PaperQA2

PaperQA2 is an agentic RAG (Retrieval-Augmented Generation) model designed for scientific papers. Key features include:

• Grounded Responses: Provides answers with in-text citations.
• Advanced RAG: Uses document metadata-awareness in embeddings, LLM-based re-ranking, and Contextual Summarization (RCS).
• Agentic Workflows: Supports iterative query refinement and answer generation via language agents.
• Metadata Enrichment: Automatically fetches citation and journal quality data from providers like Semantic Scholar and Crossref.
• Search Engine: Includes a full-text search engine for local PDF/text repositories.
• Extensibility: Supports any LiteLLM-compatible model or embedding provider.

You can quickly perform RAG (Retrieval Augmented Generation) on a folder of research papers using the pqa command-line interface. This process automatically handles metadata extraction (including citation counts and retraction checks), parses PDFs, caches them into a full-text search index, and uses an LLM agent to answer questions.

Follow these steps to set up a local directory of papers and ask a question:

pip install paper-qa
mkdir my_papers
curl -o my_papers/PaperQA2.pdf https://arxiv.org/pdf/2409.13740
cd my_papers
pqa ask 'What is PaperQA2?'

Install the paper-qa-nemotron package or include it as an extra in the main paper-qa installation. If you intend to use nemotron-parse hosted on AWS SageMaker, install the sagemaker extra.

# Install with paper-qa extras
pip install paper-qa[nemotron]

# Or install the package directly
pip install paper-qa-nemotron

# For AWS SageMaker support
pip install paper-qa-nemotron[sagemaker]

pip install paper-qa[nemotron]
# Or
pip install paper-qa-nemotron

# If you want to prompt nemotron-parse hosted on AWS SageMaker:
pip install paper-qa-nemotron[sagemaker]

Install the base paper-qa-pypdf package to use PyPDF for PDF reading within PaperQA.

pip install paper-qa-pypdf

To reproduce results from the research papers, use the question IDs and paper DOIs specified in docs/2024-10-16_litqa2-splits.json5.

• Train and eval splits: Question IDs are sourced from LAB-Bench's LitQA2.
• Test split: Question IDs are sourced from aviary-paper-data.

Specific papers for reproduction:

• skarlinski2024language: Supports train and eval splits.
• narayanan2024aviarytraininglanguageagents: Supports train, eval, and test splits.

PaperQA2 supports querying clinical trials via the clinical_trials_search tool, which retrieves data from the clinicaltrials.gov API. As of January 2025, this tool is not enabled by default.

You can query clinical trials exclusively (without using local documents) by using the search_only_clinical_trials setting name.

from paperqa import Settings, agent_query

answer_response = await agent_query(
    query="What drugs have been found to effectively treat Ulcerative Colitis?",
    settings=Settings.from_name("search_only_clinical_trials"),
)

print(answer_response.session.answer)

To combine clinical trial searches with your own documents, add clinical_trials_search to the tool_names list within the agent configuration of your Settings object. You can use DEFAULT_TOOL_NAMES from paperqa.agents.tools as a base.

from pathlib import Path
from paperqa import Settings, agent_query
from paperqa.agents.tools import DEFAULT_TOOL_NAMES

# Query using standard tools + clinical_trials and a local paper directory
answer_response = await agent_query(
    query="What drugs have been found to effectively treat Ulcerative Colitis?",
    settings=Settings(
        paper_directory="my_papers",
        agent={"tool_names": DEFAULT_TOOL_NAMES + ["clinical_trials_search"]},
    ),
)

print(answer_response.session.formatted_answer)

The fastest way to test PaperQA2 is via the pqa CLI. Navigate to a directory containing PDF papers and use the ask command to query them. PaperQA2 will index the local files (using Crossref and Semantic Scholar for metadata), chunk the evidence, and generate an answer. Subsequent queries in the same directory will skip indexing if no changes are detected.

To run with a specific parameter like temperature:

pqa --temperature 0.5 ask 'What is PaperQA2?'

pqa ask 'What is PaperQA2?'