kreuzberg 4.10.2 → 4.10.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 6f9e454c85db9b78c82dc30871278c9faa132008ce57fb0b50d09712165a8e79
-  data.tar.gz: e1c38eccc06b401149d0a5f6863867da5b9abbc9f95895b0e2e278509f30f047
+  metadata.gz: dfacdc270872e87de7d3be21475808d50ec8f5ca34655eef45c740007de1b222
+  data.tar.gz: bb632e33fc5dc77dbe3740eac10075541974c95999bd6af65dff138c0acb5dbf
 SHA512:
-  metadata.gz: bad7bfa9ff4200bb49704677ac64aa6c4144df3e62bf402e664c1f2dafe2c88a6ab4a88597fabbb530d2c7c5c62923a7678f7babb592bf3432269645d0a9b9a0
-  data.tar.gz: d63b9df95db3f5d632b454611c278e6eda637598a149ca1648b92709d96a3a869c9db9f2739bcc65acb31526d33d0db628a0abc05c5e0759272c88d828e0be0b
+  metadata.gz: 34206c5ddf11c69313124109d5cd81ae67f6ed8b3250636359f556a94fd79cc3fd37def7754e4dd46656aa7e286c0cb1bc88c0c2f2a36d841eb941da23cfb8b4
+  data.tar.gz: 2a1ff1c6d8a20045c329feae6d64e0a98f23dd11e1bf6887f1ef788031a2b191c08533c21e4ff149a7474b2257d6325dad28fe04e2fef594ee69dd9bd2e078d8

data/ext/kreuzberg_rb/native/.cargo/config.toml

@@ -3,18 +3,18 @@ rustflags = ["-A", "unpredictable-function-pointer-comparisons", "-A", "fn_ptr_e
 
 [target.x86_64-pc-windows-gnu]
 rustflags = [
-    "-A",
-    "unpredictable-function-pointer-comparisons",
-    "-A",
-    "fn_ptr_eq",
-    "-C",
-    "link-arg=-lbcrypt",
-    "-C",
-    "link-arg=-lcrypt32",
-    "-C",
-    "link-arg=-lws2_32",
-    "-C",
-    "link-arg=-luserenv",
+  "-A",
+  "unpredictable-function-pointer-comparisons",
+  "-A",
+  "fn_ptr_eq",
+  "-C",
+  "link-arg=-lbcrypt",
+  "-C",
+  "link-arg=-lcrypt32",
+  "-C",
+  "link-arg=-lws2_32",
+  "-C",
+  "link-arg=-luserenv",
 ]
 
 [target.x86_64-pc-windows-msvc]

data/ext/kreuzberg_rb/native/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "kreuzberg-rb"
-version = "4.10.2"
+version = "4.10.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -22,7 +22,7 @@ default = ["embeddings"]
 embeddings = ["kreuzberg/embeddings"]
 
 [dependencies]
-async-trait = "0.1.89"
+async-trait = "0.1.92"
 kreuzberg = { path = "../../../vendor/kreuzberg", default-features = false, features = [
     "pdf",
     "excel",
@@ -49,13 +49,13 @@ kreuzberg-ffi = { path = "../../../vendor/kreuzberg-ffi" }
 magnus = { git = "https://github.com/matsadler/magnus", rev = "f6db11769efb517427bf7f121f9c32e18b059b38", features = [
     "rb-sys",
 ] }
-rb-sys = { version = "0.9.128", default-features = false, features = [
+rb-sys = { version = "0.9.130", default-features = false, features = [
     "stable-api-compiled-fallback",
 ] }
-serde_json = "1.0.150"
-toml = "1.1.2"
+serde_json = "1.0.151"
+toml = "1.1.5"
 serde_yaml_ng = "0.10"
-tokio = { version = "1.52.3", features = [
+tokio = { version = "1.53.1", features = [
     "rt",
     "rt-multi-thread",
     "macros",
@@ -65,7 +65,7 @@ tokio = { version = "1.52.3", features = [
     "time",
     "io-util",
 ] }
-html-to-markdown-rs = { version = "3.8.3", default-features = false }
+html-to-markdown-rs = { version = "3.14.3", default-features = false }
 
