add more reasonable arxiv retriever (#13327)

libs/langchain/langchain/retrievers/arxiv.py

@@ -12,7 +12,12 @@ class ArxivRetriever(BaseRetriever, ArxivAPIWrapper):
     It uses all ArxivAPIWrapper arguments without any change.
     """
 
+    get_full_documents: bool = False
+
     def _get_relevant_documents(
         self, query: str, *, run_manager: CallbackManagerForRetrieverRun
     ) -> List[Document]:
+        if self.get_full_documents:
             return self.load(query=query)
+        else:
+            return self.get_summaries_as_docs(query)

libs/langchain/langchain/utilities/arxiv.py

@@ -90,6 +90,43 @@ class ArxivAPIWrapper(BaseModel):
             )
         return values
 
+    def get_summaries_as_docs(self, query: str) -> List[Document]:
+        """
+        Performs an arxiv search and returns list of
+        documents, with summaries as the content.
+
+        If an error occurs or no documents found, error text
+        is returned instead. Wrapper for
+        https://lukasschwab.me/arxiv.py/index.html#Search
+
+        Args:
+            query: a plaintext search query
+        """  # noqa: E501
+        try:
+            if self.is_arxiv_identifier(query):
+                results = self.arxiv_search(
+                    id_list=query.split(),
+                    max_results=self.top_k_results,
+                ).results()
+            else:
+                results = self.arxiv_search(  # type: ignore
+                    query[: self.ARXIV_MAX_QUERY_LENGTH], max_results=self.top_k_results
+                ).results()
+        except self.arxiv_exceptions as ex:
+            return [Document(page_content=f"Arxiv exception: {ex}")]
+        docs = [
+            Document(
+                page_content=result.summary,
+                metadata={
+                    "Published": result.updated.date(),
+                    "Title": result.title,
+                    "Authors": ", ".join(a.name for a in result.authors),
+                },
+            )
+            for result in results
+        ]
+        return docs
+
     def run(self, query: str) -> str:
         """
         Performs an arxiv search and A single string