kreuzberg 4.9.1 → 4.9.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 559b3104e6e21f2f14a92949d427703b51ea8b35b7a643d8964b6953785aa6e1
-  data.tar.gz: bfce92579c45ecba0da0d8e1f077ecca0bb9dd6a1e96c950e8beb8d0a39b5884
+  metadata.gz: 3d8a203168595f6b316a165f500818abed75d89c7a82c46b5b20df996a4bb841
+  data.tar.gz: 28fd19fecd9b18597f17a783923ec3ec08cfa7b99612fec1ca8790aa5cdddbdc
 SHA512:
-  metadata.gz: 1ea8af57d65eb5008126758041df2bfe07acca9d47ebfbf9c9de79f12b5d5ff2336d55b643268d1ea420db1825eaf9bfef6e5deb7335bd2449e9ccb62800492d
-  data.tar.gz: '019f2abaa7dcaf2b91925f7ed0b5332ce569a41a5ae1f152d698423cce1276ea68e177c2088b4377c369943e235c70584f61a76da5600d6dd8b3bd075bc266ab'
+  metadata.gz: af522bff519c1082396d9a6a9480a088693791a4f50818fd1d233726082675559a3144f7758b0ee217b18cdbb1cd08236ecbb332f68c4186a26aa69b83454392
+  data.tar.gz: 4109e6dbc32c5fed518ba84940a7bc732553a178d3b67b69dad6eff5b998aad12b0a53cfe6c6d0784848cbb484e7b79a7abe9154f23d2100786f38414d8286c0

data/README.md

@@ -22,7 +22,7 @@
     <img src="https://img.shields.io/maven-central/v/dev.kreuzberg/kreuzberg?label=Java&color=007ec6" alt="Java">
   </a>
   <a href="https://github.com/kreuzberg-dev/kreuzberg/releases">
-    <img src="https://img.shields.io/github/v/tag/kreuzberg-dev/kreuzberg?label=Go&color=007ec6&filter=v4.9.1" alt="Go">
+    <img src="https://img.shields.io/github/v/tag/kreuzberg-dev/kreuzberg?label=Go&color=007ec6&filter=v4.9.4" alt="Go">
   </a>
   <a href="https://www.nuget.org/packages/Kreuzberg/">
     <img src="https://img.shields.io/nuget/v/Kreuzberg?label=C%23&color=007ec6" alt="C#">

data/ext/kreuzberg_rb/native/Cargo.lock

@@ -2127,9 +2127,9 @@ dependencies = [
 
 [[package]]
 name = "html-to-markdown-rs"
-version = "3.2.5"
+version = "3.2.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bcb619abe81160bba2e2185823e10f6c0793220a266f16791aa715287de322cd"
+checksum = "bc4b9f5076d013aac34a0369c73035cf68f3d9e0771ce96a99e5a02e7e3bf9d4"
 dependencies = [
  "ahash",
  "astral-tl",
@@ -2916,7 +2916,7 @@ dependencies = [
 
 [[package]]
 name = "kreuzberg-rb"
-version = "4.9.1"
+version = "4.9.3"
 dependencies = [
  "async-trait",
  "html-to-markdown-rs",
@@ -3040,9 +3040,9 @@ checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
 
 [[package]]
 name = "liter-llm"
-version = "1.2.1"
+version = "1.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1884be380e45da823105c85ef0fa188af81d57be7de9b65016576e1774fdd5f8"
+checksum = "4e4ce5d2d0b09f2e63537ba40b15b0a95c2d6818ed0454eb04d9593ba4a0cad3"
 dependencies = [
  "base64 0.22.1",
  "bytes",
@@ -3634,9 +3634,9 @@ dependencies = [
 
