[dny238]

Fixes #1068

Problem

Automatic search for a book could return 0 results even when the book is available, because the auto-built query is Title Author (Series) and many release names don't carry the author.

I verified this read-only against NZBgeek's API: …Full Cast Edition returned 5 results, but appending J.K. Rowling dropped it to 0 (release names omit the author); and over-specific multi-word queries AND-matched to nothing. (De-hyphenating made no difference — that's a separate normalization, see #1070.)

Fix

AutomaticSearchService now retries once with an author-free query (title + series) when the full query returns no results.

Safety: the automatic path scores results by quality only — relevance comes entirely from the query — so broadening the query could otherwise let a same-title/different-book release pass scoring and be grabbed. The fallback therefore constrains its results to titles that still match the book (ConstrainFallbackResultsToTitle, built on TitleUtils.IsMatchingTitle) before they reach scoring. Distinctive multi-word/long titles recover their releases; short or generic titles (where same-title collisions are likely) fail the match and are safely dropped, so nothing unrelated is grabbed.

Testing

AutomaticSearchResultClassifierTests:

• the author-free query drops the author but keeps title + series;
• the title constraint keeps a matching release (Andy Weir - Artemis - 2017 - M4B for Artemis) and drops an unrelated one (… The Martian …);
• empty input returns empty.

The retry orchestration itself lives in the hosted-service loop (ProcessAudiobookAsync), which resolves ~7 services + a SignalR hub from a scope; happy to add a scope-driven integration test if you'd like it, but I kept the safety-critical pieces (query shape + title constraint) in the testable classifier.

🤖 Generated with Claude Code

[therobbiedavis]

The author-free retry is useful, but the title guard still lets a different book reach automatic download selection. I reproduced this in a live container pass. This needs an identity check before the broadened results reach quality scoring.

[dny238]

Closing in favor of #825, which already tracks the title-only retry — thanks for the pointer. I agree the retry is a band-aid over the real selection issue (ranking by quality before establishing the release matches the requested book), and #1054 is reworking the search path in that direction. The one piece worth carrying forward is constraining results to titles that match the requested book before quality ranking (the ConstrainFallbackResultsToTitle guard here); happy to fold that relevance idea into #825/#1054 if it's useful. Apologies for the duplicate.

[Yoldark34]

Backend matching is delegated to #1076

Any match revision before the step of doing a backend matching is welcomed IMO. We should revise matching because I will not change the buisness logic rules during the backend refactoring.

This PR should stay open then.

[dny238]

Reopened — sorry, I'd closed this before your review landed. Thanks for catching the collision and for the detailed candidate; that's a real wrong-book hazard and the containment guard clearly wasn't enough.

Replaced IsMatchingTitle with CanAutomaticallySelectFallback following your direction: it parses only Title and Author - Title layouts (stripping trailing year/format suffixes), requires the complete remaining title to match the book exactly after normalization, and requires at least one corroborating signal — parsed author, Artist, or ASIN — with conflicting evidence vetoing. Artist only corroborates when it matches (never treated as confirmation just for being populated), and uncorroborated authorless results are skipped.

Tests (AutomaticSearchResultClassifierTests): a valid author-matched fallback is kept; both of your collisions are rejected — Eoin Colfer - Artemis Fowl for Artemis (title mismatch) and Chris Weitz - Revival for Stephen King's Revival (conflicting author); an uncorroborated authorless Artemis is skipped; and an Artist-field match corroborates.

On the full-cycle coverage: the guard is what gates fallback results before they reach selection, so these assert the collisions can't be selected. I kept them at the classifier level — the retry orchestration lives in the hosted-service loop (resolves ~7 services + a SignalR hub from a scope), which your own candidate's tests also stopped short of. Happy to add a scope-driven handoff test if you'd like that layer too.

[dny238]

Added the full-cycle regression coverage you asked for. The previous tests stopped at the classifier (ConstrainFallbackResultsToTitle); the new AutomaticSearchProcessorFallbackTests drive the whole automatic-search cycle via AutomaticSearchProcessor.RunCycleAsync and assert at the download-service handoff:

• Same-word different book — Eoin Colfer's Artemis Fowl for Andy Weir's Artemis → StartDownloadAsync is never called (strict IDownloadService mock).
• Same-title different author — Chris Weitz's Revival for Stephen King's → StartDownloadAsync is never called.
• Correctly attributed fallback — Andy Weir - Artemis → StartDownloadAsync is called once with the requested book's id.

Both collisions are the ones you reproduced reaching StartDownloadAsync in the live container pass; these now lock that down at the cycle level. The valid case mocks the quality scorer (its own tests cover scoring) and uses a DDL result so the client selector resolves without configuring a download client. Pushed as 1a98902.

[therobbiedavis]

@dny238 The updated guard addresses the wrong-book cases I reproduced. I have one non-blocking follow-up suggestion for numeric titles: the missed recovery below also occurs before this PR, so I wouldn't hold this change up for it.

[therobbiedavis]

Good follow-up PR candidate: could we preserve numeric book titles before treating trailing numbers as publication years? For George Orwell - 1984 - M4B, this loop removes both M4B and 1984, leaving George Orwell as the parsed title and dropping the correct release. David McCullough - 1776 - 2005 - M4B behaves the same way. This edge case was also in the prototype I suggested.

I compared the pre-PR path and this head in a live HTTP indexer container. When the full query returns nothing, neither selects these numeric-title releases; when it returns a valid 1984 result, both select it. The suffix parsing is new, but the missed recovery is not a user-visible regression—the previous path never retried.

For a follow-up, stop removing a segment when it matches the requested title:

while (parts.Count > 1
    && !string.Equals(
        NormalizeIdentity(parts[^1]),
        NormalizeIdentity(audiobook.Title),
        StringComparison.Ordinal)
    && Regex.IsMatch(
        parts[^1],
        @"^(?:\d{4}|\[?(?:m4b|mp3|flac|aac|ogg|opus)\]?)$",
        RegexOptions.IgnoreCase))
{
    parts.RemoveAt(parts.Count - 1);
}

That candidate recovered both numeric titles in the same container while still rejecting conflicting-author and uncorroborated authorless results. Add cases for a numeric title alone, with a format suffix, and with a separate publication year, retaining those negative controls. Validation stops at the recorded StartDownloadAsync handoff; no actual media download or import was exercised.