Compare papers
mosaic compare puts cached papers side by side in a table, one row per paper and one column per dimension you name: method, dataset, code, key finding. The language model of [llm] reads each title and abstract and fills the cells. Without one, mosaic fills only the columns it has as metadata.
Usage
mosaic compare --query "hall thruster" -n 10 # cached papers matching every word
mosaic compare --query "hall thruster" --sort citations -n 10 # the 10 most cited of them
mosaic compare --from refs.bib --dimensions "method,geometry,code,key finding"
mosaic compare --query "hall thruster" -n 10 --output hall.md # also write Markdown (.csv, .json)The papers come from the cache: with --query, those whose title or abstract contains every word; with --from, those of a .bib or .csv file that are in the cache (run mosaic get --from first for the others); with neither, the whole cache. --sort citations or --sort year orders them before --max (-n, default 20) keeps the first ones. The default dimensions are method,dataset,metric,result.
The terminal always shows the table (number, title, year, authors, then your dimensions); --output writes it as well, as a Markdown table, CSV with a header row, or a JSON list with one object per paper.
How it works
With a language model configured (see setting up the models), mosaic sends the papers in batches of 20, each as its title and the first 300 characters of its abstract, and asks for a JSON list of objects with exactly your dimensions as keys, - where the text says nothing. Any dimension name works: the model reads it as a question about the abstract.
Without a model, mosaic says so in one line and fills the table from metadata. When a request fails or the reply is not JSON, it falls back to metadata for the whole table and logs a warning (shown with mosaic --verbose). These dimension names are filled from the cache, every other one is –:
| Dimension | Filled from |
|---|---|
year | the publication year |
source | the source that returned the paper |
journal | the journal |
doi | the DOI |
authors | the short author list (Reza et al.) |
citations, citation_count, cited | the citation count, when a source reported one |
Limits
- The model sees 300 characters of each abstract, not the paper. A cell can only be as good as that text, and a model can misread it: treat the table as a first pass to check against the papers.
- Each batch of 20 papers is one request; a local model on a large selection takes minutes.
- Wide tables wrap in the terminal. Write
--output table.mdor.csvfor more than five or six dimensions.
See also: tutorial chapter 8, recipes, mosaic compare.