Filter exact music identities before lexical and semantic candidate limits

This commit is contained in:
karl committed 2026-10-13 18:36:54 +02:00
1 parent 3ac1201909
commit 2eace15a33
7 files changed
+458 -19

No files matched your search

+6
View File
@@ -41,3 +41,9 @@ Audio timestamps represent segment starts, not exact word locations. Smaller
windows alone have not demonstrated accurate spoken navigation. No-answer
calibration, music discrimination, exact spoken positions, process isolation,
unchanged-vector reuse and production platform checks remain release gates.
Exact artist/title constraints use the core's normalized structured probe keys
before ANN limits and in the eligibility cache key. Metadata-only writes advance
the corpus revision and dirty path so warmed readers cannot retain old eligibility.
Empty metadata-only hybrid requests use core results without model inference.
These explicit constraints do not establish abstention for free-form topic queries.
+19 -1
View File
@@ -513,7 +513,7 @@ impl HybridSearch {
.lexical
.search_with_options_traced(query, candidates, options, trace, parent)?;
if query.trim().is_empty() || limit == 0 {
return Ok(Vec::new());
return Ok(lexical.into_iter().take(limit.min(1000)).collect());
}
trace.timed("hybrid.identifier_priority", parent, |_, _| {
promote_identifiers(query.trim(), &mut lexical)
@@ -776,6 +776,24 @@ mod tests {
let directory = tempfile::tempdir()?;
let db = directory.path().join("index.db");
let mut service = HybridSearch::with_provider(&db, Box::new(FailingProvider))?;
// The failure-only provider proves metadata-only searches never require inference.
let metadata = service.search_measured(
"",
5,
&SearchOptions {
media_artist: Some("Absent Artist".into()),
..Default::default()
},
true,
)?;
assert!(metadata.results.is_empty());
assert!(!metadata
.metrics
.trace
.unwrap()
.iter()
.any(|span| span.name == "semantic.query_embedding"));
let mut trace = TraceCollector::new(MetricsMode::Debug);
assert!(service
.search_traced(
+2 -8
View File
@@ -190,15 +190,9 @@ impl HybridSearch {
metadata: &MediaProbe,
change_revision: i64,
) -> Result<()> {
let tags = metadata
.tags
.iter()
.map(|(k, v)| format!("{k}: {v}"))
.collect::<Vec<_>>()
.join("\n");
self.lexical.set_media_metadata_at_revision(
self.lexical.set_structured_media_metadata_at_revision(
&path.to_string_lossy(),
&tags,
&metadata.tags,
metadata.duration_ms,
identity.bytes,
identity.modified_ns,
+143 -2
View File
@@ -31,6 +31,8 @@ struct FilterKey {
max_duration: Option<u32>,
source_type: Option<String>,
code_only: bool,
media_artist: Option<String>,
media_title: Option<String>,
}
struct EligibilityCache {
@@ -184,6 +186,17 @@ impl Store {
}
}
initialization.execute_batch("CREATE TABLE IF NOT EXISTS semantic_media_errors(generation INTEGER NOT NULL REFERENCES semantic_generations(id) ON DELETE CASCADE,path TEXT NOT NULL,message TEXT NOT NULL,PRIMARY KEY(generation,path))")?;
// Metadata-only writes change strict eligibility even when file bytes do not change.
for (name, event, row) in [
("insert", "INSERT", "new"),
("update", "UPDATE", "new"),
("delete", "DELETE", "old"),
] {
initialization.execute_batch(&format!("CREATE TRIGGER IF NOT EXISTS semantic_media_metadata_{name} AFTER {event} ON media_metadata BEGIN
UPDATE semantic_meta SET corpus_revision=corpus_revision+1 WHERE singleton=1;
INSERT INTO semantic_dirty_paths(path,revision) VALUES({row}.path,(SELECT corpus_revision FROM semantic_meta WHERE singleton=1))
ON CONFLICT(path) DO UPDATE SET revision=excluded.revision; END;"))?;
}
incremental::migrate(&initialization)?;
initialization.commit()?;
let canonical = db.canonicalize()?;
@@ -644,6 +657,16 @@ impl Store {
max_duration: options.max_duration_ms,
source_type: source_type.map(str::to_owned),
code_only: options.code_only,
media_artist: options
.media_artist
.as_deref()
.map(searchi_core::normalize_media_field)
.transpose()?,
media_title: options
.media_title
.as_deref()
.map(searchi_core::normalize_media_field)
.transpose()?,
};
trace.end(filter_span, Default::default());
let eligibility_span = trace.begin("semantic.eligibility", parent);
@@ -883,6 +906,8 @@ impl Store {
AND (?3 IS NULL OR f.extension=?3)
AND (?4 IS NULL OR f.size>=?4) AND (?5 IS NULL OR f.size<=?5)
AND (?6 IS NULL OR f.mtime/1000000>=?6) AND (?7 IS NULL OR f.mtime/1000000<=?7)
AND (?9 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.artist_key=?9))
AND (?10 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.title_key=?10))
