📄 Document Intelligence Crew

!PineCone Vector DB https://app.pinecone.io/organizations

A production-grade multi-agent AI system for enterprise contract analysis. Upload vendor contracts (PDF, DOCX, XLSX) and get instant risk assessment, compliance flags, and executive recommendations — powered by Agentic RAG and a 5-agent CrewAI pipeline.

🏗️ Architecture

📄 Documents (PDF + DOCX + XLSX) ↓ [RAG Pipeline] → chunk → embed → Pinecone Vector DB ↓ [RAG Agent] → autonomous search + rerank + reflect ↓ [Extractor] → ExtractedDoc (Pydantic) ↓ [Analyst] → AnalysisResult (Pydantic) ↓ [Action Planner] → ActionPlan (Pydantic) ↓ [Summary Agent] → Executive Report ↓ [FastAPI] → REST endpoints [Streamlit] → Interactive UI

✨ Key Features

• Multi-Agent Orchestration — 5 specialized CrewAI agents with typed handoffs
• Agentic RAG — autonomous retrieval with re-ranking and self-reflection
• Multi-Format Ingestion — PDF, DOCX, XLSX document support
• Pydantic Typed Contracts — strongly typed inter-agent schemas
• Vector Semantic Search — Pinecone VectorDB + Google text-embedding-001
• Re-ranking — cross-encoder scoring for precision retrieval
• Self-Reflection — agent evaluates its own retrieval completeness
• REST API — FastAPI with 4 endpoints + Swagger docs
• Interactive UI — Streamlit with file upload and live agent traces
• PII Detection — sensitive data flagged in contract analysis
• Structured Logging — timestamped log files per run

🛠️ Tech Stack

| Layer | Technology | |---|---| | Agent Orchestration | CrewAI | | LLM | Google Gemini 2.5 Flash | | Embeddings | Google gemini-embedding-001 | | Vector Store | Pinecone VectorDB | | RAG Strategy | Agentic RAG + Re-ranking + Reflection | | Typed Contracts | Pydantic AI | | Backend API | FastAPI + Uvicorn | | Frontend UI | Streamlit | | Document Parsing | PyMuPDF, python-docx, openpyxl | | Observability | Structured logging | | Language | Python 3.12 |

📁 Project Structure

doc-intelligence-crew/ ├── agents/ # CrewAI agent definitions │ ├── rag_agent.py # Agentic RAG with search + rerank + reflect │ ├── extractor_agent.py # Contract data extraction │ ├── analyst_agent.py # Risk classification │ ├── action_agent.py # Action planning │ └── summary_agent.py # Executive summary ├── tasks/ # Task definitions per agent ├── schemas/ # Pydantic typed inter-agent contracts │ └── contract_schemas.py # ExtractedDoc, AnalysisResult, ActionPlan ├── tools/ # CrewAI tools │ ├── search_tool.py # Pinecone VectorDB semantic search │ ├── reranker_tool.py # Cross-encoder re-ranking │ └── reflection_tool.py # Coverage self-reflection ├── config/ # Configuration │ └── llm_config.py # LLM + PineCone VectorDB setup ├── pipeline/ # RAG ingestion pipeline │ └── rag_pipeline.py # Load → chunk → embed → store ├── api/ # FastAPI routes │ └── routes/ │ ├── analyze.py # POST /analyze │ ├── documents.py # GET /documents, POST /upload │ └── health.py # GET /health ├── data/ # Contract documents (PDF, DOCX, XLSX) ├── logs/ # Timestamped run logs ├── app.py # FastAPI entry point ├── streamlit_app.py # Streamlit UI └── orchestrator.py # Full crew orchestration

🚀 Quick Start

Prerequisites

• Python 3.12+
• Google Gemini API key (get one here)

Installation

# clone the repo
git clone https://github.com/YOUR_USERNAME/doc-intelligence-crew.git
cd doc-intelligence-crew

# create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# install dependencies
pip install -r requirements.txt

Configuration

# create env file
mkdir env
cp .env.example env/.env

# add your API keys
GEMINI_API_KEY=your-gemini-api-key
GOOGLE_API_KEY=your-gemini-api-key
PINECONE_API_KEY-your-pinecone-api-key
PINECONE_INDEX=doc-intelligence

Run

# Step 1 — ingest documents into PineCone VectorDB
python pipeline/rag_pipeline.py

# Step 2 — start FastAPI backend
uvicorn app:app --reload --port 8000

# Step 3 — start Streamlit UI (new terminal)
streamlit run streamlit_app.py

Open 👉 http://localhost:8501

🔌 API Endpoints

| Method | Endpoint | Description | |---|---|---| | GET | /health | System health check | | GET | /documents | List ingested documents | | POST | /documents/upload | Upload new contracts | | POST | /ingest | Re-ingest all documents | | POST | /analyze | Run full 5-agent analysis |

Interactive docs: http://localhost:8000/docs

📊 Results

| Metric | Value | |---|---| | Documents supported | PDF, DOCX, XLSX | | Agents in pipeline | 5 | | Avg retrieval precision | High (re-ranked) | | Contracts analyzed | 6 vendor contracts | | Total contract value analyzed | $25.1M | | Risk findings identified | 35+ across all contracts | | PII elements detected | SSN, Email, Phone, Bank Account |

🤖 Agent Pipeline

| Agent | Role | Output Schema | |---|---|---| | RAG Agent | Autonomous retrieval + rerank + reflect | Raw context | | Extractor Agent | Structured data extraction | ExtractedDoc | | Analyst Agent | Risk classification by severity | AnalysisResult | | Action Agent | Prioritized action planning | ActionPlan | | Summary Agent | Executive brief | Plain prose |

👩‍💻 Author

Dipti Joshi Senior Software Engineer → Agentic AI Engineer LinkedIn | GitHub