sci-etl-core

sci-etl-core is a Python library for turning scientific papers from any field into structured, searchable data. It fetches papers from arXiv, PubMed, OpenAlex, and Semantic Scholar, reads their full text, extracts the values you ask for with an LLM, and keeps the papers searchable on your own machine.

udg-catalogue shows the result. It screens astrophysics papers on arXiv, extracts measurements of ultra-diffuse galaxies, and publishes a cross-matched catalog of 1,285 objects, with keyword and semantic search over the papers behind it. sci-etl-core supplies the fetching, parsing, extraction, caching, resumable state, and search, while udg-catalogue adds the astronomy: prompts, validation rules, sky-position matching, and the dashboard.

Nothing in the library is tied to astronomy. Field knowledge lives in your prompts, validators, and normalizers, so the same building blocks work for any field of science.

Documentation: https://xueromll.github.io/sci-etl-core/

Motivation

I built sci-etl-core while working with scientific papers during my undergraduate physics studies, after rewriting the same fetch–parse–extract–cache machinery one too many times. It's a personal research and learning project, released free and open-source under the MIT License — not a commercial product.

Features

• Composable building blocks — extractors, parsers, LLM clients, embedding memory, processors, exporters, and state managers behind abstract base classes, with Sync*Adapter wrappers for existing blocking implementations. Run the whole pipeline, or use only the parts you need, such as search.
• Async-first orchestration through AsyncETLPipeline, with bounded concurrency, resumable crash-safe state, graceful shutdown, polite retries that honor Retry-After, shared and per-host rate limits, progress events and run metrics, and explicit failure signaling through PipelineAborted. ETLPipeline runs the same pipeline from blocking code.
• LLM response caching in memory or SQLite, so a rerun doesn't pay for the same prompt twice.
• Semantic memory and local search — embed full texts into a vector store, query a SQLite FTS5 index with Boolean syntax, fuse BM25 with embedding similarity, filter by metadata facets, and grow graphs of related papers.
• Concrete implementations included — arXiv, PubMed, Semantic Scholar, and OpenAlex extractors; OpenAI-compatible chat and embedding clients; PDF, LaTeX, HTML, DOCX, and JATS XML parsers; CSV, SQL, and Plotly exporters; dataframe processors and record validators.
• Typed configuration from YAML and .env, an offline test suite at 100% coverage, and PEP 561 type information.

Installation

Python 3.11 or newer is required.

pip install "sci-etl-core[async,llm,pdf]"   # everything the example below uses
pip install "sci-etl-core[full]"            # every bundled component except local embeddings

With Poetry or uv:

poetry add "sci-etl-core[async,llm,pdf]"
uv add "sci-etl-core[async,llm,pdf]"

The base install covers configuration, both pipelines, state, the sync adapters, HTML, LaTeX, DOCX, and JATS XML parsing, Boolean search, and the pandas processors. Components load their optional dependencies only when you import them, so add an extra for each component group you use:

The installation guide lists the packages each extra adds.

Prefer configuration to code? sci-etl-cli runs these pipelines from a single YAML file.

Example

import asyncio
import os

from sci_etl_core import (
    AsyncArxivExtractor,
    AsyncCsvUpsertExporter,
    AsyncETLPipeline,
    AsyncFileStateManager,
    AsyncLLMEntityExtractor,
    AsyncLLMRelevanceFilter,
    AsyncOpenAICompatibleClient,
)
from sci_etl_core.http_async import build_async_client
from sci_etl_core.parsers import LatexTarballParser, PdfPlumberParser
from sci_etl_core.processors import DefaultKeyNormalizer

RELEVANCE_PROMPT = 'Does the paper report measurements of galaxies? Reply with JSON: {"relevant": true} or {"relevant": false}.'
EXTRACTION_PROMPT = 'Extract every measured object. Reply with JSON: {"items": [{"name": "...", "value_a": 0.0}]}.'


