🧾 OmniExtract: Enterprise AI Invoice Data Extractor

Intelligent. Resilient. Fully Automated.

An enterprise-grade, fault-tolerant document processing and auditing system built in n8n. OmniExtract automates the ingestion, extraction, deterministic auditing, compliance classification, duplicate prevention, and archiving of complex invoice files at scale.

---

📚 Documentation & Guides

Whether you are a developer deploying the system or an operator maintaining it, everything you need is documented:

• 🚀 Zero-to-Hero Installation Guide - Step-by-step GCP setup and n8n Docker deployment.
• 🛡️ Enterprise Security Audit - Learn how this architecture scales securely.
• 🛠️ Day 2 Maintenance Protocol - Monitoring, backup, and safe update strategies.

---

🌟 Key Capabilities

• ✅ Multimodal Ingestion: Dynamically processes clean text-based PDFs, scanned image-based PDFs, and direct JPEGs/PNGs.
• ✅ Claude 3.5 Sonnet Engine: Extracts complex nested tables and metadata with custom multimodal prompting and 3-retry exponential backoff protection.
• ✅ Deterministic Auditing: Performs strict mathematical assertion checks on line items vs. subtotals.
• ✅ Compliance Triage: Classifies invoices dynamically into AUTO_APPROVED 🟢, REVIEW_RECOMMENDED 🟡, and REVIEW_REQUIRED 🔴.
• ✅ Zero-Duplicate Architecture: Auto-diverts duplicate submittals to quarantine folders while keeping the master Google Sheet pristine.
• ✅ Dead-Letter Telemetry: Logs granular processing durations and Claude token usage to dedicated audit tracking sheets.

---

📊 Enterprise Pipeline Architecture

The following diagram illustrates the 5 distinct operational layers within the n8n pipeline.

flowchart TD
    %% Styling definitions
    classDef layerTrigger fill:#f3e8ff,stroke:#9333ea,stroke-width:2px,color:#000;
    classDef layerIngest fill:#e0f2fe,stroke:#0284c7,stroke-width:2px,color:#000;
    classDef layerAI fill:#ffedd5,stroke:#ea580c,stroke-width:2px,color:#000;
    classDef layerAudit fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#000;
    classDef layerOutput fill:#ffe4e6,stroke:#e11d48,stroke-width:2px,color:#000;

    %% 1. TRIGGER LAYER
    subgraph Trigger_Layer [1. Trigger Layer]
        A[Gmail Attachment Trigger]
        B[Google Drive File Trigger]
    end
    class Trigger_Layer,A,B layerTrigger;

    %% 2. INGESTION LAYER
    subgraph Ingestion_Layer [2. Ingestion & Routing Layer]
        C[Detect File Type]
        D{Format Switch}
        E[Extract PDF Text]
        F{Length > 50 chars?}
        G[Convert PDF to JPEG]
        H[Binary to Base64]
    end
    class Ingestion_Layer,C,D,E,F,G,H layerIngest;

    %% 3. AI PROCESSING LAYER
    subgraph AI_Layer [3. AI Processing Layer]
        I[Prepare Claude Payload & Truncate]
        J[Claude 3.5 Sonnet Vision API]
        K[Parse JSON Response]
    end
    class AI_Layer,I,J,K layerAI;

    %% 4. VALIDATION & COMPLIANCE LAYER
    subgraph Validation_Layer [4. Validation Layer]
        L[Validate Math & Line Items]
        M[Validate Required Fields & Credit Notes]
        N[Determine Status Color Code]
        O{Is Duplicate?}
    end
    class Validation_Layer,L,M,N,O layerAudit;

    %% 5. OUTPUT & FILE ARCHIVING LAYER
    subgraph Output_Layer [5. Output & Archiving Layer]
        P[Google Sheets Append]
        Q{Status Check}
        R[Send Slack/Email Alert]
        S[Archive Processed File]
        T[Append to Audit_Log / Error_Log]
        U[Quarantine Duplicate File]
    end
    class Output_Layer,P,Q,R,S,T,U layerOutput;

    %% Routing logic
    A & B --> C --> D
    
    D -- "PDF" --> E --> F
    F -- "Text Heavy" --> I
    F -- "Scanned" --> G --> H --> I
    D -- "Image" --> H
    
    I --> J --> K
    
    K --> L --> M --> N --> O
    
    O -- "NO" --> P
    O -- "YES" --> U
    
    P --> Q
    Q -- "Review Needed" --> R
    
    P --> S --> T

Loading

---

📸 Visual Demo Assets (Strategy)

To make this repository visually compelling for stakeholders, place the following screenshots in the /assets/ directory (these are currently placeholders to be updated by the developer):

1. assets/01-n8n-canvas.png (The Brain): A wide, zoomed-out shot of the beautiful 26-node n8n canvas demonstrating the scale and visual branching of the logic.
2. assets/02-google-sheets-dashboard.png (The Output): A close-up of the Sheet1 ledger. Showcase the conditional formatting—rows glowing Green (Auto-Approved) alongside Red (Review Required) rows to prove the triage works.
3. assets/03-claude-multimodal.png (The Intelligence): A split-screen shot. On the left, a blurry/scanned PDF invoice. On the right, the perfectly structured JSON output generated by Claude 3.5 Sonnet.
4. assets/04-duplicate-quarantine.png (The Safety Net): An email screenshot of the high-priority "Duplicate Alert" showing that the system successfully caught and blocked a double-submittal.

Note to Junior Dev: Take these 4 screenshots on your local deployment and update the image tags below!

Platform Previews

---

📜 License

Licensed under the MIT License.