Skip to content

Latest commit

 

History

History
52 lines (40 loc) · 2.95 KB

HIGH_LEVEL_HANDOFF.md

File metadata and controls

52 lines (40 loc) · 2.95 KB

ScaffoldAI — High‑Level Research Handoff (Concise)

This document gives a plain, high‑level overview for instructors and reviewers. It avoids external links and numeric results and uses only the correct local figure.

Purpose

  • Assist educators in designing sustainability and climate‑resilience curriculum by turning course PDFs and questions into sourced, explainable suggestions.

What the system does

  • Ingests course materials (PDFs) and prepares them for retrieval.
  • Answers questions with relevant excerpts and clear citations from the uploaded materials.
  • Provides a simple classroom‑ready Windows application (no Python required for teachers).

How it works (at a glance)

  • Chunk and embed syllabus/reading PDFs into a vector index (FAISS).
  • Retrieve top‑matching passages for a user question and assemble context.
  • Run a small language model to draft an answer grounded in those passages.
  • Return the answer with citations back to the user interface.

Methods (figure)

Methods Diagram

What ships in this repository

  • Teacher application (Windows folder build) plus example index files under vector_outputs.
  • Core processing modules in scaffold_core/ (PDF processing, retrieval, LLM, citations).
  • Enhanced Flask UI under frontend/.

Running options (one‑sentence pointers)

  • Teachers: use the zipped Windows app folder and follow TEACHER_ZIP_RUN.md in this repository.
  • Developers: use Python 3.12, install requirements, and start the Enhanced UI.

Notes on access

  • Some models may require a Hugging Face token. A temporary token can be placed in .env.local next to the app; long‑term use should employ a personal token.

Maintenance pointers

  • When adding new PDFs, re‑create the vector index to keep retrieval accurate.
  • Prefer small, well‑cited answers in the UI; review and refine prompts as needed.

Next steps (concise)

  • Data corpus curation: consolidate sources, set a refresh cadence, record provenance.
  • Index maintenance: scheduled re‑embedding, version FAISS/metadata, verify integrity.
  • Retrieval quality: deduplicate by DOI/URL, tune top‑k, consider a lightweight reranker.
  • Evaluation protocol: periodic sampling with a rubric and a brief written summary.
  • Model governance: define model selection policy, manage tokens, document fallbacks.
  • UX and accessibility: ensure keyboard navigation, contrast compliance, and citation export.
  • Packaging and distribution: reproducible signed builds, checksum files, smoke‑test matrix.
  • Privacy and security: minimize PII, manage .env securely, publish a usage policy.
  • Observability: structured logs, latency breakdowns, simple health checks.
  • Performance: preload models, enable ONNX/quantization where appropriate, cache hot queries.
  • Teaching deployment: provide a one‑page quick start and classroom network guidance.
  • Documentation: keep this handoff and the teacher guide synchronized at release.