Visual Schema-Driven Chunked Data Extractor

An interactive, visual schema editor and text-extraction application. This project pairs a web-based Cappuccino (Objective-J) desktop frontend with a Mojolicious (Perl) backend to extract structured, nested data from unstructured text using Large Language Models (LLMs) and automatically resolve terminology via dense retrieval.

---

System Architecture

┌──────────────────────────────────────┐          ┌───────────────────────────┐
│         Objective-J Frontend         │          │    Mojolicious Backend    │
│  ┌────────────────────────────────┐  │  JSON    │  ┌─────────────────────┐  │
│  │ CPOutlineView (Visual Schema)  │  ├─────────►│  │ Chunking & Merging  │  │
│  └────────────────────────────────┘  │  Request │  └──────────┬──────────┘  │
│  ┌────────────────────────────────┐  │          │             ▼             │
│  │ CPTextView (Text Highlights)   │  │◄─────────┤  ┌─────────────────────┐  │
│  └────────────────────────────────┘  │  JSON    │  │ Schema Validation   │  │
│  ┌────────────────────────────────┐  │  Response│  │  & Self-Correction  │  │
│  │ CPTableView & CSV Export       │  │          │  └──────────┬──────────┘  │
│  └────────────────────────────────┘  │          │             ▼             │
└──────────────────────────────────────┘          │      Dense Retrieval      │
                                                  │    Vector Search Lookup   │
                                                  │             ▼             │
                                                  │       Target LLM API      │
                                                  └───────────────────────────┘

---

Features

• Graphical Schema Tree Editor: Create, modify, and nest schema fields dynamically (supporting string, number, array, and object types).
• Dense-Retrieval Integration: Seamlessly maps raw text extractions to canonical database entries. Extracted strings are resolved using semantic dense vector searches (top_k=1) against user-selected catalog systems to retrieve unified identifiers.
• Automatic Chunking & Merging: Automatically splits long inputs to match context window limitations and merges the results into a single structured output.
• Source Text Highlighting: Color-coded visualization mapping extracted fields back to their exact coordinates in the source text.
• Self-Correction Loop: Validates JSON outputs against the user-defined schema and performs corrective retry prompts if type or structure mismatches occur.
• Data Export: Instant tabular viewing and export of resolved paths and values as a CSV spreadsheet.

---

Getting Started

1. Prerequisites

• Perl 5.20+
• An active LLM connection (OpenAI, local Ollama, or any OpenAI-compatible gateway)
• An endpoint hosting embedding indices for semantic vector search (such as the Patchbay API gateway)

2. Setup

Clone the repository and install the Perl dependencies:

# Install Mojolicious and dependencies
cpanm Mojolicious::Lite Encode File::Spec

Set up your environment variables for your chosen LLM endpoint and dense-retrieval vector database:

export VLLM_API_KEY="your-api-key"
export VLLM_ENDPOINT="https://your-llm-gateway.example.com/v1/chat/completions"
export VLLM_MODEL="gpt-oss-120b"
export PATCHBAY_URL="http://10.210.21.201:3036"

Start the Mojolicious development server (defaults to port 3000):

morbo ./backend.pl  --listen "http://*:3000"

Open http://localhost:3000/Frontend/index.html in your web browser to run the application.

---

Usage Workflow

1. Input Text: Paste unstructured text into the left pane.
2. Configure Schema: Define keys and structural hierarchy using the Schema Tree Editor on the right.
3. Enable Dense Retrieval: Assign specific fields to lookup targets (vector collections) to resolve raw extracted text to standard, coded database equivalents.
4. Extract: Click Extract Structured Chunks. The application chunks the text, runs extraction, validates consistency, queries the dense-retrieval API, and returns highlighted results.
5. Export: Click Export Results as CSV to download a spreadsheet of the resolved extraction records.