diff --git a/src/tags.rs b/src/tags.rs
index b94cb3b..e40e13c 100644
--- a/src/tags.rs
+++ b/src/tags.rs
@@ -32,6 +32,7 @@ where
     )
     .service(web::resource("image/tags/all").route(web::get().to(get_all_tags::<TagD>)))
     .service(web::resource("image/tags/batch").route(web::post().to(update_tags::<TagD>)))
+    .service(web::resource("image/tags/lookup").route(web::post().to(lookup_tags_batch::<TagD>)))
 }
 
 async fn add_tag<D: TagDao>(
@@ -238,6 +239,51 @@ async fn update_tags<D: TagDao>(
         .into_http_internal_err()
 }
 
+#[derive(Deserialize, Debug)]
+pub struct LookupTagsBatchRequest {
+    pub paths: Vec<String>,
+}
+
+/// Bulk per-path tag lookup. Apollo's photo-match flow used to fan out
+/// one ``GET /image/tags?path=`` per record (~4k for a wide window) —
+/// each call locked the dao briefly and the round-trip cost dwarfed
+/// the actual SQL. This collapses the whole fan-out into one POST and
+/// one (chunked) JOIN. Body: ``{paths: [...]}``; response:
+/// ``{path: [{id, name, ...}]}`` with **only paths that have at least
+/// one tag** in the map (the caller treats absence as empty list).
+///
+/// Trade-off: this matches by ``rel_path`` directly and does NOT do
+/// the cross-library content-hash sibling expansion that the per-path
+/// ``GET /image/tags`` does. For Apollo's grid view the simpler match
+/// is fine — it's the common case for single-library deploys; the
+/// carousel still uses the per-path endpoint and resolves siblings on
+/// demand. If multi-library content-hash sharing becomes load-bearing
+/// for the grid, extend this to JOIN ``image_exif`` on content_hash.
+async fn lookup_tags_batch<D: TagDao>(
+    _: Claims,
+    http_request: HttpRequest,
+    body: web::Json<LookupTagsBatchRequest>,
+    tag_dao: web::Data<Mutex<D>>,
+) -> impl Responder {
+    let context = extract_context_from_request(&http_request);
+    let span = global_tracer().start_with_context("lookup_tags_batch", &context);
+    let span_context = opentelemetry::Context::current_with_span(span);
+
+    if body.paths.is_empty() {
+        return HttpResponse::Ok().json(std::collections::HashMap::<String, Vec<Tag>>::new());
+    }
+
+    let normalized: Vec<String> = body.paths.iter().map(|p| normalize_path(p)).collect();
+    let mut dao = tag_dao.lock().expect("Unable to get TagDao");
+    match dao.get_tags_grouped_by_paths(&span_context, &normalized) {
+        Ok(grouped) => {
+            span_context.span().set_status(Status::Ok);
+            HttpResponse::Ok().json(grouped)
+        }
+        Err(e) => HttpResponse::InternalServerError().body(format!("{}", e)),
+    }
+}
+
 #[derive(Serialize, Queryable, Clone, Debug, PartialEq)]
 pub struct Tag {
     pub id: i32,
@@ -317,6 +363,14 @@ pub trait TagDao: Send + Sync {
         context: &opentelemetry::Context,
         paths: &[String],
     ) -> anyhow::Result<Vec<Tag>>;
+    /// Per-path grouped lookup: ``rel_path → [tags]``. Used by the
+    /// ``/image/tags/lookup`` batch endpoint. Returns only paths that
+    /// have at least one tag; the caller treats absence as empty.
+    fn get_tags_grouped_by_paths(
+        &mut self,
+        context: &opentelemetry::Context,
+        paths: &[String],
+    ) -> anyhow::Result<std::collections::HashMap<String, Vec<Tag>>>;
     fn create_tag(&mut self, context: &opentelemetry::Context, name: &str) -> anyhow::Result<Tag>;
     fn remove_tag(
         &mut self,
@@ -470,6 +524,51 @@ impl TagDao for SqliteTagDao {
         })
     }
 
+    fn get_tags_grouped_by_paths(
+        &mut self,
+        context: &opentelemetry::Context,
+        paths: &[String],
+    ) -> anyhow::Result<std::collections::HashMap<String, Vec<Tag>>> {
+        use std::collections::HashMap;
+        let mut out: HashMap<String, Vec<Tag>> = HashMap::new();
+        if paths.is_empty() {
+            return Ok(out);
+        }
+        let mut conn = self
+            .connection
+            .lock()
+            .expect("Unable to lock SqliteTagDao connection");
+        trace_db_call(context, "query", "get_tags_grouped_by_paths", |span| {
+            span.set_attribute(KeyValue::new("path_count", paths.len() as i64));
+            // SQLite's default SQLITE_LIMIT_VARIABLE_NUMBER is 32766 in
+            // modern builds (999 in old ones). Chunk at 500 to stay
+            // safely under both — five queries for a 4k-photo grid is
+            // still ~800x cheaper than 4k single-row HTTP calls.
+            const CHUNK: usize = 500;
+            for chunk in paths.chunks(CHUNK) {
+                let rows: Vec<(String, i32, String, i64)> = tagged_photo::table
+                    .inner_join(tags::table)
+                    .filter(tagged_photo::rel_path.eq_any(chunk))
+                    .select((
+                        tagged_photo::rel_path,
+                        tags::id,
+                        tags::name,
+                        tags::created_time,
+                    ))
+                    .get_results(conn.deref_mut())
+                    .with_context(|| "Unable to get tags grouped from Sqlite")?;
+                for (rel_path, id, name, created_time) in rows {
+                    out.entry(rel_path).or_default().push(Tag {
+                        id,
+                        name,
+                        created_time,
+                    });
+                }
+            }
+            Ok(out)
+        })
+    }
+
     fn create_tag(&mut self, context: &opentelemetry::Context, name: &str) -> anyhow::Result<Tag> {
         let mut conn = self
             .connection
@@ -893,6 +992,23 @@ mod tests {
             Ok(out)
         }
 
+        fn get_tags_grouped_by_paths(
+            &mut self,
+            _context: &opentelemetry::Context,
+            paths: &[String],
+        ) -> anyhow::Result<std::collections::HashMap<String, Vec<Tag>>> {
+            let tagged = self.tagged_photos.borrow();
+            let mut out = std::collections::HashMap::new();
+            for p in paths {
+                if let Some(tags) = tagged.get(p)
+                    && !tags.is_empty()
+                {
+                    out.insert(p.clone(), tags.clone());
+                }
+            }
+            Ok(out)
+        }
+
         fn create_tag(
             &mut self,
             _context: &opentelemetry::Context,