Conversation
This commit introduces workflow_dispatch inputs for specifying branch names or tags for multiple repositories in the CI pipeline. It enables flexibility for testing specific versions of dependencies, with defaults set to 'main.' Additionally, a conditional check ensures local publishing to GitHub Pages based on input or event type.
Update ci for workflow build
update to test the empty list fix in the index.
Implemented parsing logic to resolve repository and reference overrides for multiple dependencies dynamically based on inputs. Updated checkout steps to use resolved values, improving flexibility and maintainability.
- Restored intuitive “type‑ahead” results for underscore/hyphenated terms (e.g., Vulkan extension names) without changing the index format. - Begins‑with fallback: if a query term contains `_` or `-`, we now use the raw lowercased prefix (`original + '*'`, pipeline disabled) so composite tokens like `vk_nv_copy` match early during incremental typing (`vk`, `vk_`, `vk_nv`, …). - Contains fallback: continues to use a normalized/stemmed last segment substring (`*segStem*`, pipeline disabled) for robust matching across inflections while keeping recall. - Preserved previous fixes: - Case‑insensitive wildcard matching. - Vendor‑order retry (`NV_VK_*` ↔ `VK_NV_*`) only on zero results. - Split‑index rendering guard: no blank rows; show a single “No results found …” dataset when there are no hits. - Index‑side fallbacks ensure non‑empty `title` and `text` in all docs.
- Segment‑aware wildcard search for underscore/hyphen tokens - Begins‑with fallback: if a query term contains `_` or `-`, use the raw, lower‑cased composite prefix (`original + '*'`) so partial typing like `vk_nv_c`, `vk_nv_copy`, `vk_nv_copy_mem` yields matches early. - Contains fallback: use the last segment’s normalized/stemmed token with leading/trailing wildcards (`*seg*`) to match inflections like `copy`→`copi`. - Case‑insensitivity preserved; pipeline is bypassed for wildcard phases but we pre‑normalize as needed. - Stabilized results for underscore/hyphen queries - Skip trie prefiltering when the input contains `_` or `-` to avoid over‑restrictive candidate pruning that caused result churn during incremental typing. - Ranking improvements - Increased field boost for `name` (page/extension name) so extension pages rank above general mentions: `name: 15`, `title: 10`, `keyword: 5`. Also mirrored this boost in the temporary filtered index. - Vendor‑order retry retained - Zero‑result safety net re‑tries `NV_VK_*` ↔ `VK_NV_*` queries once (allow‑listed vendors), unchanged.
|
Telecon: @oddhack will build locally to review how search works, can be merged with their approval. |
|
@gpx1000 I built locally and compared to the current published docs site. The immediate issue, of the empty list results, does seem not to appear in superficial testing - e.g. I can put in "VK_KHR_spir" and see just 3 matching results, without empty list boxes afters. But there appear to be some pretty serious regressions, too. For example, putting in 'format' in the docs site returns several pages of matching results, while with this branch, it shows nothing - not even an empty results box, but nothing at all under the searchbox. However, putting in a prefix like 'forma' does return much the same list as the docs site with the same term (in somewhat different order, which may not mean anything). I don't think we can publish as this stands. Using a common word as a search term and returning nothing is not OK. Related, having a test framework that we could hang various common queries off to look for sensible results would be very helpful and likely would have caught this behavior immediately, as 'format' is a pretty common query. |
oddhack
left a comment
oddhack
left a comment
There was a problem hiding this comment.
Needs to return relevant results for simple search terms, as noted in my previous comment, before we can accept this.
2 participants
DO NOT MERGE THIS DIRECTLY.
This points back to a PR repo under review. When the index changes in antora-lunr-extension are merged from my repo, then update to point at that commit and then merge.