Nomad changes

crates/kreuzberg/tests/pdf_markdown_quality.rs

@@ -0,0 +1,157 @@
+//! PDF markdown quality smoke tests: verify extraction produces structural elements.
+//!
+//! These are lightweight assertions — detailed quality scoring and A/B comparisons
+//! live in `tools/benchmark-harness` (subcommands: `compare`, `pipeline-benchmark`).
+//!
+//! Usage:
+//!   cargo test -p kreuzberg --features "pdf" \
+//!     --test pdf_markdown_quality -- --nocapture
+
+#![cfg(feature = "pdf")]
+
+mod helpers;
+
+use helpers::*;
+use kreuzberg::core::config::{ExtractionConfig, OutputFormat};
+use kreuzberg::extract_file_sync;
+
+/// Documents with markdown ground truth.
+const MARKDOWN_GT_DOCS: &[(&str, &str)] = &[("docling", "pdf/docling.pdf")];
+
+fn extract_markdown(pdf_path: &std::path::Path) -> String {
+    let config = ExtractionConfig {
+        output_format: OutputFormat::Markdown,
+        ..Default::default()
+    };
+    extract_file_sync(pdf_path, None, &config)
+        .expect("extraction should succeed")
+        .content
+}
+
+#[cfg(feature = "layout-detection")]
+fn extract_markdown_with_layout(pdf_path: &std::path::Path) -> String {
+    use kreuzberg::core::config::layout::LayoutDetectionConfig;
+
+    let config = ExtractionConfig {
+        output_format: OutputFormat::Markdown,
+        layout: Some(LayoutDetectionConfig::default()),
+        ..Default::default()
+    };
+    extract_file_sync(pdf_path, None, &config)
+        .expect("layout extraction should succeed")
+        .content
+}
+
+/// Count structural elements in markdown content.
+fn count_headings(md: &str) -> usize {
+    md.lines().filter(|l| l.starts_with('#')).count()
+}
+
+fn count_table_rows(md: &str) -> usize {
+    md.lines()
+        .filter(|l| l.starts_with('|') && l.ends_with('|') && !l.contains("---"))
+        .count()
+}
+
+fn count_list_items(md: &str) -> usize {
+    md.lines()
+        .filter(|l| {
+            let t = l.trim_start();
+            t.starts_with("- ") || t.starts_with("* ") || t.starts_with("+ ")
+        })
+        .count()
+}
+
+fn has_code_blocks(md: &str) -> bool {
+    md.contains("```")
+}
+
+#[test]
+fn test_baseline_produces_structural_markdown() {
+    if !test_documents_available() {
+        println!("Skipping: test_documents not available");
+        return;
+    }
+
+    for &(name, pdf_rel) in MARKDOWN_GT_DOCS {
+        let pdf_path = get_test_file_path(pdf_rel);
+        if !pdf_path.exists() {
+            println!("Skipping {}: file not found", name);
+            continue;
+        }
+
+        let content = extract_markdown(&pdf_path);
+
+        // Basic structural assertions
+        assert!(
+            !content.trim().is_empty(),
+            "{}: extraction produced empty content",
+            name
+        );
+        assert!(
+            content.len() > 500,
+            "{}: content too short ({} chars)",
+            name,
+            content.len()
+        );
+        assert!(count_headings(&content) > 0, "{}: expected at least one heading", name);
+
+        println!(
+            "{}: {} chars, {} headings, {} table rows, {} list items, code={}",
+            name,
+            content.len(),
+            count_headings(&content),
+            count_table_rows(&content),
+            count_list_items(&content),
+            has_code_blocks(&content),
+        );
+    }
+}
+
+#[cfg(feature = "layout-detection")]
+#[test]
+fn test_layout_does_not_regress_text_content() {
+    if !test_documents_available() {
+        println!("Skipping: test_documents not available");
+        return;
+    }
+
+    for &(name, pdf_rel) in MARKDOWN_GT_DOCS {
+        let pdf_path = get_test_file_path(pdf_rel);
+        if !pdf_path.exists() {
+            println!("Skipping {}: file not found", name);
+            continue;
+        }
+
+        let baseline = extract_markdown(&pdf_path);
+        let layout = extract_markdown_with_layout(&pdf_path);
+
+        // Layout extraction should not lose significant content
+        let baseline_len = baseline.len();
+        let layout_len = layout.len();
+
+        // Allow up to 20% content loss (layout may restructure)
+        assert!(
+            layout_len as f64 >= baseline_len as f64 * 0.8,
+            "{}: layout content ({} chars) is significantly shorter than baseline ({} chars)",
+            name,
+            layout_len,
+            baseline_len,
+        );
+
+        // Layout should still have headings
+        assert!(
+            count_headings(&layout) > 0,
+            "{}: layout extraction lost all headings",
+            name
+        );
+
+        println!(
+            "{}: baseline={} chars, layout={} chars, layout headings={}",
+            name,
+            baseline_len,
+            layout_len,
+            count_headings(&layout),
+        );
+    }
+}