quick_search/crates/quicksearch-core/tests/fixtures/legacy

LibreOffice fixtures

sample.doc, sample.xls, sample.ppt, sample.odt, sample.ods, sample.odp, sample.pptx and sample.rtf, written by LibreOffice and committed. Read by tests/extraction_corpus.rs through the corpus module in tests/corpus/legacy.rs.

Why these are committed when the rest of the corpus is generated

Every other format in the extraction corpus is written at test time by code that is not the code that reads it back — the whole point being that a fixture built with the reader's own library can only prove the two agree with each other. The pre-2007 binary formats have no such library: cfb is the only Rust crate that writes OLE2 compound files, and cfb is what extract::ole reads them with.

So a foreign producer writes them once, here, and the output is committed. That buys more than independence — a LibreOffice .doc is a real FIB with a real piece table, its .xls a real BIFF stream with a real SST, and its .ppt drags the master slide's placeholder prompts (Click to edit the title text format, ___PPT10) into the text stream alongside the content. None of those shapes come out of a minimal synthetic fixture, and the last one is why the corpus asserts ordered containment rather than equality.

The .odt, .ods, .odp, .pptx and .rtf are the same argument one step removed. No Rust crate writes them either, so the corpus hand-assembles its own from the spec (tests/corpus/odf.rs, tests/corpus/ooxml.rs, tests/corpus/rtf.rs): independent of the reader's code, but written by the reader's author from the same reading of the spec. LibreOffice's are a real producer's: ODF packages with their styles, settings and thumbnails, a .pptx with its master, layouts, theme and relationships, and an .rtf with font, colour and style tables and each non-ASCII character as a \uN escape followed by a \'hh fallback the reader must skip.

The unit tests in src/extract/ole_tests.rs are the complement, not a duplicate: they build deliberately malformed streams to check bounds and error paths, which needs byte-level control that only cfb gives.

Files

file	role
prose.txt	source for sample.doc, sample.odt and sample.rtf
sheet.csv	source for sample.xls and sample.ods — one cell per row, no commas
deck.fodp	source for sample.ppt, sample.odp and sample.pptx — flat ODF, so the text is reviewable
sample.doc sample.xls sample.ppt sample.odt sample.ods sample.odp sample.pptx sample.rtf	LibreOffice output, committed
regen.sh	regenerates the eight from the sources

The sources are the source of truth: legacy.rs reads the expected fragments out of them rather than restating the text, so a regenerated fixture that dropped a line fails the test instead of quietly redefining what it should contain. Each source carries a needle (chalcedony9001-chalcedony9003) that the end-to-end search looks for; legacy.rs asserts it is still present. The other five fixtures are converted from copies of their source whose needle regen.sh renumbers (chalcedony9004-chalcedony9008), so every needle still names exactly one file.

Regenerating

./regen.sh          # needs libreoffice (Writer, Calc, Impress) on PATH
cargo test -p quicksearch-core --test extraction_corpus

The .doc, .xls and .ppt were written by 26.2.4.2, the other five by 26.2.6.3.

Only the .doc, .xls, .ppt and .rtf are byte-reproducible (26.2.6.3 rewrites 26.2.4.2's three unchanged). LibreOffice stamps the time into each .odt, .ods, .odp and .pptx, so re-running changes their bytes without changing the text. Commit the result only when the sources changed; the test is what says the files are still right.

sample.ppt is ~460 KB because the PowerPoint 97 export filter embeds the master slide. There is no filter option that trims it, and it is well inside the 50 MiB maximum_text_file_size the end-to-end run indexes with.

Unlike the generated half of the corpus, the text in these eight is fixed: QUICKSEARCH_CORPUS_SEED does not affect them.