Literature · FutureHouse
PaperQA2 (Crow)
Grounded retrieval-augmented generation agent that answers scientific queries using cited literature.
What PaperQA2 (Crow) is
PaperQA2 is an agentic retrieval-augmented generation (RAG) framework developed by FutureHouse, specifically designed for querying and synthesising scientific literature. In typical configurations or literature layers, it queries external biomedical sources like Europe PMC or searches across local document collections to retrieve candidate papers without fabricating references.
The system addresses the challenge of hallucinations and unsourced claims in large language model outputs. By anchoring every claim in retrieved passages and verifying paper metadata, it provides working scientists with synthesised, evidence-backed answers accompanied by full provenance and in-text citations.
PaperQA2 operates either via an automated language agent that dynamically refines searches and gathers evidence, or as a manual querying pipeline. When an LLM provider is connected, its Crow agent synthesises structured narratives; without an LLM provider, it still supplies ranked, citable evidence contexts.
How it works
PaperQA2 follows a structured three-phase algorithm: Paper Search, Gather Evidence, and Generate Answer. In the initial phase, candidate papers are identified using LLM-generated keyword queries, after which candidate texts are chunked, embedded, and added to the document index. Document metadata is gathered and verified using services such as Crossref, Semantic Scholar, or OpenAlex.
During evidence gathering, the query is embedded into a dense, sparse, or hybrid vector space to retrieve the top candidate chunks. The system then performs re-ranking and contextual summarisation (RCS), using an LLM to score each chunk's relevance to the query. The highest-scoring contextual summaries are assembled into a prompt, allowing the language model to generate a final grounded answer with corresponding citations.
Key facts
- Version
- 5.0 (PaperQA2 / CalVer)
- Domain
- Scientific literature
- Supported modalities
- Text, tables, figures, mathematics
- Hardware requirement
- Runs on CPU
- Default LLM
- OpenAI gpt-4o-2024-11-20 via LiteLLM
- Default embedding model
- OpenAI text-embedding-3-small
What you can use it for
- Grounded biomedical question answering
- Formulate scientific questions to receive synthesised textual responses where every factual assertion is linked to specific retrieved papers and page numbers.
- Literature evidence retrieval
- Search local documents or Europe PMC to extract and rank relevant text passages with associated abstracts, citations, and retraction checks.
- Scientific contradiction detection
- Run the contracrow configuration to evaluate a specific scientific claim across a corpus and flag whether literature contradicts it.
- Article and review synthesis
- Use the wikicrow configuration to draft comprehensive, evidence-grounded scientific overviews similar to structured reference entries.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Run this model
PaperQA2 (Crow)
Strengths and limits
Good at
- Guarantees that every generated claim is tied to specific passages, preventing reference fabrication.
- Integrates automatic metadata fetching, journal quality assessment, and retraction checks via providers such as Crossref and Semantic Scholar.
- Supports multiple input formats beyond PDFs, including Microsoft Office files, text files, HTML, and source code.
- Permits hybrid retrieval strategies combining dense vector embeddings with sparse keyword representations.
- Provides flexibility across LLM providers via LiteLLM as well as local execution using llamafile, Ollama, or sentence-transformers.
Watch out for
- Requires substantial instruction-following capability from the underlying LLM, making smaller models (such as 7B parameter models) perform poorly.
- Cannot automatically infer metadata or citations from source code files, requiring users to supply them manually.
- Pickled Docs state objects from PaperQA version 4 and earlier are incompatible and must be rebuilt for version 5.
Inputs and outputs
You provide
Research question
Papers to retrieveoptional
You get back
Ranked papers with citation counts and abstracts.
Grounded narrative when an LLM provider is configured.
Licence
Commercial use permitted · Apache-2.0
The Apache-2.0 licence allows use in commercial research and products.
Questions
Can I use PaperQA2 (Crow) commercially?
Commercial use permitted. The Apache-2.0 licence allows use in commercial research and products.
How do I run PaperQA2 (Crow) without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers, typically in 2.5s, and a run costs free. Visitors without an account get 1,000 credits a month.
What is the difference between PaperQA and PaperQA2?
PaperQA2 refers to version 5 and later of the repository, transitioning from earlier major versions to incorporate agentic workflows, multi-tool loops, and multimodal contextual summarisation.
Can PaperQA2 run with locally hosted models?
Yes, it supports local inference through tools like llamafile or Ollama for LLMs, and sentence-transformers for local dense and hybrid embeddings.
What formats can be indexed?
It can index PDFs, plain text, Markdown, HTML, Microsoft Office documents (.docx, .xlsx, .pptx), and various source code files.
How does the system prevent fabricated citations?
It generates answers strictly using extracted, scored summaries from verified candidate documents, retaining full provenance and metadata throughout the workflow.
How to cite it
Skarlinski et al., Language agents achieve superhuman synthesis of scientific knowledge, arXiv:2409.13740, 2024.
Written from PaperQA2 (Crow)'s own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.