paper-qa

PaperQA2

PaperQA2 is a package for doing high-accuracy retrieval augmented generation (RAG) on PDFs, text files, Microsoft Office documents, and source code files, with a focus on the scientific literature. See our recent 2024 paper to see examples of PaperQA2's superhuman performance in scientific tasks like question answering, summarization, and contradiction detection.

---

Table of Contents

• Quickstart
  • Example Output
• What is PaperQA2
  • PaperQA2 vs PaperQA
  • PaperQA2 Goes CalVer in December 2025
  • What's New in Version 5 (aka PaperQA2)?
  • What's New in December 2025?
  • PaperQA2 Algorithm
• Installation
• CLI Usage
  • Bundled Settings
  • Rate Limits
• Library Usage
  • Agentic Adding/Querying Documents
  • Manual (No Agent) Adding/Querying Documents
  • Async
  • Choosing Model
    • Locally Hosted
  • Embedding Model
    • Specifying the Embedding Model
    • Local Embedding Models (Sentence Transformers)
  • Adjusting number of sources
  • Using Code or HTML
  • Multimodal Support
  • Using External DB/Vector DB and Caching
  • Creating Index
    • Manifest Files
  • Reusing Index
  • Using Clients Directly
• Settings Cheatsheet
• Where do I get papers?
• Callbacks
  • Caching Embeddings
• Customizing Prompts
  • Pre and Post Prompts
• FAQ
  • How come I get different results than your papers?
  • How is this different from LlamaIndex or LangChain?
  • Can I save or load?
• Reproduction
• Citation

---

Quickstart

In this example we take a folder of research paper PDFs, magically get their metadata - including citation counts with a retraction check, then parse and cache PDFs into a full-text search index, and finally answer the user question with an LLM agent.

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?'

Example Output

Question: Has anyone designed neural networks that compute with proteins or DNA?

The claim that neural networks have been designed to compute with DNA is supported by multiple sources. The work by Qian, Winfree, and Bruck demonstrates the use of DNA strand displacement cascades to construct neural network components, such as artificial neurons and associative memories, using a DNA-based system (Qian2011Neural pages 1-2, Qian2011Neural pages 15-16, Qian2011Neural pages 54-56). This research includes the implementation of a 3-bit XOR gate and a four-neuron Hopfield associative memory, showcasing the potential of DNA for neural network computation. Additionally, the application of deep learning techniques to genomics, which involves computing with DNA sequences, is well-documented. Studies have applied convolutional neural networks (CNNs) to predict genomic features such as transcription factor binding and DNA accessibility (Eraslan2019Deep pages 4-5, Eraslan2019Deep pages 5-6). These models leverage DNA sequences as input data, effectively using neural networks to compute with DNA. While the provided excerpts do not explicitly mention protein-based neural network computation, they do highlight the use of neural networks in tasks related to protein sequences, such as predicting DNA-protein binding (Zeng2016Convolutional pages 1-2). However, the primary focus remains on DNA-based computation.

What is PaperQA2

PaperQA2 is engineered to be the best agentic RAG model for working with scientific papers. Here are some features:

• A simple interface to get good answers with grounded responses containing in-text citations.
• State-of-the-art implementation including document metadata-awareness in embeddings and LLM-based re-ranking and contextual summarization (RCS).
• Support for agentic RAG, where a language agent can iteratively refine queries and answers.
• Automatic redundant fetching of paper metadata, including citation and journal quality data from multiple providers.
• A usable full-text search engine for a local repository of PDF/text files.
• A robust interface for customization, with default support for all LiteLLM models.

By default, it uses OpenAI embeddings and models with a Numpy vector DB to embed and search documents. However, you can easily use other closed-source, open-source models or embeddings (see details below).

PaperQA2 depends on some awesome libraries/APIs that make our repo possible. Here are some in no particular order:

1. Semantic Scholar
2. Crossref
3. Unpaywall
4. Pydantic
5. tantivy
6. LiteLLM
7. pybtex

PaperQA2 vs PaperQA

We've been working hard on fundamental upgrades for a while and mostly followed SemVer, until December 2025. Meaning we've incremented the major version number on each breaking change. This brings us to the current major version number v5. So why call is the repo now called PaperQA2? We wanted to remark on the fact though that we've exceeded human performance on many important metrics. So we arbitrarily call version 5 and onward PaperQA2, and versions before it as PaperQA1 to denote the significant change in performance. We recognize that we are challenged at naming and counting at FutureHouse, so we reserve the right at any time to arbitrarily change the name to PaperCrow.

PaperQA2 Goes CalVer in December 2025

Prior to December 2025 we used semantic versioning. This eventually led to confusion in two ways:

1. Developers: should we major version bump based on settings or fundamental system capabilities? What if a bug fix requires breaking changes to the agent's behaviors?
2. Speaking: should one use terminology from our publications (e.g. PaperQA1, PaperQA2) or the Git tags (e.g. v5) from this repo/package? When someone says "PaperQA" -- what version do they mean?

