Fix: Embedding API Validation Error

Fix: Embedding API Validation Error

Date: 2026-02-10
Issue: Pipeline processing failed: Embedding error: API error: ‘$.input’ is invalid
Status: ✅ Fixed
Commit: 5b6bcd6a

Problem

When processing documents, the pipeline occasionally failed with:

Pipeline processing failed: Embedding error: API error: '$.input' is invalid. Please check the AP...

This error occurred when embedding providers (OpenAI, Ollama, etc.) received arrays containing empty or whitespace-only strings. API validation rejected these invalid inputs.

Root Cause

The embedding pipeline was passing all text strings to the API without filtering, including:

• Empty strings ("")
• Whitespace-only strings (" ", "\n", "\t")
• Strings that became empty after .trim()

External APIs (OpenAI, Gemini, Jina, etc.) validate input and reject empty strings in the input array.

Solution

All embedding providers now:

1. Filter invalid inputs before API calls:

   let valid_texts: Vec<(usize, &String)> = texts
       .iter()
       .enumerate()
       .filter(|(_, text)| !text.trim().is_empty())
       .collect();

2. Handle all-empty case gracefully:

   if valid_texts.is_empty() {
       return Ok(vec![vec![0.0; self.embedding_dimension]; texts.len()]);
   }

3. Map results back to original indices:

   let mut result = vec![vec![0.0; self.embedding_dimension]; texts.len()];
   for ((orig_idx, _), embedding) in valid_texts.iter().zip(api_embeddings) {
       result[*orig_idx] = embedding;
   }

Affected Providers

All embedding providers were updated:

• ✅ OpenAI (openai.rs)
• ✅ Ollama (ollama.rs)
• ✅ Gemini (gemini.rs)
• ✅ Jina (jina.rs)
• ✅ Azure OpenAI (azure_openai.rs)
• ✅ LM Studio (lmstudio.rs)
• ✅ Mock Provider (mock.rs)

Testing

Unit Tests

All 201 tests pass:

cd edgequake
cargo test --package edgequake-llm --lib
# Result: ok. 201 passed; 0 failed; 0 ignored

Manual Testing

1. Start the backend:

   make dev

2. Upload a problematic PDF document

3. Verify the document processes successfully without embedding errors

4. Check backend logs:

   tail -f /tmp/edgequake-backend.log

Expected: No “Embedding error” messages, document status shows “Completed”

Edge Cases Handled

Input Case	Behavior
All strings valid	Normal processing, all strings embedded
Some strings empty	Empty strings get zero vectors, others processed normally
All strings empty	Return array of zero vectors (dimension-matched)
Whitespace-only	Treated as empty, receives zero vector
Mixed valid/invalid	Valid strings embedded, invalid get zero vectors

Performance Impact

• Negligible overhead: One additional filter() pass over input array
• API call reduction: Fewer strings sent to API when some are empty
• Consistency: Output array size always matches input array size

Code Quality

✅ Clippy: No warnings
✅ Tests: All 201 tests pass
✅ Consistency: All providers use same pattern

Future Improvements

Consider:

1. Log warning when many empty strings are filtered (potential data quality issue)
2. Add telemetry to track how often filtering occurs
3. Upstream validation in chunking/extraction to prevent empty strings earlier

Related Issues

This fix prevents:

• OpenAI API errors: $.input is invalid
• Ollama API errors: invalid input
• Gemini API errors: empty text not allowed

Verification Checklist

• All providers filter empty strings
• Results mapped back to correct indices
• Zero vectors returned for empty inputs
• Array size consistency maintained
• Tests pass
• Clippy clean
• Documentation updated