AI-ML2025

FlowDocs

Modular Agentic Document Intelligence Platform

FastAPIReact 19TypeScriptChromaDBOllamaPaddleOCRPyMuPDF

Discuss this project

Video Demonstration

Three live scenarios — document parsing, corpus topic clustering & RAG QA chat

FlowDocs Architecture Pipeline

Visual DAG • Parsing • Topic Clustering • RAG QA

Pipeline Live in Demo

Step 01

Document Ingestion

Input Stream

Asynchronously accepts multilingual inputs (Arabic, French, English) via FastAPI upload streams.

Step 02

3-Tier OCR & Extraction

Multilingual Engine

Performs layout analysis, tabular data parsing, and multilingual character recognition with high confidence.

Step 03

Vector Embeddings

ChromaDB + Ollama

Chunks content into semantically coherent vectors stored in local ChromaDB for offline retrieval.

Step 04

ORCA Critic Agent

Anti-Hallucination

Cross-references claims against source chunks, rejecting ungrounded inferences to suppress hallucination.

Step 05

Live Execution & Visual DAG

FastAPI + React 19

Executes typed-node graphs in asynchronous parallel waves with cycle detection, streaming live execution logs and answers to the React 19 visualizer.

System Key Advantage:Guarantees zero external cloud reliance through fully local Ollama inference, while the Critic Agent suppresses hallucinations by enforcing source verification across extracted tabular and text evidence.

Screenshots & Interface

The Challenge

Organizations deal with multilingual documents requiring intelligent extraction, analysis, and Q&A — but cloud AI tools leak sensitive data, hallucinate without verification, and treat AI as a black box.

Engineered Solution

Built a typed-node workflow engine (42 node types, plugin auto-discovery) with a custom SVG DAG editor, 3-tier multilingual OCR (PaddleOCR/Tesseract/EasyOCR), PyMuPDF PII redaction, TF-IDF/K-Means clustering, and offline RAG over ChromaDB + Ollama — all running 100% locally.

Outcome & Impact

Made agentic AI tangible as typed, composable, verifiable workflows — 100% local processing with no cloud and no API keys, shipped as TAIS NextGen's Project-1 and verified end-to-end with Playwright.

Technical Deep Dive

A visual, no-code document-intelligence platform where anyone composes OCR, parsing, RAG and clustering into drag-and-drop node graphs — 42 typed node types with plugin auto-discovery, executed by a Kahn-scheduling engine with parallel waves. Built around a 100% local vision: no cloud, no API keys, data never leaves the machine — 3-tier OCR (PaddleOCR/Tesseract/EasyOCR), RAG with Ollama nomic-embed-text over ChromaDB, TF-IDF/K-Means topic clustering, and PyMuPDF PII redaction. Includes an ORCA-inspired Critic Agent that cross-references generated claims against vector-retrieved source evidence to suppress hallucinations. Shipped as TAIS NextGen's Project-1 with Docker Compose and Playwright end-to-end tests. The demo walks through three end-to-end scenarios: parsing a 10-K document into complete text and structured data, clustering a document corpus by topic with exportable results (JSON/CSV), and RAG-powered QA chat answering questions over 536 indexed chunks.

PreviousTAIS NextGen

NextTextile ERP