ORDER BY c.passage_id,c.ordinal LIMIT ?8")?;
// Fetch one extra to detect the baseline cap instead of silently hiding
// candidates. Metadata filtering precedes this cap and all top-k limits.
@@ -894,7 +919,17 @@ impl Store {
max_size,
options.modified_after_unix_ms,
options.modified_before_unix_ms,
MAX_SCAN_CHUNKS as i64 + 1
MAX_SCAN_CHUNKS as i64 + 1,
options
.media_artist
.as_deref()
.map(searchi_core::normalize_media_field)
.transpose()?,
options
.media_title
.as_deref()
.map(searchi_core::normalize_media_field)
.transpose()?
])?;
let mut hashes = HashMap::<i64, String>::new();
let mut best = HashMap::<String, SearchResult>::new();
@@ -1008,6 +1043,8 @@ fn collect_eligibility(
AND (?6 IS NULL OR f.mtime/1000000>=?6) AND (?7 IS NULL OR f.mtime/1000000<=?7)
AND (?10 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.duration_ms>=?10))
AND (?11 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.duration_ms<=?11))
AND (?13 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.artist_key=?13))
AND (?14 IS NULL OR EXISTS(SELECT 1 FROM media_metadata m WHERE m.path=f.path AND m.title_key=?14))
AND (?12 IS NULL OR c.source_type=?12)
AND (?9 IS NULL OR c.path=?9) {code} ORDER BY c.passage_id,c.ordinal LIMIT ?8"))?;
let mut rows = statement.query(params![
@@ -1022,7 +1059,9 @@ fn collect_eligibility(
path,
filter.min_duration,
filter.max_duration,
filter.source_type
filter.source_type,
filter.media_artist,
filter.media_title
])?;
let mut hashes = HashMap::<i64, String>::new();
let mut keys = HashSet::new();
@@ -1293,6 +1332,108 @@ mod tests {
Ok(())
}
#[test]
fn exact_metadata_filters_apply_before_ann_and_invalidate_warmed_eligibility() -> Result<()> {
let dir = tempfile::tempdir()?;
let root = dir.path().join("media");
std::fs::create_dir(&root)?;
std::fs::write(root.join("first.ogg"), b"metadata ANN eligibility fixture")?;
std::fs::write(root.join("other.ogg"), b"metadata distractor fixture")?;
let db = dir.path().join("index.db");
let mut lexical = SearchIndex::open(&db)?;
lexical.index_folder(&root)?;
let mut store = Store::open(&db, DIMENSIONS)?;
let mut files = store.media_batch(0)?;
files.sort_by(|a, b| a.1.cmp(&b.1));
let (other_id, other_path, _, _) = files.pop().context("distractor fixture")?;
let other_identity = searchi_media::MediaIdentity::read(Path::new(&other_path), 1024)?;
let (id, path, _, _) = files.remove(0);
let identity = searchi_media::MediaIdentity::read(Path::new(&path), 1024)?;
let revision =
lexical.media_change_revision(&path, identity.bytes, identity.modified_ns)?;
let mut tags = std::collections::BTreeMap::from([
("artist".into(), "Björk".into()),
("title".into(), "Jóga".into()),
]);
lexical.set_structured_media_metadata_at_revision(
&path,
&tags,
1000,
identity.bytes,
identity.modified_ns,
revision,
)?;
// Deterministic geometry tests storage/filter behavior, not model quality.
let mut vector = vec![0.; DIMENSIONS];
vector[0] = 1.;
let generation = store.begin("metadata-filter-proof")?;
let mut weaker = vector.clone();
weaker[0] = 0.8;
weaker[1] = 0.6;
store.insert_media(generation, id, 0, &identity, "audio", 0, &weaker)?;
store.insert_media(
generation,
other_id,
0,
&other_identity,
"audio",
0,
&vector,
)?;
store.activate(generation)?;
let mut options = SearchOptions {
media_artist: Some("BJÖRK".into()),
media_title: Some("JÓGA".into()),
..Default::default()
};
assert_eq!(
store
.search(&vector, "metadata-filter-proof", 1, &options)?
.len(),
1
);
options.media_artist = Some("Absent".into());
assert!(store
.search(&vector, "metadata-filter-proof", 1, &options)?
.is_empty());
options.media_artist = Some("BJÖRK".into());
assert_eq!(
store
.search(&vector, "metadata-filter-proof", 1, &options)?
.len(),
1
);
tags.insert("artist".into(), "Replacement".into());
lexical.set_structured_media_metadata_at_revision(
&path,
&tags,
1000,
identity.bytes,
identity.modified_ns,
revision,
)?;
assert!(store
.search(&vector, "metadata-filter-proof", 1, &options)?
.is_empty());
options.media_artist = Some("Replacement".into());
assert_eq!(
store
.search(&vector, "metadata-filter-proof", 1, &options)?
.len(),
1
);
lexical.index_paths_cancellable(
&root,
&[std::path::PathBuf::from(&path)],
&searchi_core::IndexOptions::default(),
&std::sync::atomic::AtomicBool::new(false),
)?;
assert!(store
.search(&vector, "metadata-filter-proof", 1, &options)?
.is_empty());
Ok(())
}
#[test]
fn media_origins_persist_filters_and_changes_remove_stale_candidates() -> Result<()> {
let directory = tempfile::tempdir()?;