To resolve these confusions, in December 2025, we moved to calendar versioning. The developer burden is diminished because we're basically removing guarantees of backwards compatibility across releases (as CalVer is ZeroVer bound to dates). It solves the "speaking" issue because Git tags are now quite different from publication terminology (e.g. PaperQA2 vs v2025.12.17). When someone says "PaperQA" it will just refer to the system, not a particular snapshot of agentic behaviors. When someone says "PaperQA2" it will refer to paper-qa>=5, which applies to both SemVer tags v5.0.0 and the new CalVer tags v2025.12.17.

This switch is backwards compatible for version 5's SemVer, as the year 2025 is strictly greater than major version 5.

What's New in Version 5 (aka PaperQA2)?

Version 5 added:

• A CLI pqa
• Agentic workflows invoking tools for paper search, gathering evidence, and generating an answer
• Removed much of the statefulness from the Docs object
• A migration to LiteLLM for compatibility with many LLM providers as well as centralized rate limits and cost tracking
• A bundled set of configurations (read this section here)) containing known-good hyperparameters

Note that Docs objects pickled from prior versions of PaperQA are incompatible with version 5, and will need to be rebuilt. Also, our minimum Python version was increased to Python 3.11.

What's New in December 2025?

The last four months since version 5.29.1 have seen many changes:

• New modalities: tables, figures, non-English languages, math equations
• More and better readers
  • Two new model-based PDF readers: Docling and Nvidia nemotron-parse
  • All PDF readers now can parse images and tables, report page numbers, support DPI
  • A reader for Microsoft Office data types
• Multimodal contextual summarization
  • Media objects are also passed to the summary_llm during creation
  • Media objects' embedding space is enhanced using an enrichment_llm prompt
• Simpler and performant HTTP stack
  • Consolidation from aiohttp and httpx to just httpx
  • Integration with httpx-aiohttp for performance
• Context relevance is simplified and some assumptions were removed
• Many minor features such as retrying Context creation upon invalid JSON, compatibility with fall 2025's frontier LLMs, and improved prompt templates
• Multiple fixes in metadata processing via Semantic Scholar and OpenAlex, and metadata processing (e.g. incorrectly inferring identical document IDs for main text and SI)
• Completed the deprecations accrued over the past year

PaperQA2 Algorithm

To understand PaperQA2, let's start with the pieces of the underlying algorithm. The default workflow of PaperQA2 is as follows:

Phase	PaperQA2 Actions
1. Paper Search	- Get candidate papers from LLM-generated keyword query
	- Chunk, embed, and add candidate papers to state
2. Gather Evidence	- Embed query into vector
	- Rank top k document chunks in current state
	- Create scored summary of each chunk in the context of the current query
	- Use LLM to re-score and select most relevant summaries
3. Generate Answer	- Put best summaries into prompt with context
	- Generate answer with prompt

The tools can be invoked in any order by a language agent. For example, an LLM agent might do a narrow and broad search, or using different phrasing for the gather evidence step from the generate answer step.

Installation

For a non-development setup, install PaperQA2 (aka version 5) from PyPI. Note version 5 requires Python 3.11+.

pip install paper-qa>=5

For development setup, please refer to the CONTRIBUTING.md file.

PaperQA2 uses an LLM to operate, so you'll need to either set an appropriate API key environment variable (i.e. export OPENAI_API_KEY=sk-...) or set up an open source LLM server (i.e. using llamafile. Any LiteLLM compatible model can be configured to use with PaperQA2.

If you need to index a large set of papers (100+), you will likely want an API key for both Crossref and Semantic Scholar, which will allow you to avoid hitting public rate limits using these metadata services. Those can be exported as CROSSREF_API_KEY and SEMANTIC_SCHOLAR_API_KEY variables.

CLI Usage

The fastest way to test PaperQA2 is via the CLI. First navigate to a directory with some papers and use the pqa cli:

pqa ask 'What is PaperQA2?'

You will see PaperQA2 index your local PDF files, gathering the necessary metadata for each of them (using Crossref and Semantic Scholar), search over that index, then break the files into chunked evidence contexts, rank them, and ultimately generate an answer. The next time this directory is queried, your index will already be built (save for any differences detected, like new added papers), so it will skip the indexing and chunking steps.

All prior answers will be indexed and stored, you can view them by querying via the search subcommand, or access them yourself in your PQA_HOME directory, which defaults to ~/.pqa/.

pqa -i 'answers' search 'ranking and contextual summarization'

PaperQA2 is highly configurable, when running from the command line, pqa --help shows all options and short descriptions. For example, GPT-4o accepts a custom temperature:

pqa --llm gpt-4o --summary_llm gpt-4o --temperature 0.5 ask 'What is PaperQA2?'