 [dev-dependencies]
 pretty_assertions = "1.4"

data/lib/kreuzberg/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Kreuzberg
-  VERSION = "4.10.2"
+  VERSION = "4.10.4"
 end

data/vendor/Cargo.toml

@@ -2,7 +2,7 @@
 members = ["kreuzberg", "kreuzberg-ffi", "kreuzberg-tesseract", "kreuzberg-paddle-ocr", "kreuzberg-pdfium-render"]
 
 [workspace.package]
-version = "4.10.2"
+version = "4.10.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -12,15 +12,15 @@ homepage = "https://kreuzberg.dev"
 
 [workspace.dependencies]
 ahash = { version = "0.8.12", features = ["serde"] }
-anyhow = "1.0.103"
-async-trait = "0.1.89"
-base64 = "0.22.1"
+anyhow = "1.0.104"
+async-trait = "0.1.92"
+base64 = "0.23.1"
 blake3 = "1"
 bytes = { version = "1", features = ["serde"] }
-cfb = "0.14"
+cfb = "0.15"
 chrono = "0.4"
 clap = { version = "4.6", features = ["derive", "color", "suggestions"] }
-comrak = { version = "0.54", default-features = false }
+comrak = { version = "0.55", default-features = false }
 console_error_panic_hook = "0.1"
 criterion = { version = "0.8", features = ["html_reports"] }
 ctor = "1.0"
@@ -28,35 +28,35 @@ dbase = "0.8"
 futures = "0.3"
 getrandom = { version = "0.4.3", features = ["wasm_js"] }
 hex = "0.4.3"
-html-to-markdown-rs = { version = "3.8.3", default-features = false }
+html-to-markdown-rs = { version = "3.14.3", default-features = false }
 image = { version = "0.25.10", default-features = false }
 itertools = "0.15"
 js-sys = "0.3"
-kreuzberg = { path = "./crates/kreuzberg", version = "4.10.2", default-features = false }
-kreuzberg-ffi = { path = "./crates/kreuzberg-ffi", version = "4.10.2" }
+kreuzberg = { path = "./crates/kreuzberg", version = "4.10.4", default-features = false }
+kreuzberg-ffi = { path = "./crates/kreuzberg-ffi", version = "4.10.4" }
 lazy_static = "1.5.0"
-libc = "0.2.186"
-liter-llm = { version = "1.9", features = ["native-http", "tracing"], default-features = false }
+libc = "0.2.189"
+liter-llm = { version = "2.0.3", features = ["native-http"], default-features = false }
 log = "0.4"
-lzma-rust2 = { version = "0.16.5" }
+lzma-rust2 = { version = "0.21" }
 memmap2 = "0.9.11"
 minijinja = "2"
 num_cpus = "1.17.0"
 once_cell = "1.21.4"
-ort = { version = "2.0.0-rc.12", features = ["std", "api-18"], default-features = false }
+ort = { version = "=2.0.0-rc.12", features = ["std", "api-18"], default-features = false }
 parking_lot = "0.12.5"
 pdfium-render = { package = "kreuzberg-pdfium-render", path = "crates/kreuzberg-pdfium-render", version = "4.3" }
 rayon = "1.12.0"
-reqwest = { version = "0.13.4", default-features = false }
-serde = { version = "1.0.228", features = ["derive"] }
-serde_json = { version = "1.0.150" }
+reqwest = { version = "0.13.5", default-features = false }
+serde = { version = "1.0.229", features = ["derive"] }
+serde_json = { version = "1.0.151" }
 serde_toon_format = "0.1"
 tempfile = "3.27.0"
-thiserror = "2.0.18"
-tokio = { version = "1.52.3", features = ["rt", "rt-multi-thread", "macros", "sync", "process", "fs", "time", "io-util"] }
-toml = "1.1.2"
+thiserror = "2.0.20"
+tokio = { version = "1.53.1", features = ["rt", "rt-multi-thread", "macros", "sync", "process", "fs", "time", "io-util"] }
+toml = "1.1.6"
 tracing = "0.1"
-tree-sitter-language-pack = { version = "1.12.5", features = ["serde"], default-features = false }
+tree-sitter-language-pack = { version = "1.20.0", features = ["serde"], default-features = false }
 wasm-bindgen = { version = "0.2", features = ["enable-interning"] }
 wasm-bindgen-futures = "0.4"
 web-sys = { version = "0.3", features = ["Blob", "File", "FileReader", "console", "TextDecoder", "ImageData", "Window", "Response"] }

