# ScholasticAI

> A local, CPU-friendly RAG application that answers questions from uploaded PDFs and OpenAlex using the open Pleias-Pico 350M model.

ScholasticAI is a retrieval-augmented generation application from Pleias that lets users upload local PDFs and ask questions answered with quotes from those documents. It can also query external knowledge databases such as OpenAlex, which the README describes as holding 250M+ documents. It runs locally on a Pleias-Pico (350M) model served through llamafile, and the README says it is designed to run on consumer laptops or desktops without a GPU.

## What It Is

ScholasticAI is a desktop-style Python application for document question answering. Users load PDFs, ask questions in a chat window, and receive generated answers grounded in quotes from the source documents. The README says it was built mainly for research articles but can be used with other kinds of PDFs.

## How It Works

A llamafile server (v0.8.13) loads the Pleias-Pico GGUF model and exposes a local API endpoint at 127.0.0.1:8080. The app is then started with `python -m src.main`. On first launch it downloads the BGE-M3 embedding model. Documents can be added at any time. A fine-tuned YOLOv8 model is used for layout analysis. The README recommends a temperature of 0 for best answer quality.

## The Pleias-Pico Model

Pleias-Pico 0.1 is a 350M-parameter model continued from Pleias-350m on a dataset of about 45 billion tokens of retrieval tasks derived from Common Corpus. According to the README, it is trained on permissibly licensed open data and is compliant with the European AI Act. It supports multilingual input across main European languages. It uses a standardized structure for queries, sources and references, and performs source analysis that acts as an integrated reranker. The README says it was originally conceived as an experimental model, and that it works properly only with prompts matching its RAG design.

## Setup and Requirements

Setup involves cloning the repository, downloading llamafile and the model, installing requirements with pip, and running the app. The README lists Windows or macOS, a Ryzen 5 3000 series, Intel Core i5 (8th gen) or Apple M1 processor, and 8GB of RAM as minimum specs. The interface is designed for bright mode. The README credits the Mozilla Local AI Builders program with supporting development.

## Access and scope

The application repository is public, but it does not specify an application license. Model openness does not automatically license the application. Inference can run locally; downloading models and querying OpenAlex still use external services. The local llamafile endpoint is a dependency, not a public ScholasticAI API. AI Act compliance is a vendor claim, not an independently established legal conclusion.

## Features
- Upload and analyze local PDFs
- Answers grounded in high-fidelity quotes from documents
- Query external knowledge databases such as OpenAlex
- Local CPU inference; external downloads and OpenAlex queries require network access
- Uses open Pleias-Pico 350M RAG model
- Multilingual support for main European languages
- Source analysis with integrated reranking
- Layout analysis with fine-tuned YOLOv8
- BGE-M3 embeddings for retrieval

## Integrations
OpenAlex, llamafile, Hugging Face, BGE-M3, YOLOv8

## Platforms
WINDOWS, MACOS

## Pricing
Free

## Links
- Website: https://github.com/Pleias/pleias_ScholasticAI
- Documentation: https://github.com/Pleias/pleias_ScholasticAI
- Repository: https://github.com/Pleias/pleias_ScholasticAI
- EveryDev.ai: https://www.everydev.ai/tools/scholasticai