[darshjoshi]

The category filter of the Forge ArxivSearchTool leaks results from other categories whenever the query has more than one word.

The tool builds all:<query> AND cat:<category>. arXiv only applies the all: prefix to the first word, and AND binds tighter than the implicit operator between words. So all:diffusion model AND cat:cs.CV effectively means "diffusion anywhere, or model in cs.CV". The same happens with prefixed queries like ti:RAG OR abs:agents.

Searching "diffusion model" with category="cs.CV" and 50 results:

main:  50 results, 9 outside cs.CV [('1109.2262', ['cond-mat.mtrl-sci']), ('1109.2261', ['cond-mat.mtrl-sci']), ('1701.00257', ['physics.chem-ph'])]
fixed: 50 results, 0 outside cs.CV []

Over 100 results against the raw API, 26 fall outside cs.CV with the current query and none with the grouped one.

The fix wraps the query in parentheses before appending the category, producing (all:diffusion model) AND cat:cs.CV. Queries without a category are unchanged. I updated the two existing expectations and added a test for multi-word and prefixed queries; both fail on main. The tool's tests pass, and black --check and flake8 are clean.

[Eigenwise]

Grouping the full query before AND applies the category to both OR branches. Verified the regression failures on main and all 10 arXiv tool tests pass with this change. No live API request was needed. Thanks.