You can view all settings with pqa view. Another useful thing is to change to other templated settings - for example fast is a setting that answers more quickly and you can see it with pqa -s fast view

Maybe you have some new settings you want to save? You can do that with

pqa -s my_new_settings --temperature 0.5 --llm foo-bar-5 save

and then you can use it with

pqa -s my_new_settings ask 'What is PaperQA2?'

If you run pqa with a command which requires a new indexing, say if you change the default chunk_size, a new index will automatically be created for you.

pqa --parsing.chunk_size 5000 ask 'What is PaperQA2?'

You can also use pqa to do full-text search with use of LLMs view the search command. For example, let's save the index from a directory and give it a name:

pqa -i nanomaterials index

Now I can search for papers about thermoelectrics:

pqa -i nanomaterials search thermoelectrics

or I can use the normal ask

pqa -i nanomaterials ask 'Are there nm scale features in thermoelectric materials?'

Both the CLI and module have pre-configured settings based on prior performance and our publications, they can be invoked as follows:

pqa --settings <setting name> \
    ask 'Are there nm scale features in thermoelectric materials?'

Bundled Settings

Inside src/paperqa/configs we bundle known useful settings:

Setting Name	Description
high_quality	Highly performant, relatively expensive (due to having evidence_k = 15) query using a ToolSelector agent.
fast	Setting to get answers cheaply and quickly.
wikicrow	Setting to emulate the Wikipedia article writing used in our WikiCrow publication.
contracrow	Setting to find contradictions in papers, your query should be a claim that needs to be flagged as a contradiction (or not).
debug	Setting useful solely for debugging, but not in any actual application beyond debugging.
tier1_limits	Settings that match OpenAI rate limits for each tier, you can use tier<1-5>_limits to specify the tier.

Rate Limits

If you are hitting rate limits, say with the OpenAI Tier 1 plan, you can add them into PaperQA2. For each OpenAI tier, a pre-built setting exists to limit usage.

pqa --settings 'tier1_limits' ask 'What is PaperQA2?'

This will limit your system to use the tier1_limits, and slow down your queries to accommodate.

You can also specify them manually with any rate limit string that matches the specification in the limits module:

pqa --summary_llm_config '{"rate_limit": {"openai/gpt-6-luna": "30000 per 1 minute"}}' \
    ask 'What is PaperQA2?'

Or by adding into a Settings object, if calling imperatively:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm_config={"rate_limit": {"openai/gpt-6.1-sol": "30000 per 1 minute"}},
        summary_llm_config={"rate_limit": {"openai/gpt-6-luna": "30000 per 1 minute"}},
    ),
)

Library Usage

PaperQA2's full workflow can be accessed via Python directly:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(agent={"index": {"paper_directory": "my_papers"}}),
)

Please see our installation docs for how to install the package from PyPI.

Agentic Adding/Querying Documents

The answer object has the following attributes: formatted_answer, answer (answer alone), question , and context (the summaries of passages found for answer). ask will use the SearchPapers tool, which will query a local index of files, you can specify this location via the Settings object:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(agent={"index": {"paper_directory": "my_papers"}}),
)

ask is just a convenience wrapper around the real entrypoint, which can be accessed if you'd like to run concurrent asynchronous workloads:

from paperqa import Settings, agent_query

answer_response = await agent_query(
    query="What is PaperQA2?",
    settings=Settings(agent={"index": {"paper_directory": "my_papers"}}),
)

The default agent will use an LLM based agent, but you can also specify a "fake" agent to use a hard coded call path of search -> gather evidence -> answer to reduce token usage.

Manual (No Agent) Adding/Querying Documents

Normally via agent execution, the agent invokes the search tool, which adds documents to the Docs object for you behind the scenes. However, if you prefer fine-grained control, you can directly interact with the Docs object.

Note that manually adding and querying Docs does not impact performance. It just removes the automation associated with an agent picking the documents to add.

from paperqa import Docs, Settings

# valid extensions include .pdf, .txt, .md, .html, .docx, .xlsx, .pptx, and code files (e.g., .py, .ts, .yaml)
doc_paths = ("myfile.pdf", "myotherfile.pdf")

# Prepare the Docs object by adding a bunch of documents
docs = Docs()
for doc_path in doc_paths:
    await docs.aadd(doc_path)

# Set up how we want to query the Docs object
settings = Settings()
settings.llm = "claude-3-5-sonnet-20240620"
settings.answer.answer_max_sources = 3

# Query the Docs object to get an answer
session = await docs.aquery("What is PaperQA2?", settings=settings)
print(session)

Async

PaperQA2 is written to be used asynchronously. The synchronous API is just a wrapper around the async. Here are the methods and their async equivalents:

Sync	Async
Docs.add	Docs.aadd
Docs.add_file	Docs.aadd_file
Docs.add_url	Docs.aadd_url
Docs.get_evidence	Docs.aget_evidence
Docs.query	Docs.aquery