data/vendor/kreuzberg/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "kreuzberg"
-version = "4.10.2"
+version = "4.10.4"
 edition = "2024"
 rust-version = "1.91"
 authors = ["Na'aman Hirschfeld <naaman@kreuzberg.dev>"]
@@ -191,29 +191,29 @@ cli = [
 
 [dependencies]
 ahash = { version = "0.8.12", features = ["serde"] }
-async-trait = "0.1.89"
+async-trait = "0.1.92"
 axum = { version = "0.8", features = ["macros", "json", "multipart"], optional = true }
-base64 = "0.22.1"
+base64 = "0.23.1"
 biblatex = { version = "0.12", optional = true }
-biblib = { version = "0.7", default-features = false, features = ["ris", "pubmed", "xml"], optional = true }
+biblib = { version = "0.8", default-features = false, features = ["ris", "pubmed", "xml"], optional = true }
 bitvec = "1.1"
 blake3 = "1"
 bytes = { version = "1", features = ["serde"] }
-calamine = { version = "0.36.0", features = ["dates"], optional = true }
-cfb = { version = "0.14", optional = true }
+calamine = { version = "0.36.1", features = ["dates"], optional = true }
+cfb = { version = "0.15", optional = true }
 chardetng = { version = "1.0.0", optional = true }
 chrono = { version = "0.4", optional = true }
-comrak = { version = "0.54", default-features = false }
+comrak = { version = "0.55", default-features = false }
 dashmap = "6.2"
 dbase = { version = "0.8", optional = true }
-dirs = "6"
-encoding_rs = { version = "0.8.35" }
-fast_image_resize = { version = "6.0.0", optional = true }
+dirs = "7"
+encoding_rs = { version = "0.8.41" }
+fast_image_resize = { version = "6.1.0", optional = true }
 flate2 = { version = "1.1", optional = true }
 hayro-jbig2 = { version = "0.3", default-features = false, features = ["std"], optional = true }
 hayro-jpeg2000 = { version = "0.3", default-features = false, features = ["std", "simd"], optional = true }
 hex = "0.4.3"
-html-to-markdown-rs = { version = "3.8.3", default-features = false, features = ["inline-images", "metadata"], optional = true }
+html-to-markdown-rs = { version = "3.14.3", default-features = false, features = ["inline-images", "metadata"], optional = true }
 image = { version = "0.25.10", default-features = false, features = [
     "png",
     "jpeg",
@@ -224,17 +224,17 @@ image = { version = "0.25.10", default-features = false, features = [
     "pnm",
     "rayon",
 ], optional = true }
-indexmap = "2.14.0"
-infer = "0.19.0"
+indexmap = "2.14.2"
+infer = "0.22.0"
 jotdown = "0.10"
 kamadak-exif = { version = "0.6.1", optional = true }
 