 [[package]]
 name = "openssl"
-version = "0.10.77"
+version = "0.10.78"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bfe4646e360ec77dff7dde40ed3d6c5fee52d156ef4a62f53973d38294dad87f"
+checksum = "f38c4372413cdaaf3cc79dd92d29d7d9f5ab09b51b10dded508fb90bb70b9222"
 dependencies = [
  "bitflags",
  "cfg-if",
@@ -3666,9 +3666,9 @@ checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
 
 [[package]]
 name = "openssl-sys"
-version = "0.9.113"
+version = "0.9.114"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ad2f2c0eba47118757e4c6d2bff2838f3e0523380021356e7875e858372ce644"
+checksum = "13ce1245cd07fcc4cfdb438f7507b0c7e4f3849a69fd84d52374c66d83741bb6"
 dependencies = [
  "cc",
  "libc",
@@ -4619,9 +4619,9 @@ checksum = "f87165f0995f63a9fbeea62b64d10b4d9d8e78ec6d7d51fb2125fda7bb36788f"
 
 [[package]]
 name = "rustls-webpki"
-version = "0.103.12"
+version = "0.103.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8279bb85272c9f10811ae6a6c547ff594d6a7f3c6c6b02ee9726d1d0dcfcdd06"
+checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
 dependencies = [
  "aws-lc-rs",
  "ring",
@@ -5506,7 +5506,7 @@ dependencies = [
  "toml_datetime 1.1.1+spec-1.1.0",
  "toml_parser",
  "toml_writer",
- "winnow 1.0.1",
+ "winnow 1.0.2",
 ]
 
 [[package]]
@@ -5533,7 +5533,7 @@ version = "1.1.2+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a2abe9b86193656635d2411dc43050282ca48aa31c2451210f4202550afb7526"
 dependencies = [
- "winnow 1.0.1",
+ "winnow 1.0.2",
 ]
 
 [[package]]
@@ -6577,9 +6577,9 @@ checksum = "df79d97927682d2fd8adb29682d1140b343be4ac0f08fd68b7765d9c059d3945"
 
 [[package]]
 name = "winnow"
-version = "1.0.1"
+version = "1.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09dac053f1cd375980747450bfc7250c264eaae0583872e845c0c7cd578872b5"
+checksum = "2ee1708bef14716a11bae175f579062d4554d95be2c6829f518df847b7b3fdd0"
 
 [[package]]
 name = "wit-bindgen"

data/ext/kreuzberg_rb/native/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "kreuzberg-rb"
-version = "4.9.1"
+version = "4.9.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -65,7 +65,7 @@ tokio = { version = "1.52.1", features = [
     "time",
     "io-util",
 ] }
-html-to-markdown-rs = { version = "3.2.5", default-features = false }
+html-to-markdown-rs = { version = "3.2.6", default-features = false }
 
 [dev-dependencies]
 pretty_assertions = "1.4"

data/ext/kreuzberg_rb/native/src/config/types.rs

@@ -404,6 +404,12 @@ pub fn parse_image_extraction_config(ruby: &Ruby, hash: RHash) -> Result<ImageEx
         true
     };
 
+    let max_images_per_page = if let Some(val) = get_kw(ruby, hash, "max_images_per_page") {
+        Some(u32::try_convert(val)?)
+    } else {
+        None
+    };
+
     let config = ImageExtractionConfig {
         extract_images,
         target_dpi,
@@ -412,6 +418,7 @@ pub fn parse_image_extraction_config(ruby: &Ruby, hash: RHash) -> Result<ImageEx
         auto_adjust_dpi,
         min_dpi,
         max_dpi,
+        max_images_per_page,
     };
 
     Ok(config)

data/lib/kreuzberg/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Kreuzberg
-  VERSION = '4.9.1'
+  VERSION = '4.9.4'
 end

data/vendor/Cargo.toml

@@ -2,7 +2,7 @@
 members = ["kreuzberg", "kreuzberg-ffi", "kreuzberg-tesseract", "kreuzberg-paddle-ocr", "kreuzberg-pdfium-render"]
 
