feat: integrate Google semantic search and branded desktop launcher

crates/searchi-semantic/tests/gemma_cpu.rs

@@ -0,0 +1,102 @@
+use anyhow::Result;
+use searchi_semantic::{Device, GemmaEmbedder, GEMMA_DIMENSIONS};
+
+#[test]
+#[ignore = "Requires explicitly provisioned Google model and CPU worker"]
+fn google_cpu_real_embeddings_and_verified_tail_windows() -> Result<()> {
+    let model = std::env::var_os("SEARCHI_GEMMA_MODEL_DIR").expect("Set SEARCHI_GEMMA_MODEL_DIR");
+    let worker = std::env::var_os("SEARCHI_GEMMA_WORKER").expect("Set SEARCHI_GEMMA_WORKER");
+    let mut embedder = GemmaEmbedder::load_with_device(model, worker, 1, Device::Cpu)?;
+    assert_eq!(embedder.info.device, "cpu");
+    assert_eq!(embedder.info.dimensions, GEMMA_DIMENSIONS);
+    let reference: serde_json::Value =
+        serde_json::from_str(include_str!("fixtures/digup-text-reference.json"))?;
+    let long_reference=(1..=60).map(|i|format!("Note {i}: the review moved to room {} because the projector in the main hall failed again, and the team agreed to send the slides the night before.",i%7+2)).collect::<Vec<_>>().join(" ");
+    let mut made = Vec::new();
+    for (name, query) in [
+        ("query", "payment declined error"),
+        ("query-bn", "ইলিশ মাছ রান্না"),
+        ("query-ar", "متى تغلق المكتبة"),
+    ] {
+        made.push((name, embedder.embed_query(query)?));
+    }
+    made.push(("document",embedder.embed_documents(&["Flight to Tokyo Haneda on November 3. Then the train to Kamakura to see the Great Buddha, and back for dinner in Shibuya."],"Tokyo itinerary")?.remove(0)));
+    made.push((
+        "document-long",
+        embedder.embed_documents(&[&long_reference], "")?.remove(0),
+    ));
+    for (name, vector) in made {
+        let expected = reference["vectors"][name].as_array().unwrap();
+        let cosine: f64 = vector
+            .iter()
+            .zip(expected)
+            .map(|(a, b)| f64::from(*a) * b.as_f64().unwrap())
+            .sum();
+        println!("DigUp text reference {name}: cosine {cosine:.8}");
+        assert!(
+            cosine >= 0.999,
+            "Google CPU reference parity failed for {name}: {cosine}"
+        );
+    }
+    let documents = [
+        "The supplier agreement may be ended with sixty days written notice.",
+        "Business train tickets and hotel stays can be reimbursed by submitting receipts.",
+        "Die Gartenplanung sieht ein Hochbeet für Gemüse und eine Regentonne vor.",
+    ];
+    let vectors = embedder.embed_documents(&documents, "none")?;
+    for (query, expected) in [
+        ("Wie beenden wir die Zusammenarbeit mit dem Lieferanten?", 0),
+        ("How do I claim my accommodation and railway expenses?", 1),
+        (
+            "Where is the plan for the vegetable bed and rainwater barrel?",
+            2,
+        ),
+    ] {
+        let query = embedder.embed_query(query)?;
+        let scores: Vec<f32> = vectors
+            .iter()
+            .map(|vector| vector.iter().zip(&query).map(|(a, b)| a * b).sum())
+            .collect();
+        let best = scores
+            .iter()
+            .enumerate()
+            .max_by(|a, b| a.1.total_cmp(b.1))
+            .unwrap()
+            .0;
+        assert_eq!(
+            best, expected,
+            "Semantic cross-language ranking failed: {scores:?}"
+        );
+    }
+    let long = format!("{}\nHotel accommodation on business trips is reimbursed after submitting the overnight stay receipt.",
+        "The routine maintenance log records ordinary uptime checks and cable inspections. ".repeat(350));
+    let windows = embedder.split_windows(&long, "maintenance-log.txt")?;
+    assert!(windows.len() > 1);
+    assert_eq!(windows.last().unwrap().end_byte, long.len());
+    assert!(windows
+        .last()
+        .unwrap()
+        .text
+        .contains("overnight stay receipt"));
+    let texts: Vec<_> = windows.iter().map(|window| window.text.as_str()).collect();
+    let vectors = embedder.embed_documents(&texts, "maintenance-log.txt")?;
+    let query = embedder.embed_query(
+        "Where does the policy say that lodging costs during work travel can be claimed?",
+    )?;
+    let scores: Vec<f32> = vectors
+        .iter()
+        .map(|vector| vector.iter().zip(&query).map(|(a, b)| a * b).sum())
+        .collect();
+    let best = scores
+        .iter()
+        .enumerate()
+        .max_by(|a, b| a.1.total_cmp(b.1))
+        .unwrap()
+        .0;
+    assert_eq!(
+        best,
+        windows.len() - 1,
+        "Tail retrieval did not find final window: {scores:?}"
+    );
+    Ok(())
+}