 kreuzberg-tesseract = { path = "../kreuzberg-tesseract", optional = true }
-libc = "0.2.186"
-liter-llm = { version = "1.9", features = ["native-http", "tracing"], default-features = false, optional = true }
+libc = "0.2.189"
+liter-llm = { version = "2.0.3", features = ["native-http"], default-features = false, optional = true }
 log = "0.4"
-lopdf = { version = "0.44.0", default-features = false, features = ["chrono", "rayon"], optional = true }
-mail-parser = { version = "0.11.5", optional = true }
+lopdf = { version = "0.45.0", default-features = false, features = ["chrono", "rayon"], optional = true }
+mail-parser = { version = "0.11.9", optional = true }
 memchr = "2.8.3"
 memmap2 = "0.9.11"
 mime_guess = "2.0"
@@ -245,17 +245,17 @@ once_cell = "1.21.4"
 opentelemetry = { version = "0.32", features = ["trace"], optional = true }
 opentelemetry_sdk = { version = "0.32", features = ["rt-tokio"], optional = true }
 org = { version = "0.3", optional = true }
-ort = { version = "2.0.0-rc.12", default-features = false, features = ["std", "ndarray", "api-18"], optional = true }
+ort = { version = "=2.0.0-rc.12", default-features = false, features = ["std", "ndarray", "api-18"], optional = true }
 outlook-pst = { version = "1.2.0", optional = true }
 parking_lot = "0.12.5"
 pastey = "0.2"
 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 = "0.41.0", features = ["serialize"], optional = true }
+quick-xml = { version = "=0.41.0", features = ["serialize"], optional = true }
 rake = { version = "0.3.6", optional = true }
 rayon = "1.12.0"
-regex = "1.13.0"
-rmcp = { version = "2.2.0", features = [
+regex = "1.13.1"
+rmcp = { version = "3.4.0", features = [
     "server",
     "macros",
     "base64",
@@ -266,8 +266,8 @@ rmcp = { version = "2.2.0", features = [
 rmp-serde = "1.3"
 
 roxmltree = { version = "0.21.1", optional = true }
-serde = { version = "1.0.228", features = ["derive"] }
-serde_json = { version = "1.0.150" }
+serde = { version = "1.0.229", features = ["derive"] }
+serde_json = { version = "1.0.151" }
 serde_toon_format = "0.1"
 serde_yaml_ng = "0.10.0"
 sevenz-rust2 = { version = "0.20.2", optional = true }
@@ -277,13 +277,13 @@ snap = { version = "1.1", optional = true }
 tar = { version = "0.4.46", optional = true }
 tempfile = { version = "3.27.0", optional = true }
 text-splitter = { version = "0.32.0", features = ["markdown"], optional = true }
-thiserror = "2.0.18"
+thiserror = "2.0.20"
 tiff = { version = "0.11", optional = true }
 # Keep aligned with text-splitter's optional tokenizers integration so ChunkSizer
 # is implemented for the same Tokenizer type used by Kreuzberg.
-tokenizers = { version = "0.23.1", optional = true, default-features = false, features = ["http", "fancy-regex"] }
-tokio = { version = "1.52.3", features = ["rt", "rt-multi-thread", "macros", "sync", "process", "fs", "time", "io-util"], optional = true }
-toml = "1.1.2"
+tokenizers = { version = "0.23.2", optional = true, default-features = false, features = ["http", "fancy-regex"] }
+tokio = { version = "1.53.1", features = ["rt", "rt-multi-thread", "macros", "sync", "process", "fs", "time", "io-util"], optional = true }
+toml = "1.1.6"
 tower = { version = "0.5", features = ["timeout", "limit", "util"], optional = true }
 tower-http = { version = "0.7", features = [
     "cors",
@@ -310,13 +310,13 @@ hf-hub = { version = "1.0", default-features = false, features = ["blocking", "r
 # PaddleOCR via ONNX Runtime - not available on WASM (vendored from paddle-ocr-rs)
 kreuzberg-paddle-ocr = { path = "../kreuzberg-paddle-ocr", optional = true }
 pprof = { version = "0.15.0", features = ["flamegraph"], optional = true }
-ureq = { version = "3.3", default-features = false, features = ["rustls", "json"], optional = true }
+ureq = { version = "3.4", default-features = false, features = ["rustls", "json"], optional = true }
 