 [workspace.package]
-version = "4.9.1"
+version = "4.9.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -28,12 +28,12 @@ dbase = "0.7"
 futures = "0.3"
 getrandom = { version = "0.4.2", features = ["wasm_js"] }
 hex = "0.4.3"
-html-to-markdown-rs = { version = "3.2.5", default-features = false }
+html-to-markdown-rs = { version = "3.2.6", default-features = false }
 image = { version = "0.25.10", default-features = false }
 itertools = "0.14"
 js-sys = "0.3"
-kreuzberg = { path = "./crates/kreuzberg", version = "4.9.1", default-features = false }
-kreuzberg-ffi = { path = "./crates/kreuzberg-ffi", version = "4.9.1" }
+kreuzberg = { path = "./crates/kreuzberg", version = "4.9.4", default-features = false }
+kreuzberg-ffi = { path = "./crates/kreuzberg-ffi", version = "4.9.4" }
 lazy_static = "1.5.0"
 libc = "0.2.185"
 liter-llm = { version = "1.2", features = ["native-http", "tracing"], default-features = false }
@@ -45,7 +45,7 @@ num_cpus = "1.17.0"
 once_cell = "1.21.4"
 ort = { version = "2.0.0-rc.12", features = ["std", "api-18"], default-features = false }
 parking_lot = "0.12.5"
-pdf_oxide = { version = "0.3.34", default-features = false }
+pdf_oxide = { version = "0.3.37", default-features = false }
 pdfium-render = { package = "kreuzberg-pdfium-render", path = "crates/kreuzberg-pdfium-render", version = "4.3" }
 rayon = "1.12.0"
 reqwest = { version = "0.13.2", default-features = false }

data/vendor/kreuzberg/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "kreuzberg"
-version = "4.9.1"
+version = "4.9.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -271,7 +271,7 @@ hayro-jpeg2000 = { version = "0.3", default-features = false, features = [
     "simd",
 ], optional = true }
 hex = "0.4.3"
-html-to-markdown-rs = { version = "3.2.5", default-features = false, features = [
+html-to-markdown-rs = { version = "3.2.6", default-features = false, features = [
     "inline-images",
     "metadata",
 ], optional = true }
@@ -287,7 +287,7 @@ image = { version = "0.25.10", default-features = false, features = [
 ], optional = true }
 indexmap = "2.14.0"
 infer = "0.19.0"
-jotdown = "0.9"
+jotdown = "0.10"
 kamadak-exif = { version = "0.6.1", optional = true }
 
 kreuzberg-tesseract = { path = "../kreuzberg-tesseract", optional = true }
@@ -314,7 +314,7 @@ ort = { version = "2.0.0-rc.12", default-features = false, features = [
 outlook-pst = { version = "1.2.0", optional = true }
 parking_lot = "0.12.5"
 pastey = "0.2"
-pdf_oxide = { version = "0.3.34", default-features = false, optional = true }
+pdf_oxide = { version = "0.3.37", default-features = false, optional = true }
 pdfium-render = { package = "kreuzberg-pdfium-render", path = "../kreuzberg-pdfium-render", features = ["thread_safe", "image_latest"], optional = true }
 pulldown-cmark = { version = "0.13" }
 quick-xml = { version = "0.39.2", features = ["serialize"], optional = true }

data/vendor/kreuzberg/README.md

@@ -18,7 +18,7 @@ High-performance document intelligence library for Rust. Extract text, metadata,
 
 This is the core Rust library that powers the Python, TypeScript, and Ruby bindings.
 
