Return errors instead of raising

[joelpaulkoch]

We raise in Rag.Embedding.generate_embeddings_batch/3, we should return an error tuple instead to let the library user decide how to handle the error.

[onestardao]

yep — this exact pattern is one of the silent killers in LLM pipelines.
once you raise inside embedding logic, especially batch ops, any upstream retry/recovery strategy dies immediately.
it’s a classic case of asymmetric error propagation — the system looks like it failed “early,” but in real use (especially RAG), it’s fatal downstream because recovery becomes opaque.

we actually cataloged this as Problem No.1 in our open-source reasoning engine — and added a patch that lets the failure flow semantically with full traceability, without disrupting upstream logic.

if you’re interested i can point you to the module. it’s fully MIT and tested across RAG+OCR edge cases. let me know.

[joelpaulkoch]

Sure, always interested ;)

[onestardao]

great to hear! we actually mapped this exact failure as Problem No.1 in our diagnostic map ~~ and now maintain a public Semantic Clinic where most AI infra & logic failures are cataloged, debugged, and patched.
all MIT, open, and already validated across RAG, OCR, batch logic, and more.

if it helps your work, feel free to star the repo ~~~no pressure, just happy to see folks finding it useful. thanks!