The synchronous version just calls the async version in a loop. Most modern python environments support async natively (including Jupyter notebooks!). So you can do this in a Jupyter Notebook:

import asyncio
from paperqa import Docs


async def main() -> None:
    docs = Docs()
    # valid extensions include .pdf, .txt, .md, .html, .docx, .xlsx, .pptx, and code files (e.g., .py, .ts, .yaml)
    for doc in ("myfile.pdf", "myotherfile.pdf"):
        await docs.aadd(doc)

    session = await docs.aquery("What is PaperQA2?")
    print(session)


asyncio.run(main())

Choosing Model

By default, PaperQA2 uses OpenAI's gpt-6-luna for summary_llm, gpt-6.1-sol for llm and media enrichment, and gpt-6-sol for agent_llm. Please see the Settings Cheatsheet for more information on these settings. PaperQA2 also defaults to using OpenAI's text-embedding-3-small model for the embedding setting. If you don't have an OpenAI API key, you can use a different embedding model. More information about embedding models can be found in the "Embedding Model" section.

We use the lmi package for our LLM interface, which in turn uses litellm to support many LLM providers. You can adjust this easily to use any model supported by litellm:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="openai/gpt-6.1-sol", summary_llm="openai/gpt-6-luna", agent={"index": {"paper_directory": "my_papers"}}
    ),
)

To use Claude, make sure you set the ANTHROPIC_API_KEY environment variable. In this example, we also use a different embedding model. Please make sure to pip install paper-qa[local] to use a local embedding model.

from paperqa import Settings, ask
from paperqa.settings import AgentSettings

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="claude-3-5-sonnet-20240620",
        summary_llm="claude-3-5-sonnet-20240620",
        agent=AgentSettings(agent_llm="claude-3-5-sonnet-20240620"),
        # SEE: https://huggingface.co/sentence-transformers/multi-qa-MiniLM-L6-cos-v1
        embedding="st-multi-qa-MiniLM-L6-cos-v1",
    ),
)

Or Gemini, by setting the GEMINI_API_KEY from Google AI Studio

from paperqa import Settings, ask
from paperqa.settings import AgentSettings

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="gemini/gemini-2.0-flash",
        summary_llm="gemini/gemini-2.0-flash",
        agent=AgentSettings(agent_llm="gemini/gemini-2.0-flash"),
        embedding="gemini/text-embedding-004",
    ),
)

Locally Hosted

You can use llama.cpp to be the LLM. Note that you should be using relatively large models, because PaperQA2 requires following a lot of instructions. You won't get good performance with 7B models.

The easiest way to get set-up is to download a llama file and execute it with -cb -np 4 -a my-llm-model --embedding which will enable continuous batching and embeddings.

from paperqa import Settings, ask

local_llm_config = dict(
    model_list=[
        dict(
            model_name="my_llm_model",
            litellm_params=dict(
                model="my-llm-model",
                api_base="http://localhost:8080/v1",
                api_key="sk-no-key-required",
                temperature=0.1,
                frequency_penalty=1.5,
                max_tokens=512,
            ),
        )
    ]
)

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="my-llm-model",
        llm_config=local_llm_config,
        summary_llm="my-llm-model",
        summary_llm_config=local_llm_config,
    ),
)

Models hosted with ollama are also supported. To run the example below make sure you have downloaded llama3.2 and mxbai-embed-large via ollama.

from paperqa import Settings, ask

local_llm_config = {
    "model_list": [
        {
            "model_name": "ollama/llama3.2",
            "litellm_params": {
                "model": "ollama/llama3.2",
                "api_base": "http://localhost:11434",
            },
        }
    ]
}

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(
        llm="ollama/llama3.2",
        llm_config=local_llm_config,
        summary_llm="ollama/llama3.2",
        summary_llm_config=local_llm_config,
        embedding="ollama/mxbai-embed-large",
    ),
)

Embedding Model

Embeddings are used to retrieve k texts (where k is specified via Settings.answer.evidence_k) for re-ranking and contextual summarization. If you don't want to use embeddings, but instead just fetch all chunks, disable "evidence retrieval" via the Settings.answer.evidence_retrieval setting.

PaperQA2 defaults to using OpenAI (text-embedding-3-small) embeddings, but has flexible options for both vector stores and embedding choices.

Specifying the Embedding Model

The simplest way to specify the embedding model is via Settings.embedding:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(embedding="text-embedding-3-large"),
)

embedding accepts any embedding model name supported by litellm. PaperQA2 also supports an embedding input of "hybrid-<model_name>" i.e. "hybrid-text-embedding-3-small" to use a hybrid sparse keyword (based on a token modulo embedding) and dense vector embedding, where any litellm model can be used in the dense model name. "sparse" can be used to use a sparse keyword embedding only.

