AI / LLMOps2025

DontREADME

Smart RAG chatbot for conversational document analysis with built-in Prefect orchestration.

Mistral AIRAGChromaDBPrefectGradioLangChainPython

Concept & Architecture

DontREADME is a conversational platform that lets you query any document (PDF, DOCX, TXT) through a chatbot powered by Mistral AI. Problem solved: Reading and analyzing long documents (theses, reports, technical documentation) is time-consuming. DontREADME provides precise, sourced answers in a few seconds. RAG architecture (Retrieval-Augmented Generation): 1. Ingestion — The document is split into smart chunks while preserving its structure 2. Vectorization — Embeddings generated and stored in ChromaDB 3. Retrieval — Hybrid semantic search (dense + sparse) with reranking 4. Generation — Mistral AI generates a contextual answer with source citations Tech stack: • Backend: Python, LangChain, ChromaDB • LLM: Mistral AI (mistral-tiny) via API • Interface: Gradio with a custom theme (dark gray / orange) • Orchestration: Prefect for automated workflows
3
Supported formats
<5s
Processing time
~90%
RAG accuracy
5
Templates
Configuration interface
1 / 2
Document tab: upload, chunking settings, template selection

Advanced Features

The application goes beyond a simple chatbot with production-oriented features. Smart document processing: • Automatic detection of the document type (academic, technical, legal, general) • Adaptive chunking that preserves the structure (headings, paragraphs) • Keyword and metadata extraction • Specialized prompt templates based on the detected context Conversation management: • Conversation memory with contextual history • Real-time metadata (template used, sources consulted, performance) • Session export to JSON for later analysis Monitoring and performance: • Real-time system metrics (CPU, RAM, disk) • Per-request performance tracking (latency, tokens, success) • Built-in system health dashboard
4
Detected types
Contextual
Memory
JSON
Export
Real time
Monitoring

Prefect Orchestration

Prefect integration turns DontREADME into a complete MLOps platform with automated workflows. 12 Prefect workflows, grouped into 4 families: • Health Check — Full system diagnosis (CPU, RAM, ChromaDB, directories) • Database Maintenance — ChromaDB backup, optimization and cleanup • Smoke Tests — Quick functional validation with the Mistral API • Batch Processing — Automated ingestion of document folders Orchestration architecture: • Prefect Server mode — Scheduling, persistence, web interface • Direct mode (fallback) — Synchronous execution without a dedicated server • Built-in Gradio interface — One-click actions, real-time tracking Observability: • Structured logs with levels (DEBUG, INFO, WARNING, ERROR) • Run history with durations and statuses • Automatic alerts and recommendations
12
Workflows
>95%
Success rate
3 docs
Parallelism
Built-in
Interface

User Interface

Clean Gradio interface with a professional visual identity (dark gray / orange). Document tab: • File upload (PDF, DOCX, TXT) with automatic validation • Chunking configuration (size, overlap) • Selection of the number of retrieved documents (k) • Template choice (auto, general, academic, technical, legal) • Display of processing information and performance Smart Chat tab: • Conversational interface with history • Real-time answers with cited sources • Metadata for each answer (template, sources, latency) • Export of the full session Prefect Orchestration tab: • System dashboard • One-click quick actions • Execution and tracking of custom workflows • Detailed logs and diagnostics
3
Tabs
Custom
Theme
Cited
Sources
Simplified
UX
Chat interface
Smart Chat tab: conversation with answer metadata
  _____  _____ _____
 |  __ \|  __ \_   _|
 | |__) | |__) || |
 |  _  /|  ___/ | |
 | | \ \| |    _| |_
 |_|  \_\_|   |_____|

Self-hosted on Raspberry Pi 5

~/links

$ echo "Made with passion"

© 2026 Abdel Touati