---

name: paperai description: "Guides paperai workflows for indexing paper corpora, querying scientific articles, and generating RAG-backed Markdown, CSV, or PDF-annotation reports." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

paperai

paperai is a Python application for searching and analyzing scientific and medical paper corpora. It combines a paperetl-style SQLite database with a txtai embeddings index and optional LLM/RAG pipelines. Use this skill when a task asks to build a paperai index, search indexed articles, expose enriched search results, or generate repeatable reports.

Install and inspect

Use Python 3.10+ and install the public package:

python -m pip install "paperai==2.6.0"
python -m pip check
# from the generated paperai skill directory
python scripts/check_install.py

The last command is a bundled, no-download import check; it does not load model weights or validate a corpus. Read references/install.md for dependency and runtime boundaries, and references/troubleshooting.md for cross-cutting failures.

Route by task

• Build or inspect an index, validate articles.sqlite, train static vectors, or export section text: read indexing.
• Run one-shot searches, use paperai shell, integrate the enriched API, or make a search UI: read querying.
• Define YAML tasks and produce Markdown/CSV/PDF-annotation reports: read reporting.

A normal end-to-end handoff is indexing → querying or indexing → reporting. The generated skill contains the needed API details and safe helpers; do not reopen the original repository's source, examples, tests, or checkout-specific paths.

Shared model-directory contract

Most workflows receive one directory containing articles.sqlite and saved txtai artifacts (config or config.json plus model/index files). The database is normally produced by paperetl and is not created by paperai. Validate its tables and columns before model construction. Model identifiers can trigger network downloads and optional GPU use; state the cache, device, data, and runtime constraints before starting a large operation.

Verification boundary

A successful import proves package plumbing only. Validate actual model loading, SQLite schema compatibility, retrieval quality, report rendering, and optional PDF annotation separately. Keep maxsize, toprank, topn, and report context bounded for smoke runs. Markdown/CSV are the portable default; ant annotation requires original PDFs and additional runtime support.

Read references/repo-provenance.md before using this skill with a changed checkout. If the commit, package version, or evidence paths differ, refresh the repo skill instead of assuming the guidance is current.