 [target.'cfg(all(target_os = "windows", not(target_arch = "wasm32")))'.dependencies]
 hf-hub = { version = "1.0", default-features = false, features = ["blocking", "rustls-tls"], optional = true }
 # PaddleOCR via ONNX Runtime - not available on WASM (vendored from paddle-ocr-rs)
 kreuzberg-paddle-ocr = { path = "../kreuzberg-paddle-ocr", optional = true }
-ureq = { version = "3.3", default-features = false, features = ["native-tls", "json"], optional = true }
+ureq = { version = "3.4", default-features = false, features = ["native-tls", "json"], optional = true }
 
 [target.'cfg(not(target_arch = "wasm32"))'.dependencies.tree-sitter-language-pack]
 workspace = true
@@ -327,7 +327,7 @@ optional = true
 # Override getrandom to enable js feature for WASM targets
 # This is needed because ring/rustls (via ureq) depend on getrandom without js feature
 getrandom = { version = "0.4.3", features = ["wasm_js"] }
-tree-sitter-language-pack = { version = "1.12.5", features = ["serde"], default-features = false, optional = true }
+tree-sitter-language-pack = { version = "1.20.0", features = ["serde"], default-features = false, optional = true }
 wasm-bindgen-rayon = { version = "1.3", optional = true }
 
 [build-dependencies]
@@ -335,16 +335,16 @@ pkg-config = "0.3"
 tracing = "0.1"
 
 [dev-dependencies]
-anyhow = "1.0.103"
+anyhow = "1.0.104"
 criterion = { version = "0.8", features = ["html_reports"] }
 dotenvy = "0.15"
 filetime = "0.2"
 image = { version = "0.25.10", default-features = false, features = ["png"] }
-jsonschema = "0.47"
+jsonschema = "0.56"
 serial_test = "3.5.0"
 tar = "0.4.46"
 tempfile = "3.27.0"
-tokio = { version = "1.52.3", features = ["macros", "time"] }
+tokio = { version = "1.53.1", features = ["macros", "time"] }
 tokio-test = "0.4"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 zip = { version = ">=7.0.0, <8.6.0", default-features = false, features = ["deflate-flate2"] }

data/vendor/kreuzberg/README.md

@@ -60,7 +60,7 @@ 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 4.10.2 Release**
+> **🚀 Version 4.10.4 Release**
 > This is a pre-release version. We invite you to test the library and [report any issues](https://github.com/kreuzberg-dev/kreuzberg-lts/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/core/config/extraction/loaders.rs

@@ -175,16 +175,36 @@ impl ExtractionConfig {
         Ok((*config_arc).clone())
     }
 
-    /// Discover configuration file in parent directories.
+    /// Discover a configuration file.
     ///
-    /// Searches for `kreuzberg.toml` in current directory and parent directories.
+    /// Resolution order:
+    /// 1. `kreuzberg.toml` in the current directory, then each parent up to the
+    ///    filesystem root (project-local config wins).
+    /// 2. The user-level config directory as a fallback, resolved
+    ///    platform-natively via [`dirs::config_dir`] (same approach as the cache
+    ///    dir): `$XDG_CONFIG_HOME/kreuzberg/kreuzberg.toml` or
+    ///    `~/.config/kreuzberg/kreuzberg.toml` on Linux,
+    ///    `~/Library/Application Support/kreuzberg/kreuzberg.toml` on macOS,
+    ///    `%APPDATA%\kreuzberg\kreuzberg.toml` on Windows.
     ///
     /// # Returns
     ///
     /// - `Some(config)` if found
     /// - `None` if no config file found
     pub fn discover() -> Result<Option<Self>> {
-        let mut current = std::env::current_dir().map_err(KreuzbergError::Io)?;
+        let cwd = std::env::current_dir().map_err(KreuzbergError::Io)?;
+        let global = dirs::config_dir().map(|dir| dir.join("kreuzberg").join("kreuzberg.toml"));
+        Self::discover_from(&cwd, global.as_deref())
+    }
+
+    /// Walk up from `start` looking for `kreuzberg.toml`, then fall back to the
+    /// user-level `global` config path if provided.
+    ///
+    /// Split out from [`discover`](Self::discover) so the resolution order is
+    /// unit-testable without mutating process-global state (the working
+    /// directory or `HOME`).
+    fn discover_from(start: &Path, global: Option<&Path>) -> Result<Option<Self>> {
+        let mut current = start.to_path_buf();
 