-> **🚀 Version 4.9.1 Release**
+> **🚀 Version 4.9.4 Release**
 > This is a pre-release version. We invite you to test the library and [report any issues](https://github.com/kreuzberg-dev/kreuzberg/issues) you encounter.
 >
 > **Note**: The Rust crate is not currently published to crates.io for this RC. Use git dependencies or language bindings (Python, TypeScript, Ruby) instead.

data/vendor/kreuzberg/src/chunking/semantic/mod.rs

@@ -25,10 +25,6 @@ const SEGMENT_SIZE: usize = 200;
 #[cfg(feature = "embeddings")]
 const DEFAULT_TOPIC_THRESHOLD: f32 = 0.75;
 
-/// Safety ceiling for auto-budget when no embedding model is configured.
-/// Prevents unbounded chunks in header-less documents.
-const AUTO_BUDGET_CEILING: usize = 4000;
-
 /// Split text into semantically coherent chunks.
 ///
 /// Splits text into fine-grained segments, detects structural (and optionally
@@ -46,6 +42,8 @@ pub fn chunk_semantic(
         });
     }
 
+    warn_if_fallback_path(config);
+
     let seg_size = SEGMENT_SIZE;
     let has_markdown_headers = text.lines().any(crate::utils::markdown_utils::is_markdown_header);
     let splitter_segments: Vec<&str> = if has_markdown_headers {
@@ -165,11 +163,33 @@ fn compute_boundaries(_segments: &[Segment<'_>], forced: &[bool], _config: &Chun
     Ok(forced.to_vec())
 }
 
-/// Resolve the safety ceiling for chunk size.
+/// Warn when the semantic chunker is invoked without an embedding model.
+///
+/// Without an embedding, `chunk_semantic` falls back to a structural-boundary
+/// heuristic (ALL-CAPS headers, numbered sections, blank-line paragraphs).
+/// Topic-similarity chunking requires an embedding model. This warning makes
+/// the fallback mode discoverable to callers who think they're getting
+/// embedding-driven topic detection.
+#[cfg(feature = "embeddings")]
+fn warn_if_fallback_path(config: &ChunkingConfig) {
+    if config.embedding.is_none() {
+        tracing::warn!(
+            "chunker_type='semantic' without an EmbeddingConfig falls back to a \
+             structural-boundary heuristic; topic-similarity chunking requires an \
+             embedding model. Either configure `embedding` or switch to \
+             chunker_type='text'/'markdown' to silence this warning."
+        );
+    }
+}
+
+#[cfg(not(feature = "embeddings"))]
+fn warn_if_fallback_path(_config: &ChunkingConfig) {}
+
+/// Resolve the size ceiling for merged chunks.
 ///
-/// When an embedding preset is configured, use its chunk_size as the ceiling
-/// (chunks must fit in the model's context window). Otherwise use a generous
-/// default that prevents unbounded chunks in header-less documents.
+/// When an embedding preset is configured, use its `chunk_size` so chunks fit
+/// in the model's context window. Otherwise honor the caller's configured
+/// `max_characters`.
 fn resolve_ceiling(config: &ChunkingConfig) -> usize {
     #[cfg(feature = "embeddings")]
     if let Some(ref emb) = config.embedding
@@ -178,8 +198,7 @@ fn resolve_ceiling(config: &ChunkingConfig) -> usize {
     {
         return size;
     }
-    let _ = config;
-    AUTO_BUDGET_CEILING
+    config.max_characters
 }
 
