[ai-form-recognizer] Lazy iterator for words of a line #18444
Conversation
|
API changes have been detected in API changes + words?: () => IterableIterator<DocumentWord>; |
|
API changes have been detected in API changes + words: () => IterableIterator<DocumentWord>; |
|
/azp run js - ai-form-recognizer - tests |
|
Azure Pipelines successfully started running 1 pipeline(s). |
| * @returns a convenience layer DocumentLine | ||
| */ | ||
| function toDocumentLineFromGenerated( | ||
| generated: GeneratedDocumentLine, |
There was a problem hiding this comment.
NIT: perhaps the generated param can be typed as DocumentLine instead of being cast to it twice in the body.
There was a problem hiding this comment.
It doesn't satisfy DocumentLine until after the words method is added, so I'm reluctant to do this.
| it("search finds correct index", () => { | ||
| for (const datum of TEST_DATA) { | ||
| const testSpan = { offset: datum.span.offset, length: 1 }; | ||
| assert.strictEqual( | ||
| (iteratorFromFirstMatchBinarySearch(testSpan, TEST_DATA).next().value as TestData)?.id, | ||
| naiveFindFirst(testSpan, TEST_DATA)?.id | ||
| ); | ||
| } | ||
| }); |
There was a problem hiding this comment.
NIT: did you consider a quickcheck-style tests? I am not sure if JS has a good lib for it. Found https://github.com/gampleman/quick_check.js/ but not sure how good it is.
There was a problem hiding this comment.
I didn't consider this, but it's an interesting idea. When we support more generic kinds of element relationships, I'd like to add some fuzzing to this test matrix, but I will explore that during MQ.
sdk/formrecognizer/ai-form-recognizer/test/private/getChildren.spec.ts
Outdated
Show resolved
Hide resolved
* [ai-form-recognizer] Lazy iterator for words of a line * Use method instead of property * Regenerate API * Polished, wrote changelog, added some more tests, samples * Updated API MD * Improved docs * Apply changes from review
This PR adds support for accessing the
wordsof a DocumentLine. The computation is lazy, supported by a fast algorithm that assumes the words and spans are sorted. When the first element is iterated, a binary search is performed to find the first possible child element n the words array. Each subsequent iteration is nearly constant-time, as the elements are sorted and the algorithm optimizes the iteration by only comparing spans/words for as long as they are possible candidates for inclusion. The tests compare the results of this algorithm with a naive implementation.It also adds samples and a CHANGELOG entry.