         loop {
             let kreuzberg_toml = current.join("kreuzberg.toml");
@@ -199,6 +219,59 @@ impl ExtractionConfig {
             }
         }
 
+        if let Some(global) = global
+            && global.exists()
+        {
+            return Ok(Some(Self::from_toml_file(global)?));
+        }
+
         Ok(None)
     }
 }
+
+#[cfg(test)]
+mod discover_tests {
+    use super::*;
+    use std::fs;
+
+    const TOML: &str = "[ocr]\nenabled = true\n";
+
+    #[test]
+    fn project_config_takes_precedence_over_global() {
+        // Project config has an [ocr] section; the global fallback does not, so
+        // `ocr.is_some()` proves the project-local file was the one loaded.
+        let project = tempfile::tempdir().expect("tempdir");
+        fs::write(project.path().join("kreuzberg.toml"), TOML).expect("write project");
+        let global = tempfile::tempdir().expect("tempdir");
+        let global_file = global.path().join("kreuzberg.toml");
+        fs::write(&global_file, "output_format = \"markdown\"\n").expect("write global");
+
+        let found = ExtractionConfig::discover_from(project.path(), Some(&global_file))
+            .expect("discover")
+            .expect("some config");
+        assert!(
+            found.ocr.is_some(),
+            "project-local config must win over the global fallback"
+        );
+    }
+
+    #[test]
+    fn falls_back_to_global_when_no_project_config() {
+        let empty = tempfile::tempdir().expect("tempdir");
+        let global = tempfile::tempdir().expect("tempdir");
+        let global_file = global.path().join("kreuzberg.toml");
+        fs::write(&global_file, TOML).expect("write global");
+
+        let found = ExtractionConfig::discover_from(empty.path(), Some(&global_file)).expect("discover");
+        assert!(found.is_some(), "should fall back to the global config file");
+    }
+
+    #[test]
+    fn returns_none_when_no_config_anywhere() {
+        let empty = tempfile::tempdir().expect("tempdir");
+        let missing = empty.path().join("nope").join("kreuzberg.toml");
+
+        let found = ExtractionConfig::discover_from(empty.path(), Some(&missing)).expect("discover");
+        assert!(found.is_none(), "no project or global config should yield None");
+    }
+}

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

@@ -156,6 +156,7 @@ impl From<&TreeSitterProcessConfig> for tree_sitter_language_pack::ProcessConfig
             diagnostics: p.diagnostics,
             chunk_max_size: p.chunk_max_size,
             data_extraction: false,
+            ..Default::default()
         }
     }
 }

data/vendor/kreuzberg/src/extraction/doc/mod.rs

@@ -8,6 +8,12 @@
 use crate::error::{KreuzbergError, Result};
 use std::io::Cursor;
 
+/// Zero-based pair index of fcClx/lcbClx in FibRgFcLcb97.
+///
+/// See [MS-DOC] section 2.5.6:
+/// <https://learn.microsoft.com/en-us/openspecs/office_file_formats/ms-doc/0c9df81f-98d0-454e-ad84-b612cd05b1a4>
+const CLX_PAIR_INDEX: usize = 33;
+
 /// Result of DOC text extraction.
 pub struct DocExtractionResult {
     /// Extracted text content.
@@ -116,10 +122,14 @@ fn extract_text_word97(word_doc: &[u8], table_stream: &[u8]) -> Result<String> {
         return Err(KreuzbergError::parsing("FIB too short for cbRgFcLcb"));
     }
 
