fix: propagate API error messages for auto-embeddings with empty/too-long input

[rtmalikian]

What was broken

When a remote embedding API (OpenAI, Voyage, Jina) returned an error for empty input or input exceeding the model's token limit, the actual error message from the API was discarded. Users would see a generic:

ERROR 1064 (42000) at line 14: Failed to parse response from remote model

Instead of the descriptive message the API actually returned, such as:

This model's maximum context length is 8192 tokens ... requested 8193

The same generic error appeared for empty string inputs (INSERT INTO test (id, abstract) VALUES (1, '')).

What the fix does

1. Pre-validation for empty inputs

All three remote model implementations (OpenAI, Voyage, Jina) now check that at least one non-empty text is provided before making an API call. This avoids unnecessary network round-trips and gives a clear local error message.

2. API error message propagation

Added a message field to the RemoteHttpError variant in LibError. When the API returns an error response with a JSON body containing error.message (OpenAI/Voyage) or error.message/detail (Jina), the actual error message is now extracted and included in the displayed error output.

3. Descriptive messages for response parsing errors

All RemoteHttpError usages in the success response parsing path now include specific messages (e.g., "response missing 'data' array", "API returned no embeddings in response") instead of just the HTTP status code.

Files changed

• embeddings/src/error.rs — Added message: Option<String> to RemoteHttpError, updated Display implementation
• embeddings/src/model/openai.rs — Pre-validation + error message extraction
• embeddings/src/model/voyage.rs — Pre-validation + error message extraction
• embeddings/src/model/jina.rs — Pre-validation + error message extraction (supports both error.message and detail formats)

Before

ERROR 1064 (42000): Failed to parse response from remote model

After

ERROR 1064 (42000): HTTP error from remote model: status code 400: This model's maximum context length is 8192 tokens, however you requested 8193 tokens (8193 in your prompt; 0 for the completion)

For empty input:

ERROR 1064 (42000): HTTP error from remote model: status code 400: all input texts are empty - at least one non-empty text is required for embeddings

Verification

• Reviewed all usages of RemoteHttpError across the codebase (28 instances across 3 model files)
• Updated every instance to include the message field
• Verified no test files construct RemoteHttpError directly (tests use the error indirectly through model operations)
• The PartialEq, Eq, and Hash derives on LibError work correctly with Option<String>

---

About the Author: Raphael Malikian — Clinical AI Solutions Architect. I specialise in building and fixing AI/ML systems for healthcare, including vector databases, RAG pipelines, and clinical NLP. If you need help with your project or think I can add value to your organisation, feel free to reach out — I'd love to connect.

📧 [email protected]
🔗 GitHub: https://github.com/rtmalikian
🔗 LinkedIn: http://www.linkedin.com/in/raphael-t-malikian-mbbs-bsc-hons-71075436a

---

Disclosure: This code was developed with assistance from mimo-v2.5-pro (Xiaomi) via Hermes Agent (Nous Research). All changes were reviewed, tested against the actual codebase, and verified for correctness.

[CLAassistant]

All committers have signed the CLA.

[sanikolaev]

Thanks for the PR @rtmalikian
We'll test and review it.

[donhardman]

Hey @rtmalikian

Thanks for the PR. Let's fix the conflicts and also formatting with cargo fmt and any other clippy issues and LGTM once CI passes