Embedding models are used to create PaperQA2's index of the full-text embedding vectors (texts_index argument). The embedding model can be specified as a setting when you are adding new papers to the Docs object:

from paperqa import Docs, Settings

docs = Docs()
for doc in ("myfile.pdf", "myotherfile.pdf"):
    await docs.aadd(doc, settings=Settings(embedding="text-embedding-large-3"))

Note that PaperQA2 uses Numpy as a dense vector store. Its design of using a keyword search initially reduces the number of chunks needed for each answer to a relatively small number < 1k. Therefore, NumpyVectorStore is a good place to start, it's a simple in-memory store, without an index. However, if a larger-than-memory vector store is needed, you can an external vector database like Qdrant via the QdrantVectorStore class.

The hybrid embeddings can be customized:

from paperqa import (
    Docs,
    HybridEmbeddingModel,
    SparseEmbeddingModel,
    LiteLLMEmbeddingModel,
)


model = HybridEmbeddingModel(
    models=[LiteLLMEmbeddingModel(), SparseEmbeddingModel(ndim=1024)]
)
docs = Docs()
for doc in ("myfile.pdf", "myotherfile.pdf"):
    await docs.aadd(doc, embedding_model=model)

The sparse embedding (keyword) models default to having 256 dimensions, but this can be specified via the ndim argument.

Local Embedding Models (Sentence Transformers)

You can use a SentenceTransformerEmbeddingModel model if you install sentence-transformers, which is a local embedding library with support for HuggingFace models and more. You can install it by adding the local extras.

pip install paper-qa[local]

and then prefix embedding model names with st-:

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(embedding="st-multi-qa-MiniLM-L6-cos-v1"),
)

or with a hybrid model

from paperqa import Settings, ask

answer_response = ask(
    "What is PaperQA2?",
    settings=Settings(embedding="hybrid-st-multi-qa-MiniLM-L6-cos-v1"),
)

Adjusting number of sources

You can adjust the numbers of sources (passages of text) to reduce token usage or add more context. k refers to the top k most relevant and diverse (may from different sources) passages. Each passage is sent to the LLM to summarize, or determine if it is irrelevant. After this step, a limit of max_sources is applied so that the final answer can fit into the LLM context window. Thus, k > max_sources and max_sources is the number of sources used in the final answer.

from paperqa import Settings

settings = Settings()
settings.answer.answer_max_sources = 3
settings.answer.evidence_k = 5

await docs.aquery(
    "What is PaperQA2?",
    settings=settings,
)

Using Code or HTML

You do not need to use papers -- you can use code or raw HTML. Note that this tool is focused on answering questions, so it won't do well at writing code. One note is that the tool cannot infer citations from code, so you will need to provide them yourself.

import glob
import os
from paperqa import Docs

source_files = glob.glob("**/*.js")

docs = Docs()
for f in source_files:
    # this assumes the file names are unique in code
    await docs.aadd(
        f, citation="File " + os.path.basename(f), docname=os.path.basename(f)
    )
session = await docs.aquery("Where is the search bar in the header defined?")
print(session)

Multimodal Support

Multimodal support centers on:

• Standalone images
• Images or tables in PDFs

The Docs object stores media via a ParsedMedia object. When chunking a document, media are not split at chunk boundaries, so it's possible 2+ chunks can correspond with the same media. This means within PaperQA each chunk has a one-to-many relationship between ParsedMedia and chunks.

Depending on the source document, the same image can appear multiple times (e.g. each page of a PDF has a logo in the margins). Thus, clients should consider media databases to have a many-to-many relationship with chunks.

Since PaperQA's evidence gathering process centers on text-based retrieval, it's possible relevant image(s) or table(s) aren't retrieved because their associated text content is irrelevant. For a concrete example, imagine the figure in a paper has a terse caption and is placed one page after relevant main-text discussion. To solve this problem, PaperQA supports media enrichment at document read-time. Basically after reading in the PDF, the parsing.enrichment_llm is given the parsing.enrichment_prompt and co-located text to generate a synthetic caption for every image/table. The synthetic captions are used to shift the embeddings of each text chunk, but are kept separate from the actual source text. This way evidence gathering can fetch relevant images/tables without risk of polluting contextual summaries with LLM-generated captions.

If you want multimodal PDF reading, but do not want enrichment (since adds one LLM prompt/media at read-time), enrichment can be disabled by setting parsing.multimodal to ON_WITHOUT_ENRICHMENT.

When creating contextual summaries on a given chunk (a Text), the summary LLM is passed both the chunk's text and the chunk's associated media, but the output contextual summary itself remains text-only.

If you would like, specifying the prompt paperqa.prompts.summary_json_multimodal_system_prompt to the setting prompt.summary_json_system will include a used_images flag attributing usage of images in any contextual summarizations.

Using External DB/Vector DB and Caching

You may want to cache parsed texts and embeddings in an external database or file. You can then build a Docs object from those directly:

from paperqa import Docs, Doc, Text

docs = Docs()

for ... in my_docs:
    doc = Doc(docname=..., citation=..., dockey=..., citation=...)
    texts = [Text(text=..., name=..., doc=doc) for ... in my_texts]
    docs.add_texts(texts, doc)

Creating Index

Indexes will be placed in the home directory by default. This can be controlled via the PQA_HOME environment variable.

