paper-scoring-connectors

Optional data-source, document, related-work, and assessment-provider adapters for paper-scoring-core. The package contains no credentials, reads no .env files, and performs no work at import time. Applications inject configuration when they create an adapter.

Installation

Install only the integrations that an application needs:

python -m pip install "paper-scoring-connectors[arxiv,pdf,openalex]"
python -m pip install "paper-scoring-connectors[openai]"
python -m pip install "paper-scoring-connectors[ollama]"

openalex intentionally adds no third-party dependency; it uses the Python standard library. all installs every optional SDK. Dependencies between paper-scoring distributions use normal PyPI version constraints and never Git or local filesystem references.

Examples

Search arXiv without exposing SDK objects to the rest of the application:

from datetime import date

from paper_scoring_connectors.arxiv import ArxivSource

documents = ArxivSource().search("cs.AI", date(2026, 1, 15), max_results=10)

Lower-level ArxivClient, QueryBuilder, and download_pdf APIs are available when an application needs control over individual steps. Batch pipelines should call ArxivSource.search_with_failures(...): it isolates each download/extraction failure and returns ArxivBatchResult(successes=..., failures=...). Failures contain only the paper ID, title, and exception type; provider messages are never retained.

Extract a local PDF into the core document contract:

from paper_scoring_connectors.pdf import PdfDocumentLoader

document = PdfDocumentLoader().load("paper.pdf", title="A paper")

Create an assessment adapter with an arbitrary model identifier (there is no model enum):

from paper_scoring_connectors.providers import create_assessment_model

model = create_assessment_model("openai", "your-model-id", api_key="injected-at-runtime")

For Ollama, pass host= and any client options explicitly. SDK defaults may consume process environment variables that the caller injects at runtime; this package never parses .env files.

Supported contract

Version 0.1 targets these paper-scoring-core imports:

from paper_scoring.models import AssessmentRequest, AssessmentResponse, PaperDocument, PageText
from paper_scoring.ports import AssessmentModel, RelatedWorkRetriever
from paper_scoring.rubric import render_assessment_prompt

The core contract is the only mandatory runtime dependency. Each connector is independently importable, and provider SDK imports remain lazy.

Development

python -m pip install -e ".[all,test,security]"
pytest
ruff check .
mypy
bandit -c pyproject.toml -r src
python -m build

All functional tests are offline and use injected fake transports or clients. See docs/report-templates/ for the release evidence templates populated by the release workflow.