Example Usage

These examples show ArxivRetriever composed into LLM applications. They use init_chat_model so you can swap in any chat model provider.

pip install langchain-arxiv-retriever langchain langchain-openai

RAG over abstracts

The classic retrieval chain: stuff the retrieved abstracts into a prompt as context. Abstract mode keeps this fast — no PDFs are downloaded.

from langchain.chat_models import init_chat_model
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough

from langchain_arxiv import ArxivRetriever

llm = init_chat_model("openai:gpt-5.4-mini")
retriever = ArxivRetriever(k=5)

prompt = ChatPromptTemplate.from_template(
    """Answer the question based only on the context provided.
Cite papers by their arXiv ID.

Context: {context}

Question: {question}"""
)


def format_docs(docs):
    return "\n\n".join(
        f"[{doc.metadata['arxiv_id']}] {doc.metadata['title']}\n{doc.page_content}"
        for doc in docs
    )


chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

chain.invoke("ImageBind joint embedding")

Rewriting natural-language questions into arXiv queries

arXiv search is lexical, not semantic — conversational questions retrieve poorly. The retriever deliberately passes queries through verbatim, so the fix belongs in the chain: have the LLM translate the question into arXiv query syntax first.

from langchain.chat_models import init_chat_model
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate

from langchain_arxiv import ArxivRetriever

llm = init_chat_model("openai:gpt-5.4-mini")
retriever = ArxivRetriever(k=5)

rewrite_prompt = ChatPromptTemplate.from_template(
    """Convert the question into one concise arXiv search query.
Prefer a few precise keywords over full sentences. Use arXiv field
syntax where it helps: ti:"..." for title terms, abs:"..." for
abstract terms, au:"..." for authors, cat:cs.LG for categories;
combine with AND / OR / ANDNOT.
Return only the query, nothing else.

Question: {question}"""
)

rewrite = rewrite_prompt | llm | StrOutputParser()
search = rewrite | retriever

docs = search.invoke(
    {"question": "why do transformers suddenly generalize after long training?"}
)

Combined with the RAG chain above, the rewriter replaces the bare retriever:

chain = (
    {"context": search | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

Building structured arXiv queries with an LLM

The freeform rewrite above is the lightest touch. For tighter control, have the model emit a structured query and assemble the arXiv syntax in code: the fields stay machine-checkable, and the assembly is deterministic instead of hoping the model quotes and parenthesizes correctly.

from pydantic import BaseModel, Field

from langchain.chat_models import init_chat_model
from langchain_arxiv import ArxivRetriever


class ArxivQuery(BaseModel):
    """A structured arXiv search query."""

    keywords: list[str] = Field(
        description="2-5 precise technical keywords or short phrases"
    )
    categories: list[str] = Field(
        default_factory=list,
        description='arXiv category codes, e.g. ["cs.LG", "stat.ML"]',
    )
    authors: list[str] = Field(
        default_factory=list,
        description="Author names, only if the question names them",
    )


def to_query_string(query: ArxivQuery) -> str:
    groups = []
    if query.keywords:
        groups.append("(" + " OR ".join(f'abs:"{k}"' for k in query.keywords) + ")")
    if query.categories:
        groups.append("(" + " OR ".join(f"cat:{c}" for c in query.categories) + ")")
    if query.authors:
        groups.append("(" + " OR ".join(f'au:"{a}"' for a in query.authors) + ")")
    return " AND ".join(groups)


structured_llm = init_chat_model("openai:gpt-5.4-mini").with_structured_output(
    ArxivQuery
)
retriever = ArxivRetriever(k=10)


def search(question: str):
    query = structured_llm.invoke(
        "Extract an arXiv search query from this question: " + question
    )
    return retriever.invoke(to_query_string(query))


docs = search("recent work on state space models for long-context language modeling")

For a question like the one above, the model might produce keywords=["state space models", "long context"] and categories=["cs.CL", "cs.LG"], which assembles to:

(abs:"state space models" OR abs:"long context") AND (cat:cs.CL OR cat:cs.LG)