Indexes are made by reading files in the IndexSettings.paper_directory. By default, we recursively read from subdirectories of the paper directory, unless disabled using IndexSettings.recurse_subdirectories. The paper directory is not modified in any way, it's just read from.

Manifest Files

The indexing process attempts to infer paper metadata like title and DOI using LLM-powered text processing. You can avoid this point of uncertainty using a "manifest" file, which is a CSV containing DocDetails fields (order doesn't matter). For example:

• file_location: relative path to the paper's PDF within the index directory
• doi: DOI of the paper
• title: title of the paper

By providing this information, we ensure queries to metadata providers like Crossref are accurate.

To ease creating a manifest, there is a helper class method Doc.to_csv, which also works when called on DocDetails.

Reusing Index

The local search indexes are built based on a hash of the current Settings object. So make sure you properly specify the paper_directory to your IndexSettings object. In general, it's advisable to:

1. Pre-build an index given a folder of papers (can take several minutes)
2. Reuse the index to perform many queries

import os

from paperqa import Settings
from paperqa.agents.main import agent_query
from paperqa.agents.search import get_directory_index


async def amain(folder_of_papers: str | os.PathLike) -> None:
    settings = Settings(agent={"index": {"paper_directory": folder_of_papers}})

    # 1. Build the index. Note an index name is autogenerated when unspecified
    built_index = await get_directory_index(settings=settings)
    print(settings.get_index_name())  # Display the autogenerated index name
    print(await built_index.index_files)  # Display the index contents

    # 2. Use the settings as many times as you want with ask
    answer_response_1 = await agent_query(
        query="What is a cool retrieval augmented generation technique?",
        settings=settings,
    )
    answer_response_2 = await agent_query(
        query="What is PaperQA2?",
        settings=settings,
    )

Using Clients Directly

One of the most powerful features of PaperQA2 is its ability to combine data from multiple metadata sources. For example, Unpaywall can provide open access status/direct links to PDFs, Crossref can provide bibtex, and Semantic Scholar can provide citation licenses. Here's a short demo of how to do this:

from paperqa.clients import DocMetadataClient, ALL_CLIENTS

client = DocMetadataClient(metadata_clients=ALL_CLIENTS)
details = await client.query(title="Augmenting language models with chemistry tools")

print(details.formatted_citation)
# Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari,
# Andrew D. White, and Philippe Schwaller.
#  Augmenting large language models with chemistry tools. Nature Machine Intelligence,
# 6:525-535, May 2024. URL: https://doi.org/10.1038/s42256-024-00832-8,
# doi:10.1038/s42256-024-00832-8.
# This article has 243 citations and is from a domain leading peer-reviewed journal.

print(details.citation_count)
# 243

print(details.license)
# cc-by

print(details.pdf_url)
# https://www.nature.com/articles/s42256-024-00832-8.pdf

the client.query is meant to check for exact matches of title. It's a bit robust (like to casing, missing a word). There are duplicates for titles though - so you can also add authors to disambiguate. Or you can provide a doi directly client.query(doi="10.1038/s42256-024-00832-8").

If you're doing this at a large scale, you may not want to use ALL_CLIENTS (just omit the argument) and you can specify which specific fields you want to speed up queries. For example:

details = await client.query(
    title="Augmenting large language models with chemistry tools",
    authors=["Andres M. Bran", "Sam Cox"],
    fields=["title", "doi"],
)

will return much faster than the first query and we'll be certain the authors match.

