ArxivAgent Documentation

ArxivAgent is an acquisition agent for arXiv papers. It subclasses BaseAcquisitionAgent, so it uses the same acquire-then-summarize/RAG graph as WebSearchAgent and OSTIAgent.

ArxivAgent searches the arXiv API, downloads matching paper PDFs, extracts text, optionally augments PDF text with image descriptions, and then summarizes the acquired papers or runs the RAG path when an embedding model is configured.

See also: Acquisition Agents.

Basic usage

from langchain.chat_models import init_chat_model
from ursa.agents import ArxivAgent

llm = init_chat_model("openai:gpt-5.4-mini")
agent = ArxivAgent(llm=llm)

state = agent.invoke({
    "query": "neutron star radius constraints",
    "context": "What are the constraints on neutron star radius, and what uncertainties affect those constraints?",
})

print(agent.format_result(state))

You can also pass a plain string. In that case, the string becomes context, and the agent asks the LLM to generate a short arXiv search query:

state = agent.invoke("Summarize recent arXiv papers about neutron star radius constraints.")
print(agent.format_result(state))

Parameters

ArxivAgent uses the shared BaseAcquisitionAgent parameters with arXiv-specific defaults:

Parameter	Type	Default	Description
llm	BaseChatModel	required	Language model used for query generation and summarization.
process_images	bool	True	For downloaded PDFs, optionally append image interpretations when image-description support is available.
max_results	int	3	Maximum number of arXiv results to acquire.
download	bool	True	If True, search arXiv and download PDFs. If False, read cached files from database_path.
rag_embedding	optional embedding object	None	If provided, use the RAG path instead of direct per-paper summarization.
database_path	str	"arxiv_papers"	Directory under the agent den for downloaded/cached papers.
summaries_path	str	"arxiv_generated_summaries"	Directory under the agent den for per-paper and final summaries.
vectorstore_path	str	"arxiv_vectorstores"	Stored vector-store path configuration inherited from the acquisition base. The current shared RAG node constructs a RAGAgent over the acquired database when rag_embedding is provided.
**kwargs	dict	{}	Passed to BaseAgent / BaseAcquisitionAgent, including workspace/den and persistence options.

Inherited acquisition parameters such as summarize and num_threads can also be supplied through the shared base class path.

How it works

ArxivAgent implements the acquisition hooks required by BaseAcquisitionAgent:

• _search(query) — queries http://export.arxiv.org/api/query and normalizes entries into lightweight hits containing arXiv IDs and titles.
• _id(hit_or_item) — uses the arXiv ID as the stable item ID.
• _materialize(hit) — downloads https://arxiv.org/pdf/<arxiv_id>.pdf, stores it in database_path, and extracts PDF text.
• _citation(item) — formats citations as ArXiv ID: <id>.

After acquisition, the shared graph either:

• summarizes each paper and aggregates the summaries into state["final_summary"], or
• invokes the RAG path when rag_embedding is configured, or
• stops after state["items"] when summarize=False.

Output state

Important fields in the returned state include:

• query — final arXiv search query.
• context — task/question used for summarization.
• items — acquired paper metadata, including local PDF paths and extracted text.
• summaries — per-paper summaries when direct summarization is used.
• final_summary — aggregate response to the requested context.

agent.format_result(state) returns state["final_summary"] when present.

Cached files

By default, the agent writes artifacts under the agent den:

• PDFs: arxiv_papers/
• per-paper summaries: arxiv_generated_summaries/
• combined direct-summarization file: arxiv_generated_summaries/summaries_combined.txt
• final direct-summarization file: arxiv_generated_summaries/final_summary.txt
• RAG workflow artifacts are managed by the shared RAG path when rag_embedding is provided.

CLI

The interactive CLI registers this agent as:

arxiv