DocIntel AI — Enterprise Document Intelligence & Multi-Agent RAG Platform

DocIntel AI is a production-grade, multi-agent document intelligence platform engineered for automated document extractions, structured table parsing, and high-precision Q&A with strict zero-hallucination verification.

The platform combines a multi-agent orchestration architecture (LangGraph) with local AI execution (Ollama) to deliver fast, cost-free, and privacy-preserving document analysis across complex policy documents, financial disclosures, and legal contracts up to 500 pages.

---

Key Features

• Automated Key Highlights Extraction: Automatically parses and categorizes critical document sections (Coverage Scope, Deductibles, Exclusions, Claim Deadlines, and Financial Terms) upon upload.
• 5-Layer Multi-Agent Verification Architecture: Operates a sequential multi-agent execution pipeline (Planner, Router, Retriever, Synthesizer, Verifier) to eliminate hallucinated responses.
• Page-Level & Bounding Box Source Citation: Every generated response includes verifiable citations referencing source pages, section paths, and document chunks.
• Deep Document & Table Parsing: Integrates Docling layout parsing with PaddleOCR for robust structural extraction from scanned PDFs, DOCX files, and spreadsheets.
• 100% Local Inference Engine: Fully compatible with local LLMs (Ollama qwen2.5, llama3.2) and local embedding models (bge-m3), ensuring complete data privacy and 0 API rate limit bottlenecks.
• Industrial Deep Onyx UI: High-contrast, dark-mode dashboard built with Next.js 15, TailwindCSS, and Lucide icons following enterprise design standards.

---

System Architecture

[ Upload Document (PDF / DOCX / XLSX) ]
                  │
                  ▼
[ Ingestion Pipeline: Docling + PaddleOCR ] ──► [ Layout Chunker ]
                                                      │
                                                      ▼
                                           [ Dense Embedding (bge-m3) ]
                                                      │
                                                      ▼
                                           [ Vector Storage: Qdrant ]

-----------------------------------------------------------------------

[ User Query ]
      │
      ▼
[ Planner Agent ] ────► Tentukan Intent & Strategy (QA / Extraction / Clarification)
      │
      ▼
[ Router Agent ] ─────► Route Pipeline Execution
      │
      ▼
[ Retriever Agent ] ──► Hybrid Dense + BM25 Retrieval & Reranking
      │
      ▼
[ Synthesizer Agent ] ► Generate Response + Source Citations
      │
      ▼
[ Verifier Agent ] ───► 5-Layer Fact Checking & Confidence Rating
      │
      ▼
[ Final Response ]

---

Agent Pipeline Overview

1. Planner Agent: Analyzes user intent, checks query clarity, and decomposes complex queries into optimal retrieval sub-queries.
2. Router Agent: Dispatches execution paths dynamically based on query type (RAG QA, Structured Extract, or Document Overview).
3. Retriever Agent: Executes hybrid dense vector search and sparse BM25 keyword matching with Reciprocal Rank Fusion (RRF).
4. Synthesizer Agent: Generates concise, grounded responses strictly constrained to retrieved context chunks with source citations.
5. Verifier Agent: Validates claims against source passages, assigns confidence scores (High/Medium/Low), and suppresses ungrounded assertions.

---

Technology Stack

Backend

• Framework: Python 3.11, FastAPI
• Agent Orchestration: LangGraph, LangChain
• Vector Database: Qdrant Vector Search
• Metadata Database: SQLite (Async SQLAlchemy + AIOSQLite)
• Document Parsing: Docling, PaddleOCR, OpenPyXL, PyMuPDF
• Local AI Engine: Ollama (bge-m3, qwen2.5:0.5b, llama3.2)

Frontend

• Framework: Next.js 15 (App Router), React 19, TypeScript
• Styling: TailwindCSS (Deep Onyx Theme)
• Icons: Lucide React
• HTTP Client: Native Fetch API

---

Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• Ollama (installed locally)
• Qdrant Vector Engine (or Docker)

1. Local AI Engine Setup (Ollama)

Ensure Ollama is running and pull the required models:

ollama pull bge-m3
ollama pull qwen2.5:0.5b

2. Backend Setup

cd backend
python -m venv .venv
# On Windows:
.\.venv\Scripts\activate
# On Linux/macOS:
source .venv/bin/activate

pip install -e .

Configure environment settings:

cp .env.example .env

Start the FastAPI application:

uvicorn app.main:app --reload --port 8000

3. Frontend Setup

cd frontend
npm install
npm run dev

Open http://localhost:3000 in your web browser.

---

API Endpoints Reference

Document Management

• POST /api/documents/upload — Upload PDF/DOCX/XLSX files for background parsing.
• GET /api/documents — List all ingested documents and processing statuses.
• GET /api/documents/{id}/highlights — Retrieve extracted key highlights, categories, and tables.
• DELETE /api/documents/{id} — Delete document record and vector points.

Q&A & Extractions

• POST /api/query — Execute multi-agent RAG Q&A pipeline with source citations.
• POST /api/extract — Perform schema-driven structured JSON extraction.

---

License

Distributed under the MIT License. See LICENSE for details.