 #[cfg(test)]
@@ -306,30 +325,124 @@ mod tests {
     }
 
     #[test]
-    fn ceiling_caps_oversized_headerless_text() {
-        // A large block of text with no headers should be split at the ceiling,
-        // not produce one unbounded chunk.
-        let text = "word ".repeat(1500); // ~7500 chars, exceeds AUTO_BUDGET_CEILING
+    fn max_characters_caps_oversized_headerless_text() {
+        // A large block of text with no headers must be split so every chunk
+        // respects the caller's configured max_characters.
+        let text = "word ".repeat(1500); // ~7500 chars
+        let max = 1000;
         let config = ChunkingConfig {
-            max_characters: 1000, // ignored by semantic chunker
+            max_characters: max,
             overlap: 0,
             trim: true,
             chunker_type: ChunkerType::Semantic,
             ..Default::default()
         };
         let result = chunk_semantic(&text, &config, None).unwrap();
-        assert!(result.chunks.len() >= 2, "should split at ceiling, got 1 chunk");
+        assert!(result.chunks.len() >= 2, "should split at max_characters, got 1 chunk");
         for (i, chunk) in result.chunks.iter().enumerate() {
             assert!(
-                chunk.content.chars().count() <= super::AUTO_BUDGET_CEILING + 100,
-                "chunk {} exceeds ceiling: {} > {}",
+                chunk.content.chars().count() <= max,
+                "chunk {} exceeds max_characters: {} > {}",
                 i,
                 chunk.content.chars().count(),
-                super::AUTO_BUDGET_CEILING
+                max
             );
         }
     }
 
+    #[test]
+    fn max_characters_controls_fallback_chunk_size() {
+        // bb-yq35 repro: with no embedding configured, different max_characters
+        // values must produce different chunking output.
+        let sample = format!(
+            "{}{}{}",
+            "Solar panel efficiency improves. ".repeat(200),
+            "\n\nFDA clinical trials require double-blind. ".repeat(200),
+            "\n\nQuantum entanglement needs cooling. ".repeat(200),
+        );
+
+        let run = |max: usize| {
+            let config = ChunkingConfig {
+                max_characters: max,
+                overlap: 0,
+                trim: true,
+                chunker_type: ChunkerType::Semantic,
+                ..Default::default()
+            };
+            chunk_semantic(&sample, &config, None).unwrap()
+        };
+
+        let small = run(500);
+        let large = run(1500);
+
+        assert!(
+            small.chunks.len() > large.chunks.len(),
+            "smaller max_characters must yield more chunks: small={}, large={}",
+            small.chunks.len(),
+            large.chunks.len()
+        );
+        for chunk in &small.chunks {
+            assert!(
+                chunk.content.chars().count() <= 500,
+                "small chunk exceeds cap: {}",
+                chunk.content.chars().count()
+            );
+        }
+        for chunk in &large.chunks {
+            assert!(
+                chunk.content.chars().count() <= 1500,
+                "large chunk exceeds cap: {}",
+                chunk.content.chars().count()
+            );
+        }
+    }
+
+    #[cfg(feature = "embeddings")]
+    #[test]
+    fn semantic_without_embedding_warns() {
+        use std::io::Write;
+        use std::sync::{Arc, Mutex};
+
+        #[derive(Clone, Default)]
+        struct Buf(Arc<Mutex<Vec<u8>>>);
+        impl Write for Buf {
+            fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
+                self.0.lock().unwrap().extend_from_slice(buf);
+                Ok(buf.len())
+            }
+            fn flush(&mut self) -> std::io::Result<()> {
+                Ok(())
+            }
+        }
+        impl<'a> tracing_subscriber::fmt::MakeWriter<'a> for Buf {
+            type Writer = Buf;
+            fn make_writer(&'a self) -> Self::Writer {
+                self.clone()
+            }
+        }
+
+        let buffer = Buf::default();
+        let subscriber = tracing_subscriber::fmt()
+            .with_writer(buffer.clone())
+            .with_max_level(tracing::Level::WARN)
+            .with_ansi(false)
+            .finish();
+
+        tracing::subscriber::with_default(subscriber, || {
+            let config = ChunkingConfig {
+                chunker_type: ChunkerType::Semantic,
+                ..Default::default()
+            };
+            let _ = chunk_semantic("hello world", &config, None).unwrap();
+        });
+
+        let captured = String::from_utf8(buffer.0.lock().unwrap().clone()).unwrap();
+        assert!(
+            captured.contains("without an EmbeddingConfig"),
+            "expected fallback warning in captured logs, got: {captured:?}"
+        );
+    }
+
     #[test]
     fn sections_with_headers_produce_separate_chunks() {
         // Each section has enough content that the segments span multiple paragraphs.