Reranking arXiv results

arXiv ranks results lexically. When you need semantic precision, use the strategy this retriever was designed to compose with: over-fetch, then rerank locally against the original question. Here the reranker is a plain embeddings cosine similarity — swap in a dedicated reranker (e.g. Cohere Rerank) at the same seam if you have one.

from langchain.embeddings import init_embeddings
from langchain_core.documents import Document

from langchain_arxiv import ArxivRetriever

embeddings = init_embeddings("openai:text-embedding-3-small")
retriever = ArxivRetriever(k=25)  # over-fetch; arXiv's lexical ranking is pass one


def cosine(a: list[float], b: list[float]) -> float:
    dot = sum(x * y for x, y in zip(a, b))
    norm_a = sum(x * x for x in a) ** 0.5
    norm_b = sum(y * y for y in b) ** 0.5
    return dot / (norm_a * norm_b)


def rerank(question: str, docs: list[Document], top_n: int = 5) -> list[Document]:
    query_vector = embeddings.embed_query(question)
    doc_vectors = embeddings.embed_documents(
        [f"{doc.metadata['title']}\n{doc.page_content}" for doc in docs]
    )
    scored = sorted(
        zip(docs, doc_vectors),
        key=lambda pair: cosine(query_vector, pair[1]),
        reverse=True,
    )
    return [doc for doc, _ in scored[:top_n]]


question = "how do transformers learn in-context without weight updates?"
docs = rerank(question, retriever.invoke(question))

The two query-side examples and this one compose: rewrite (or structure) the query for recall, then rerank the over-fetched results for precision — all in the chain, while the retriever itself stays deterministic.

Using the retriever as agent tools

For agents, expose two tools built on two instances — a cheap search over abstracts and a full-text fetch for the paper the model selects. The docstrings teach the model to write keyword queries, so query quality becomes the agent's job:

from langchain.agents import create_agent
from langchain.tools import tool

from langchain_arxiv import ArxivRetriever

search_retriever = ArxivRetriever(k=5)
full_text_retriever = ArxivRetriever(full_text=True, max_content_chars=40_000)


@tool
def search_arxiv(query: str) -> str:
    """Search arXiv and return matching papers' IDs, titles, and abstracts.

    Write concise keyword queries, not questions. arXiv field syntax is
    supported: ti:"..." (title), abs:"..." (abstract), au:"..." (author),
    cat:cs.LG (category), combined with AND / OR / ANDNOT.
    """
    docs = search_retriever.invoke(query)
    if not docs:
        return "No results."
    return "\n\n".join(
        f"{doc.metadata['arxiv_id']} — {doc.metadata['title']}\n{doc.page_content}"
        for doc in docs
    )


@tool
def read_arxiv_paper(arxiv_id: str) -> str:
    """Fetch the full text of one paper by its arXiv ID (e.g. '2404.19756')."""
    (doc,) = full_text_retriever.invoke(arxiv_id, k=1)
    text = doc.page_content
    if doc.metadata["truncated"]:
        text += "\n\n[Text truncated at 40,000 characters.]"
    return text


agent = create_agent("openai:gpt-5.4-mini", [search_arxiv, read_arxiv_paper])

result = agent.invoke(
    {
        "messages": [
            {
                "role": "user",
                "content": "Find the paper that introduced Kolmogorov-Arnold "
                "networks and summarize what it claims about scaling laws.",
            }
        ]
    }
)

The max_content_chars cap keeps a full paper from blowing out the agent's context window — and because truncation is recorded in metadata["truncated"], the tool can tell the model the text was cut instead of letting it assume it read everything.

Switching modes at runtime

An instance's mode is fixed at construction, and instances are stateless and cheap — needing both usually means making two. If a single chain must switch modes per call, use LangChain's standard configurable_fields mechanism, which routes through the RunnableConfig that chains already propagate:

from langchain_core.runnables import ConfigurableField

from langchain_arxiv import ArxivRetriever

retriever = ArxivRetriever().configurable_fields(
    full_text=ConfigurableField(id="full_text")
)

retriever.invoke("2404.19756")  # abstract
retriever.invoke("2404.19756", config={"configurable": {"full_text": True}})  # whole paper