Settings Cheatsheet

Setting	Default	Description
llm	"openai/gpt-6.1-sol"	LLM for general use including metadata inference (see Docs.aadd) and answer generation (see Docs.aquery and gen_answer tool).
llm_config	None	Optional configuration for llm.
summary_llm	"openai/gpt-6-luna"	LLM for creating contextual summaries (see Docs.aget_evidence and gather_evidence tool).
summary_llm_config	None	Optional configuration for summary_llm.
embedding	"text-embedding-3-small"	Embedding model for embedding text chunks when adding papers.
embedding_config	None	Optional configuration for embedding.
temperature	None	Optional temperature for LLMs; omitted when unset.
batch_size	1	Batch size for calling LLMs.
texts_index_mmr_lambda	1.0	Lambda for MMR in text index.
verbosity	0	Integer verbosity level for logging (0-3). 3 = all LLM/Embeddings calls logged.
custom_context_serializer	None	Custom async function (see typing for signature) to override the default answer context serialization.
answer.evidence_k	10	Number of evidence pieces to retrieve.
answer.evidence_retrieval	True	Use retrieval vs processing all docs.
answer.evidence_summary_length	"about 100 words"	Length of evidence summary.
answer.evidence_skip_summary	False	Whether to skip summarization.
answer.evidence_text_only_fallback	False	Whether to allow context creation to retry without media present.
answer.answer_max_sources	5	Max number of sources for an answer.
answer.max_answer_attempts	None	Max attempts to generate an answer.
answer.answer_length	"about 200 words, but can be longer"	Length of final answer.
answer.max_concurrent_requests	4	Max concurrent requests to LLMs.
answer.answer_filter_extra_background	False	Whether to cite background info from model.
answer.get_evidence_if_no_contexts	True	Allow lazy evidence gathering.
answer.group_contexts_by_question	False	Groups the final contexts by the underlying gather_evidence question in the final context prompt.
answer.evidence_relevance_score_cutoff	1	Cutoff evidence relevance score to include in the answer context (inclusive)
answer.skip_evidence_citation_strip	False	Skip removal of citations from the gather_evidence contexts
parsing.page_size_limit	1,280,000	Character limit per page.
parsing.use_doc_details	True	Whether to get metadata details for docs.
parsing.reader_config	dict	Optional keyword arguments for the document reader.
parsing.multimodal	True	Control to parse both text and media from applicable documents, as well as potentially enriching them with text descriptions.
parsing.defer_embedding	False	Whether to defer embedding until summarization.
parsing.parse_pdf	paperqa_pypdf.parse_pdf_to_pages	Function to parse PDF files.
parsing.configure_pdf_parser	No-op	Callable to configure the PDF parser within parse_pdf, useful for behaviors such as enabling logging.
parsing.doc_filters	None	Optional filters for allowed documents.
parsing.use_human_readable_clinical_trials	False	Parse clinical trial JSONs into readable text.
parsing.enrichment_llm	"openai/gpt-6.1-sol"	LLM for media enrichment.
parsing.enrichment_llm_config	None	Optional configuration for enrichment_llm.
parsing.enrichment_page_radius	1	Page radius for context text in enrichment.
parsing.enrichment_prompt	image_enrichment_prompt_template	Prompt template for enriching media.
parsing.citation_prompt	citation_prompt	Prompt to create citation from peeking one chunk.
parsing.structured_citation_prompt	structured_citation_prompt	Prompt to create a citation (in JSON) from peeking one chunk.
parsing.disable_doc_valid_check	False	Flag to disable checking if a document looks like text (was parsed correctly).
prompts.summary	summary_prompt	Template for summarizing text, must contain variables matching summary_prompt.
prompts.qa	qa_prompt	Template for QA, must contain variables matching qa_prompt.
prompts.select	select_paper_prompt	Template for selecting papers, must contain variables matching select_paper_prompt.
prompts.pre	None	Optional pre-prompt templated with just the original question to append information before a qa prompt.
prompts.post	None	Optional post-processing prompt that can access PQASession fields.
prompts.system	default_system_prompt	System prompt for the model.
prompts.use_json	True	Whether to use JSON formatting.
prompts.summary_json	summary_json_prompt	JSON-specific summary prompt.
prompts.summary_json_system	summary_json_system_prompt	System prompt for JSON summaries.
prompts.context_outer	CONTEXT_OUTER_PROMPT	Prompt for how to format all contexts in generate answer.
prompts.context_inner	CONTEXT_INNER_PROMPT	Prompt for how to format a single context in generate answer. Must contain 'name' and 'text' variables.
prompts.answer_iteration_prompt	answer_iteration_prompt_template	Prompt to inject existing prior answers to allow iteration. Default injects no prior answers.
agent.agent_llm	"openai/gpt-6-sol"	LLM inside the agent making tool selections.
agent.agent_llm_config	None	Optional configuration for agent_llm.
agent.agent_type	"ToolSelector"	Type of agent to use.
agent.agent_config	None	Optional kwarg for AGENT constructor.
agent.agent_system_prompt	env_system_prompt	Optional system prompt message.
agent.agent_prompt	env_reset_prompt	Agent prompt.
agent.return_paper_metadata	False	Whether to include paper title/year in search tool results.
agent.search_count	8	Search count.
agent.timeout	500.0	Timeout on agent execution (seconds).
agent.tool_names	None	Optional override on tools to provide the agent.
agent.max_timesteps	None	Optional upper limit on environment steps.
agent.agent_evidence_n	1	Top n ranked evidences shown to the agent after gathering evidence.
agent.rebuild_index	True	Flag to rebuild the index at the start of agent runners.
agent.callbacks	{}	Named lists of callables to be invoked with environment state.
agent.index.name	None	Optional name of the index.
agent.index.paper_directory	Current working directory	Directory containing papers to be indexed.
agent.index.manifest_file	None	Path to manifest CSV with document attributes.
agent.index.index_directory	pqa_directory("indexes")	Directory to store PQA indexes.
agent.index.use_absolute_paper_directory	False	Whether to use absolute paper directory path.
agent.index.recurse_subdirectories	True	Whether to recurse into subdirectories when indexing.
agent.index.concurrency	5	Number of concurrent filesystem reads.
agent.index.sync_with_paper_directory	True	Whether to sync index with paper directory on load.
agent.index.batch_size	1	Number of files to process before committing to the index.
agent.index.files_filter	lambda f: f.suffix in {...}	Filter function to mark files in the paper directory to index.