-    let _ = u16::from_le_bytes([word_doc[cbrgfclcb_offset], word_doc[cbrgfclcb_offset + 1]]) as usize;
+    let cb_rg_fc_lcb = u16::from_le_bytes([word_doc[cbrgfclcb_offset], word_doc[cbrgfclcb_offset + 1]]) as usize;
     let rg_fc_lcb_offset = cbrgfclcb_offset + 2;
 
-    let fc_clx_offset = rg_fc_lcb_offset + 66 * 8;
+    if cb_rg_fc_lcb <= CLX_PAIR_INDEX {
+        return extract_text_contiguous(word_doc, ccp_text);
+    }
+
+    let fc_clx_offset = rg_fc_lcb_offset + CLX_PAIR_INDEX * 8;
     let lcb_clx_offset = fc_clx_offset + 4;
 
     if word_doc.len() < lcb_clx_offset + 4 {
@@ -673,4 +683,33 @@ mod tests {
         let result = extract_doc_text(b"not a doc file");
         assert!(result.is_err());
     }
+
+    #[test]
+    fn test_extract_word97_reads_clx_from_fib_rg_fc_lcb97_pair_33() {
+        let mut word_doc = vec![0_u8; 900];
+        word_doc[32..34].copy_from_slice(&14_u16.to_le_bytes());
+        word_doc[62..64].copy_from_slice(&22_u16.to_le_bytes());
+        word_doc[76..80].copy_from_slice(&1_u32.to_le_bytes());
+
+        let rg_fc_lcb_offset = 154;
+        word_doc[152..154].copy_from_slice(&34_u16.to_le_bytes());
+        let clx_pair_offset = rg_fc_lcb_offset + 33 * 8;
+        word_doc[clx_pair_offset..clx_pair_offset + 4].copy_from_slice(&8_u32.to_le_bytes());
+        word_doc[clx_pair_offset + 4..clx_pair_offset + 8].copy_from_slice(&21_u32.to_le_bytes());
+
+        let text_offset = 800_usize;
+        word_doc[text_offset] = 0xE9;
+
+        let mut table_stream = vec![0_u8; 29];
+        table_stream[8] = 0x02;
+        table_stream[9..13].copy_from_slice(&16_u32.to_le_bytes());
+        table_stream[13..17].copy_from_slice(&0_u32.to_le_bytes());
+        table_stream[17..21].copy_from_slice(&1_u32.to_le_bytes());
+        let compressed_fc = 0x4000_0000_u32 | (text_offset as u32 * 2);
+        table_stream[23..27].copy_from_slice(&compressed_fc.to_le_bytes());
+
+        let text = extract_text_word97(&word_doc, &table_stream).expect("failed to read synthetic Word 97 text");
+
+        assert_eq!(text, "é");
+    }
 }

data/vendor/kreuzberg/src/extraction/docx/parser.rs

@@ -9,6 +9,11 @@
 //! - Removed file-path based APIs (we only need bytes/reader)
 //! - Added markdown rendering and formatting support (fixes #376)
 
+// TODO(kreuzberg-dev/kreuzberg-lts#4): 4 cyclomatic-complexity and 46 size/complexity findings
+// in this file, currently excluded via the quality-debt baseline in poly.toml. Splitting
+// these needs compiler-in-the-loop verification, not a mechanical pass. Delete this
+// note and the file's baseline entry together once it goes green. Help wanted.
+
 use ahash::AHashMap;
 use std::io::{Cursor, Read, Seek};
 

data/vendor/kreuzberg/src/extractors/csv.rs

@@ -7,7 +7,7 @@ use std::borrow::Cow;
 use std::sync::LazyLock;
 