data/vendor/kreuzberg/src/core/config/extraction/types.rs

@@ -40,6 +40,18 @@ pub struct ImageExtractionConfig {
     /// Maximum DPI threshold
     #[serde(default = "default_max_dpi")]
     pub max_dpi: i32,
+
+    /// Maximum number of image objects to extract per PDF page.
+    ///
+    /// Some PDFs (e.g. technical diagrams stored as thousands of raster fragments)
+    /// can trigger extremely long or indefinite extraction times when every image
+    /// object on a dense page is decoded individually via pdfium FFI. Setting this
+    /// limit causes kreuzberg to stop collecting individual images once the count
+    /// per page reaches the cap and emit a warning instead.
+    ///
+    /// `None` (default) means no limit — all images are extracted.
+    #[serde(default)]
+    pub max_images_per_page: Option<u32>,
 }
 
 /// Token reduction configuration.
@@ -98,3 +110,44 @@ fn default_reduction_mode() -> String {
 fn default_confidence() -> f64 {
     0.8
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_max_images_per_page_defaults_none() {
+        let config = ImageExtractionConfig::default();
+        assert_eq!(config.max_images_per_page, None);
+    }
+
+    #[test]
+    fn test_max_images_per_page_serializes_as_null_when_none() {
+        let config = ImageExtractionConfig::default();
+        let json = serde_json::to_string(&config).unwrap();
+        assert!(json.contains("\"max_images_per_page\":null"));
+    }
+
+    #[test]
+    fn test_max_images_per_page_roundtrips_via_json() {
+        let config = ImageExtractionConfig {
+            max_images_per_page: Some(50),
+            ..Default::default()
+        };
+        let json = serde_json::to_string(&config).unwrap();
+        let back: ImageExtractionConfig = serde_json::from_str(&json).unwrap();
+        assert_eq!(back.max_images_per_page, Some(50));
+    }
+
+    /// Regression test for issue #766: missing field in JSON must not break
+    /// deserialization (backwards-compat — existing configs without this key
+    /// must still deserialize cleanly).
+    #[test]
+    fn test_max_images_per_page_absent_in_json_deserializes_as_none() {
+        let json = r#"{"extract_images":true,"target_dpi":300,"max_image_dimension":4096,
+                       "inject_placeholders":true,"auto_adjust_dpi":true,
+                       "min_dpi":72,"max_dpi":600}"#;
+        let config: ImageExtractionConfig = serde_json::from_str(json).unwrap();
+        assert_eq!(config.max_images_per_page, None);
+    }
+}

data/vendor/kreuzberg/src/core/config/ocr.rs

@@ -323,9 +323,12 @@ impl OcrConfig {
     /// Returns the effective pipeline config.
     ///
     /// - If `pipeline` is explicitly set, returns it.
-    /// - If `paddle-ocr` feature is compiled in and no explicit pipeline is set,
-    ///   auto-constructs a default pipeline: primary backend (priority 100) + paddleocr (priority 50).
-    /// - Otherwise returns `None` (single-backend mode, same as today).
+    /// - If `paddle-ocr` is compiled in and the backend is the default
+    ///   (tesseract), auto-constructs `[tesseract @ 100, paddleocr @ 50]`.
+    /// - Otherwise returns `None` (single-backend mode).
+    ///
+    /// Explicit non-default backend selections are honored as-is — a silent
+    /// paddleocr fallback would mask errors from the chosen backend.
     pub fn effective_pipeline(&self) -> Option<OcrPipelineConfig> {
         if self.pipeline.is_some() {
             return self.pipeline.clone();
@@ -333,25 +336,28 @@ impl OcrConfig {
 