Where do I get papers?

Well that's a really good question! It's probably best to just download PDFs of papers you think will help answer your question and start from there.

See detailed docs about zotero, openreview and parsing

Callbacks

To execute a function on each chunk of LLM completions, you need to provide a function that can be executed on each chunk. For example, to get a typewriter view of the completions, you can do:

from paperqa import Docs


def typewriter(chunk: str) -> None:
    print(chunk, end="")


docs = Docs()

# add some docs...

await docs.aquery("What is PaperQA2?", callbacks=[typewriter])

Caching Embeddings

In general, embeddings are cached when you pickle a Docs regardless of what vector store you use. So as long as you save your underlying Docs object, you should be able to avoid re-embedding your documents.

Customizing Prompts

You can customize any of the prompts using settings.

from paperqa import Docs, Settings

my_qa_prompt = (
    "Answer the question '{question}'\n"
    "Use the context below if helpful. "
    "You can cite the context using the key like (pqac-abcd1234). "
    "If there is insufficient context, write a poem "
    "about how you cannot answer.\n\n"
    "Context: {context}"
)

docs = Docs()
settings = Settings()
settings.prompts.qa = my_qa_prompt
await docs.aquery("What is PaperQA2?", settings=settings)

Pre and Post Prompts

Following the syntax above, you can also include prompts that are executed after the query and before the query. For example, you can use this to critique the answer.

FAQ

How come I get different results than your papers?

Internally at FutureHouse, we have a slightly different set of tools. We're trying to get some of them, like citation traversal, into this repo. However, we have APIs and licenses to access research papers that we cannot share openly. Similarly, in our research papers' results we do not start with the known relevant PDFs. Our agent has to identify them using keyword search over all papers, rather than just a subset. We're gradually aligning these two versions of PaperQA, but until there is an open-source way to freely access papers (even just open source papers) you will need to provide PDFs yourself.

How is this different from LlamaIndex or LangChain?

LangChain and LlamaIndex are both frameworks for working with LLM applications, with abstractions made for agentic workflows and retrieval augmented generation.

Over time, the PaperQA team over time chose to become framework-agnostic, instead outsourcing LLM drivers to LiteLLM and no framework besides Pydantic for its tools. PaperQA focuses on scientific papers and their metadata.

PaperQA can be reimplemented using either LlamaIndex or LangChain. For example, our GatherEvidence tool can be reimplemented as a retriever with an LLM-based re-ranking and contextual summary. There is similar work with the tree response method in LlamaIndex.

Can I save or load?

The Docs class can be pickled and unpickled. This is useful if you want to save the embeddings of the documents and then load them later.

import pickle

# save
with open("my_docs.pkl", "wb") as f:
    pickle.dump(docs, f)

# load
with open("my_docs.pkl", "rb") as f:
    docs = pickle.load(f)

Reproduction

Contained in docs/2024-10-16_litqa2-splits.json5 are the question IDs used in train, evaluation, and test splits, as well as paper DOIs used to build the splits' indexes.

• Train and eval splits: question IDs come from LAB-Bench's LitQA2 question IDs.
• Test split: questions IDs come from aviary-paper-data's LitQA2 question IDs.

There are multiple papers slowly building PaperQA, shown below in Citation. To reproduce:

• skarlinski2024language: train and eval splits are applicable. The test split remains held out.
• narayanan2024aviarytraininglanguageagents: train, eval, and test splits are applicable.

Example on how to use LitQA for evaluation can be found in aviary.litqa.

Citation

Please read and cite the following papers if you use this software:

@article{narayanan2024aviarytraininglanguageagents,
      title = {Aviary: training language agents on challenging scientific tasks},
      author = {
      Siddharth Narayanan and
 James D. Braza and
 Ryan-Rhys Griffiths and
 Manu Ponnapati and
 Albert Bou and
 Jon Laurent and
 Ori Kabeli and
 Geemi Wellawatte and
 Sam Cox and
 Samuel G. Rodriques and
 Andrew D. White},
      journal = {arXiv preprent arXiv:2412.21154},
      year = {2024},
      url = {https://doi.org/10.48550/arXiv.2412.21154},
}

@article{skarlinski2024language,
    title = {Language agents achieve superhuman synthesis of scientific knowledge},
    author = {
    Michael D. Skarlinski and
 Sam Cox and
 Jon M. Laurent and
 James D. Braza and
 Michaela Hinks and
 Michael J. Hammerling and
 Manvitha Ponnapati and
 Samuel G. Rodriques and
 Andrew D. White},
    journal = {arXiv prep