 use crate::Result;
-use crate::core::config::ExtractionConfig;
+use crate::core::config::{ExtractionConfig, OutputFormat};
 use crate::plugins::{DocumentExtractor, Plugin};
 use crate::text::utf8_validation;
 use crate::types::Table;
@@ -71,7 +71,7 @@ impl DocumentExtractor for CsvExtractor {
         &self,
         content: &[u8],
         mime_type: &str,
-        _config: &ExtractionConfig,
+        config: &ExtractionConfig,
     ) -> Result<InternalDocument> {
         tracing::debug!(format = "csv", size_bytes = content.len(), "extraction starting");
         let text = decode_csv_bytes(content);
@@ -114,14 +114,21 @@ impl DocumentExtractor for CsvExtractor {
         };
 
         let mut builder = InternalDocumentBuilder::new("csv");
-        let cloned_table = Table {
-            cells: table.cells.clone(),
-            markdown: table.markdown.clone(),
-            page_number: table.page_number,
-            bounding_box: table.bounding_box,
-        };
-        builder.push_table(cloned_table, None, None);
+        if matches!(config.output_format, OutputFormat::Plain) {
+            let content_text = build_content_text(&table.cells, has_header);
+            for section in content_text.split("\n\n") {
+                let section = section.trim();
+                if !section.is_empty() {
+                    builder.push_paragraph(section, vec![], None, None);
+                }
+            }
+        } else {
+            builder.push_table(table.clone(), None, None);
+        }
         let mut doc = builder.build();
+        if matches!(config.output_format, OutputFormat::Plain) {
+            doc.tables.push(table);
+        }
         doc.mime_type = Cow::Owned(mime_type.to_string());
 
         doc.metadata = Metadata {
@@ -392,6 +399,51 @@ fn infer_column_types(rows: &[Vec<String>], has_header: bool) -> Vec<String> {
         .collect()
 }
 
+/// Build embedding-friendly labeled text for a header/data-row CSV.
+///
+/// Each data row becomes a `Row N:` block with `Header: value` pairs for its
+/// non-empty cells, separated by blank lines. Rows without a header fall back
+/// to space-joined cells so plain output stays readable.
+fn build_content_text(rows: &[Vec<String>], has_header: bool) -> String {
+    if rows.is_empty() {
+        return String::new();
+    }
+
+    if !has_header || rows.len() < 2 {
+        return rows
+            .iter()
+            .map(|row| {
+                row.iter()
+                    .map(|cell| cell.trim())
+                    .filter(|cell| !cell.is_empty())
+                    .collect::<Vec<_>>()
+                    .join(" ")
+            })
+            .filter(|line| !line.is_empty())
+            .collect::<Vec<_>>()
+            .join("\n");
+    }
+
+    let headers = &rows[0];
+    let mut sections = Vec::with_capacity(rows.len() - 1);
+
+    for (i, row) in rows[1..].iter().enumerate() {
+        let mut lines = vec![format!("Row {}:", i + 1)];
+        for (header, value) in headers.iter().zip(row.iter()) {
+            let h = header.trim();
+            let v = value.trim();
+            if !h.is_empty() && !v.is_empty() {
+                lines.push(format!("  {}: {}", h, v));
+            }
+        }
+        if lines.len() > 1 {
+            sections.push(lines.join("\n"));
+        }
+    }
+
+    sections.join("\n\n")
+}
+
 /// Build a Markdown table from parsed rows.
 fn build_markdown_table(rows: &[Vec<String>]) -> String {
     if rows.is_empty() {

data/vendor/kreuzberg/src/extractors/docbook.rs

@@ -15,6 +15,11 @@
 //! - Tables
 //! - Cross-references and links
 
+// TODO(kreuzberg-dev/kreuzberg-lts#4): 4 cyclomatic-complexity and 27 size/complexity findings
+// in this file, currently excluded via the quality-debt baseline in poly.toml. Splitting
+// these needs compiler-in-the-loop verification, not a mechanical pass. Delete this
+// note and the file's baseline entry together once it goes green. Help wanted.
+
 use crate::Result;
 use crate::core::config::ExtractionConfig;
 use crate::extraction::{cells_to_markdown, cells_to_text};