async def main() -> None:
    client = build_async_client()
    llm = AsyncOpenAICompatibleClient(api_key=os.environ["LLM_API_KEY"], base_url="https://api.openai.com/v1", model="gpt-4o-mini")
    pipeline = AsyncETLPipeline(
        extractor=AsyncArxivExtractor(client=client, pdf_parser=PdfPlumberParser(), latex_parser=LatexTarballParser()),
        relevance_filter=AsyncLLMRelevanceFilter(llm_client=llm, system_prompt=RELEVANCE_PROMPT),
        entity_extractor=AsyncLLMEntityExtractor(llm_client=llm, system_prompt=EXTRACTION_PROMPT),
        exporter=AsyncCsvUpsertExporter(key_column="name", value_columns=["value_a"], normalizer=DefaultKeyNormalizer()),
        state_manager=AsyncFileStateManager("state/processed.txt", "state/metadata.json"),
        destination="results.csv",
        closeables=[client, llm],
    )
    async with pipeline:
        processed = await pipeline.run(query="all:galaxy", total_limit=50)
    print(f"Processed {processed} relevant records")


asyncio.run(main())

The quick start explains what a run does, how it resumes, and what the prompts must ask for.

Configuration

Settings load from a YAML file into Pydantic models, and the LLM API key comes from the LLM_API_KEY environment variable, which a .env file can supply. Validation errors raise ConfigurationError, naming each failing key without echoing its value.

llm:
  base_url: https://api.openai.com/v1
  model: gpt-4o-mini
http:
  user_agent: "my-project/1.0 (mailto:you@example.org)"
pipeline:
  search_query: "all:galaxy"
  total_limit: 50
  max_concurrency: 4
  newest_first: true

from pathlib import Path

from sci_etl_core import AsyncOpenAICompatibleClient, BaseAppConfig, load_config


class ProjectConfig(BaseAppConfig):
    output_csv: str = "results.csv"


config = load_config(ProjectConfig, Path("config.yaml"))
llm = AsyncOpenAICompatibleClient.from_config(config.llm)
run_arguments = config.pipeline.run_arguments()

Subclass BaseAppConfig to add typed sections of your own. from_config builders take the matching section, and run_arguments() returns the keyword arguments for run(). The configuration guide lists every key and its default.

Architecture

flowchart LR
    Source[(Literature source)] --> Extractor[AsyncExtractor]
    Extractor -->|listing page| Relevance[AsyncRelevanceFilter]
    Relevance -->|irrelevant| State[(AsyncStateManager)]
    Relevance -->|relevant| FullText[fetch_full_text]
    FullText --> Memory[MemoryIngestor]
    FullText --> Entities[AsyncEntityExtractor]
    Memory --> Vectors[(Vector memory)]
    Memory --> Index[(Text index)]
    Entities --> Exporter[AsyncExporter]
    Exporter --> State
    Relevance -.-> LLM[AsyncLLMClient]
    Entities -.-> LLM
    Vectors --> Search[AsyncHybridSearcher]
    Index --> Search

AsyncETLPipeline receives every collaborator through its constructor and depends only on the abstract interfaces, so any stage can be replaced by another implementation or a test double. It pages through the listing by cursor, processes up to max_concurrency records at a time, and marks a record processed only after its entities are exported, so a failed record is retried on the next run, until it has failed in max_attempts runs. The architecture guide describes each layer.

Documentation

Testing

pip install -e ".[full,dev,lint]"
pytest --cov=sci_etl_core --cov-report=term-missing
ruff check .
mypy

The suite runs offline, and pytest --cov fails if line coverage drops below 100%.

Contributing

Contributions are welcome — new extractors, parsers, exporters, and embedding backends especially. See CONTRIBUTING.md to get set up, and browse good first issues if you're new. Moving an existing pipeline onto the library? See MIGRATION.md. What's planned is in ROADMAP.md, and releases are recorded in CHANGELOG.md. All participation is governed by our Code of Conduct.

Security

Please report vulnerabilities privately — see SECURITY.md.

License

This is a non-commercial research and educational project, freely available under the MIT License. See LICENSE for details.