         #[cfg(feature = "paddle-ocr")]
         {
-            let mut stages = vec![OcrPipelineStage {
-                backend: self.backend.clone(),
-                priority: 100,
-                language: None,
-                tesseract_config: self.tesseract_config.clone(),
-                paddle_ocr_config: None,
-                vlm_config: self.vlm_config.clone(),
-            }];
-            // Only add paddleocr fallback if primary backend isn't already paddleocr
-            if self.backend != "paddleocr" {
-                stages.push(OcrPipelineStage {
+            if self.backend != default_tesseract_backend() {
+                return None;
+            }
+
+            let stages = vec![
+                OcrPipelineStage {
+                    backend: self.backend.clone(),
+                    priority: 100,
+                    language: None,
+                    tesseract_config: self.tesseract_config.clone(),
+                    paddle_ocr_config: None,
+                    vlm_config: self.vlm_config.clone(),
+                },
+                OcrPipelineStage {
                     backend: "paddleocr".to_string(),
                     priority: 50,
                     language: None,
                     tesseract_config: None,
                     paddle_ocr_config: self.paddle_ocr_config.clone(),
                     vlm_config: None,
-                });
-            }
+                },
+            ];
             Some(OcrPipelineConfig {
                 stages,
                 quality_thresholds: self.effective_thresholds(),
@@ -485,29 +491,21 @@ mod tests {
     }
 
     #[test]
-    fn test_effective_pipeline_paddleocr_backend_no_duplicate() {
-        // When primary backend is "paddleocr", effective_pipeline should NOT add
-        // a second paddleocr stage (issue #6 fix).
+    fn test_effective_pipeline_explicit_paddleocr_no_autofallback() {
         let config = OcrConfig {
             backend: "paddleocr".to_string(),
             ..Default::default()
         };
-        let result = config.effective_pipeline();
-        // With paddle-ocr feature: should have exactly 1 stage (no duplicate)
-        // Without paddle-ocr feature: should be None
-        #[cfg(feature = "paddle-ocr")]
-        {
-            let pipeline = result.unwrap();
-            let paddle_count = pipeline.stages.iter().filter(|s| s.backend == "paddleocr").count();
-            assert_eq!(
-                paddle_count, 1,
-                "Should not have duplicate paddleocr stages, found {paddle_count}"
-            );
-        }
-        #[cfg(not(feature = "paddle-ocr"))]
-        {
-            assert!(result.is_none());
-        }
+        assert!(config.effective_pipeline().is_none());
+    }
+
+    #[test]
+    fn test_effective_pipeline_explicit_easyocr_no_autofallback() {
+        let config = OcrConfig {
+            backend: "easyocr".to_string(),
+            ..Default::default()
+        };
+        assert!(config.effective_pipeline().is_none());
     }
 
     #[test]

data/vendor/kreuzberg/src/core/config/processing.rs

@@ -14,11 +14,13 @@ use std::path::PathBuf;
 /// * `Text` - Generic text splitter, splits on whitespace and punctuation
 /// * `Markdown` - Markdown-aware splitter, preserves formatting and structure
 /// * `Yaml` - YAML-aware splitter, creates one chunk per top-level key
-/// * `Semantic` - Topic-aware chunker that splits at natural document boundaries
-///   (headers, paragraph breaks, topic shifts). Works out of the box with no extra
-///   configuration. Optionally add an `EmbeddingConfig` for embedding-based topic
-///   detection; `topic_threshold` (default 0.75) and `max_characters` (default 1000)
-///   are automatically applied when not specified.
+/// * `Semantic` - Topic-aware chunker. With an `EmbeddingConfig`, splits at
+///   embedding-based topic shifts tuned by `topic_threshold` (default 0.75,
+///   lower = more splits). Without an embedding, falls back to a
+///   structural-boundary heuristic (ALL-CAPS headers, numbered sections,
+///   blank-line paragraphs) and merges groups into chunks capped at
+///   `max_characters` (default 1000). `topic_threshold` has no effect in the
+///   fallback path. For best results, pair with an embedding model.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
 #[serde(rename_all = "lowercase")]
 pub enum ChunkerType {