundled into 4 distinct validated assets and piped in parallel to 4 dedicated micro-models.
Total Pipeline Step 3 Latency: 2.18 seconds
END-TO-END PIPELINE ARCHITECTURE

What Happens After Classification

Classification & Splitting serves as the critical switching hub of the TaskTalk platform. Explore how each stage links together.

STEP 01 COMPLETED
Document Input

Multi-channel ingestion via Email, SFTP, WhatsApp, Webhooks, or direct REST API.

STEP 02 COMPLETED
OCR & Text Recognition

High-precision diacritic-aware Vietnamese OCR with noise and rotation correction.

STEP 03 YOU ARE HERE
Classification & Splitting

Detect page boundaries, unbundle compound PDFs, and route to specialized models.

STEP 04 NEXT STEP
Data Extraction

Extract key-values, line-item tables, entities, and signatures with schema validation.

STEP 05 DOWNSTREAM
AI Intelligence

Cross-document policy reconciliation, fraud anomaly alerts, and price benchmarks.

STEP 06 DOWNSTREAM
Review & Approval

Confidence-driven human-in-the-loop exception workbench for rapid auditor signoff.

STEP 07 DOWNSTREAM
Processing Analytics

Real-time STP dashboards, model drift monitoring, SLA tracking, and error clustering.

STEP 08 DOWNSTREAM
Export & Integration

Two-way synchronization into Core Banking, SAP, Guidewire, Oracle, and Salesforce.

TECHNICAL DOCUMENTATION FAQ

Frequently Asked Questions

Can TaskTalk split multi-page files without blank separator pages?

Yes. TaskTalk does not rely on physical patch codes, barcodes, or blank separator pages inserted by scanning staff. The engine evaluates visual layout transitions, typography scale shifts, page numbering indicators (e.g., detecting "Page 1 of 3" resets), and semantic topic modeling to autonomously cut and group multi-page documents.

What happens if a document contains multiple pages of the same type?

The continuity engine checks sequential pagination markers, table flow continuity across page boundaries, repeated reference/contract IDs, and signature block presence. When multiple sheets belong to the same multi-page item (such as a 4-page hospital breakdown or a 3-page garage estimate), they are automatically stitched into a single child document entity before extraction.

How many document types does TaskTalk support out-of-the-box?

TaskTalk comes pre-trained with 50+ out-of-the-box Vietnamese document archetypes spanning general insurance, healthcare reimbursement, automotive repair, commercial banking, and municipal tax declarations. Custom categories can also be configured with as few as 20–30 sample documents.

What happens if the AI is uncertain about a document class?

Any document with an archetype confidence score below your configurable straight-through processing threshold (e.g., < 85%) is automatically redirected to the Human-in-the-Loop review queue. Triage personnel can re-classify or verify boundaries with a single click, and their feedback immediately feeds into tenant fine-tuning.

Does TaskTalk classify documents based on visual appearance or text?

TaskTalk employs a Multimodal Vision-Language Transformer that performs both simultaneously. The spatial visual encoder inspects geometrical layouts, logos, red seals, photos, and tabular grids, while the Vietnamese semantic language model parses extracted vocabulary and sentence structure. This dual approach ensures high accuracy even with low-resolution scans.

Can we add our own proprietary document categories?

Yes. Enterprise clients can register custom archetypes through our Administrative Console or API. By providing 20–50 unredacted or synthetic training samples, our few-shot learning pipeline tunes classification weights in an isolated dedicated container within 2 hours.

ZERO MANUAL TRIAGE · UNBUNDLE FILES IN REAL TIME

Stop sorting documents by hand. Let TaskTalk classify and split your files automatically.

Transform unorganized batch uploads into structured, model-ready document queues in less than three seconds. Connect to our enterprise API or schedule a customized technical sandbox demo today.

Direct engineering inquiries: hello@tasktalk.com