Motor de Inteligência de Documentos Agentico com Cross-Reference
gayathrikumari/Agentic-Document-Intelligence-Industrialized-Cross-Reference-Engine
Automação inteligente de cross-referencing de documentos técnicos com validação determinística
Ver no GitHubSistema multi-agente em nível de produção que automatiza a análise cruzada de documentos técnicos usando Model Context Protocol (MCP) e padrões ReAct. Integra RAG multimodal com Vision-Language Models, validação via LLM-as-a-Judge e otimizações de inferência com KV Caching, alcançando 15% de melhoria na precisão de dados.
Instalação
npx convez-ai gayathrikumari/Agentic-Document-Intelligence-Industrialized-Cross-Reference-Engine1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditoria e validação cruzada de documentos de engenharia
- Extração de dados de blueprints e especificações técnicas
- Automação de processos de revisão manual de grandes repositórios
Veredito

Implementação sólida de arquitetura agentica com MCP e ReAct para resolver problemas reais de processamento de documentos técnicos. O foco em validação determinística (G-Eval) e otimização de inferência mostra profundidade técnica, embora requeira expertise significativa em agentes de IA e pipelines RAG para implementação.
README
A tradução para o português deste README ainda está na fila.
Agentic Document Intelligence & Industrialized Cross-Reference Engine
A production-grade multi-agent system leveraging the Model Context Protocol (MCP) to connect Frontier LLMs with legacy engineering repositories, automating complex cross-referencing tasks that previously required extensive manual human review.
🚀 Project Overview
This project demonstrates an end-to-end Agentic Document Intelligence pipeline built to automate engineering document cross-referencing using MCP, ReAct-based Multi-Agent Patterns, Multimodal RAG with Vision-Language Models, and an LLM-as-a-Judge (G-Eval) validation framework — achieving a 15% increase in data accuracy while eliminating manual audit bottlenecks.
📁 Repository Structure
| File | Description |
|---|---|
1_mcp_server_setup.py | MCP server exposing PDF blueprints & Excel repositories as standardized queryable resources |
2_react_agent.py | ReAct-based Sequential Multi-Agent Pattern with dedicated Extractor & Verifier agents |
3_multimodal_rag.py | Multimodal RAG pipeline using Vision-Language Models (VLMs) & semantic chunking with ChromaDB |
4_llm_judge_eval.py | LLM-as-a-Judge (G-Eval) framework for deterministic validation of agent outputs |
5_inference_optimization.py | KV Caching + temperature/top-p tuning for latency reduction and controlled generation |
requirements.txt | All dependencies required to run this project |
🛠️ Tech Stack
- Agentic Framework: Model Context Protocol (MCP), ReAct (Thought-Action-Observation), Sequential Multi-Agent Patterns
- LLM: Claude (Anthropic) — claude-sonnet-4-20250514
- RAG Pipeline: Multimodal RAG, Vision-Language Models (VLMs), LangChain, ChromaDB, Semantic Chunking
- Evaluation: LLM-as-a-Judge (G-Eval), Deterministic Validation
- Inference Optimization: KV Caching (Prompt Caching), Temperature / Top-k / Top-p Tuning
- Document Processing: PyMuPDF (fitz), OpenPyXL
- Access Control: MCP Roots for Sandboxed Resource Access
📊 Key Results
- ✅ Achieved a 15% increase in data accuracy through deterministic G-Eval validation
- ✅ Fully eliminated manual audit bottlenecks across high-volume engineering schematics
- ✅ Standardized access to legacy PDF blueprints and Excel repositories via MCP interface
- ✅ Reduced inference latency through KV Caching and optimized decoding configurations
- ✅ Preserved spatial relationships in technical drawings using Vision-Language Model analysis
⚙️ Setup & Installation
# Clone the repository
git clone https://github.com/gayathrikumari/Agentic-Document-Intelligence.git
# Navigate to the project directory
cd Agentic-Document-Intelligence
# Install dependencies
pip install -r requirements.txt
# Set your Anthropic API key
export ANTHROPIC_API_KEY=your_api_key_here
📂 Data Setup
Create the following folder structure before running:
data/
├── blueprints/ ← Place your PDF engineering blueprints here
│ ├── blueprint1.pdf
│ └── blueprint2.pdf
└── master_book.xlsx ← Place your Excel master book here
🔢 Run Order
Execute the scripts in the following order for the full pipeline:
# Step 1: Start the MCP server
python 1_mcp_server_setup.py
# Step 2: Run the ReAct Multi-Agent pipeline
python 2_react_agent.py
# Step 3: Ingest documents and test RAG queries
python 3_multimodal_rag.py
# Step 4: Validate agent outputs with G-Eval
python 4_llm_judge_eval.py
# Step 5: Benchmark inference configurations
python 5_inference_optimization.py
🏗️ Architecture
PDF Blueprints + Excel Master Book
↓
MCP Server (1_mcp_server_setup.py)
[Standardized Resource Interface]
↓
ReAct Multi-Agent Pipeline (2_react_agent.py)
[Extractor Agent → Verifier Agent]
↓
Multimodal RAG (3_multimodal_rag.py)
[VLM Image Analysis + Semantic Chunking + ChromaDB]
↓
G-Eval Validation (4_llm_judge_eval.py)
[Deterministic LLM-as-a-Judge]
↓
Inference Optimization (5_inference_optimization.py)
[KV Caching + Temperature Tuning]
👩💻 Author
Gayathri Kumar
Data Scientist & AI Engineer | Ex-Deloitte | MS Information Science @ UNT
📧 gayathrikumar.tx@gmail.com
🔗 [(Linkedin)](#) | GitHub