From ea83a2671eef7a1eb5fe1af2bdad5fd6a50fb8db Mon Sep 17 00:00:00 2001
From: Wes Chow <wesc@filament.dm>
Date: Tue, 26 May 2026 17:58:59 -0400
Subject: [PATCH 1/2] Fix handling of corrupt or empty documents.

---
 processors/src/docx_processor.rs | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

diff --git a/processors/src/docx_processor.rs b/processors/src/docx_processor.rs
index 350da0aa..8273ff27 100644
--- a/processors/src/docx_processor.rs
+++ b/processors/src/docx_processor.rs
@@ -18,7 +18,22 @@ impl DocxProcessor {
 
 impl FileProcessor for DocxProcessor {
     fn process_file(&self, path: impl AsRef<Path>) -> anyhow::Result<Document> {
-        let docs = MarkdownDocument::from_file(path);
+        // `docx-parser::MarkdownDocument::from_file` uses `panic!` instead of returning
+        // `Result` when the file is missing, corrupt, or not a valid DOCX/ZIP archive.
+        // We catch that panic here and convert it into a proper anyhow error so callers
+        // get a clean Err(…) rather than a process-level abort.
+        let path = path.as_ref().to_owned();
+        let docs =
+            std::panic::catch_unwind(move || MarkdownDocument::from_file(&path)).map_err(|e| {
+                let msg = if let Some(s) = e.downcast_ref::<String>() {
+                    s.clone()
+                } else if let Some(s) = e.downcast_ref::<&str>() {
+                    s.to_string()
+                } else {
+                    "unknown panic".to_string()
+                };
+                anyhow::anyhow!("docx_parser panicked while opening file: {}", msg)
+            })?;
         let markdown = docs.to_markdown(false);
         self.markdown_processor.process_document(&markdown)
     }

From a9188c7eeebeba74a2df6e9dbfd7246e4a205313 Mon Sep 17 00:00:00 2001
From: Wes Chow <wesc@filament.dm>
Date: Tue, 26 May 2026 18:00:41 -0400
Subject: [PATCH 2/2] Return errors when statistical functions are given bad
 input rather than panics.

---
 rust/src/chunkers/statistical.rs | 41 ++++++++++++++++++++++----------
 1 file changed, 28 insertions(+), 13 deletions(-)

diff --git a/rust/src/chunkers/statistical.rs b/rust/src/chunkers/statistical.rs
index b11c3e76..1aee24c9 100644
--- a/rust/src/chunkers/statistical.rs
+++ b/rust/src/chunkers/statistical.rs
@@ -11,27 +11,33 @@ use text_splitter::{ChunkConfig, TextSplitter};
 // use text_splitter::{ChunkConfig, TextSplitter};
 use tokenizers::Tokenizer;
 
-fn median<T>(data: &[T]) -> T
+fn median<T>(data: &[T]) -> Option<T>
 where
     T: Copy + PartialOrd + std::ops::Add<Output = T> + std::ops::Div<Output = T> + From<u8>,
 {
-    assert!(!data.is_empty(), "median requires at least one data point");
+    if data.is_empty() {
+        return None;
+    }
     let mut sorted = data.to_vec();
-    sorted.sort_by(|a, b| a.partial_cmp(b).unwrap());
+    // Use `unwrap_or` to handle NaN values (treat them as equal) instead of panicking.
+    sorted.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
     let mid = sorted.len() / 2;
-    if sorted.len() % 2 == 0 {
+    let result = if sorted.len() % 2 == 0 {
         (sorted[mid - 1] + sorted[mid]) / T::from(2u8)
     } else {
         sorted[mid]
-    }
+    };
+    Some(result)
 }
 
-fn std_dev(data: &[f32]) -> f32 {
-    assert!(data.len() > 1, "standard deviation requires at least two data points");
+fn std_dev(data: &[f32]) -> Option<f32> {
+    if data.len() < 2 {
+        return None;
+    }
     let n = data.len() as f32;
     let mean = data.iter().sum::<f32>() / n;
     let variance = data.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / n;
-    variance.sqrt()
+    Some(variance.sqrt())
 }
 
 pub struct StatisticalChunker {
@@ -255,7 +261,14 @@ impl StatisticalChunker {
         raw_similarities
     }
 
-    fn _find_optimal_threshold(&self, batch_splits: &[&str], similarities: &Vec<f32>) -> f32 {
+    fn _find_optimal_threshold(&self, batch_splits: &[&str], similarities: &[f32]) -> f32 {
+        // Guard: we need at least 2 similarity scores to compute median + std_dev.
+        // With 0 scores there are no chunk boundaries to find; return a neutral threshold.
+        // With 1 score there is no variance to measure; use that single score directly.
+        if similarities.len() < 2 {
+            return similarities.first().copied().unwrap_or(0.5);
+        }
+
         let tokens = self
             .tokenizer
             .encode_batch(batch_splits.to_vec(), true)
@@ -274,8 +287,10 @@ impl StatisticalChunker {
             .collect::<Vec<_>>();
 
         // analyze the distribution of similarity scores to set initial bounds
-        let median_score = median(similarities);
-        let std_dev = std_dev(similarities);
+        // Both median() and std_dev() return Option; the len() >= 2 guard above
+        // ensures they always return Some(_) here.
+        let median_score = median(similarities).unwrap_or(0.5);
+        let std_dev = std_dev(similarities).unwrap_or(0.0);
 
         // set initial bounds based on median and standard deviation
         let mut low = f32::max(0.0, median_score - std_dev);
@@ -300,7 +315,7 @@ impl StatisticalChunker {
                 .map(|(start, end)| cumulative_token_counts[*end] - cumulative_token_counts[*start])
                 .collect();
 
-            median_tokens = median(&split_token_counts);
+            median_tokens = median(&split_token_counts).unwrap_or(0);
 
             if self.min_split_tokens - self.split_token_tolerance <= median_tokens
                 && median_tokens <= self.max_split_tokens + self.split_token_tolerance
@@ -315,7 +330,7 @@ impl StatisticalChunker {
         }
         calculated_threshold
     }
-    fn _find_split_indices(&self, similarities: &Vec<f32>, threshold: f32) -> Vec<usize> {
+    fn _find_split_indices(&self, similarities: &[f32], threshold: f32) -> Vec<usize> {
         let mut split_indices = Vec::new();
         for (idx, score) in enumerate(similarities) {
             if *score < threshold {