Summary
On an onnx-jina-coreml backend, vector search failed for some queries but not others against the same index. Vera degrades to BM25-only results and continues, so the only signal is a WARN line.
WARN vera_core::retrieval::hybrid: vector search failed, falling back to BM25-only results
error=storage error: vector search failed
Reproduction
Index used: a clone of this repository at v1.0.0, 375 files, 4790 chunks, embeddings jinaai/jina-embeddings-v5-text-nano-retrieval (fp16), backend onnx-jina-coreml.
Counting only the WARN line, one run per query:
| query |
vector search failed |
fusion |
no |
hybrid fusion |
no |
hybrid fusion work |
no |
how does hybrid fusion work |
yes |
reranking |
no |
cross-encoder reranking of search candidates |
yes |
It is deterministic per query across repeated runs, not flaky. Longer, more natural-language queries are the ones that fail, which is the shape of query the tool is designed for.
Notes
- Reproduces identically on the released v1.0.0 binary and on a local build, so it is not related to any local change.
- The error surfaces at
vector_store.search(...) in crates/vera-core/src/retrieval/vector.rs:69, wrapped as VectorSearchError::StorageError with context "vector search failed". The underlying store error is not included in the logged message, so I could not get further without patching.
RUST_LOG=debug did not add detail beyond the same WARN.
Impact
Silent quality loss. The command exits 0 and returns plausible results, so a user has no reason to suspect the vector half of hybrid retrieval did not run. Combined with reranking being unavailable on this backend (see #33), a CoreML user can be served pure BM25 output while believing they are getting the full pipeline.
Suggestion
Propagate the underlying store error into the warning so the cause is diagnosable from the log. A hybrid search that loses a whole retrieval arm may also deserve louder treatment than a WARN.
Environment
Apple M4, macOS 26.6.1, ONNX Runtime 1.24.4, vera 1.0.0, backend onnx-jina-coreml.
Summary
On an
onnx-jina-coremlbackend,vector search failedfor some queries but not others against the same index. Vera degrades to BM25-only results and continues, so the only signal is a WARN line.Reproduction
Index used: a clone of this repository at v1.0.0, 375 files, 4790 chunks, embeddings
jinaai/jina-embeddings-v5-text-nano-retrieval(fp16), backendonnx-jina-coreml.Counting only the WARN line, one run per query:
fusionhybrid fusionhybrid fusion workhow does hybrid fusion workrerankingcross-encoder reranking of search candidatesIt is deterministic per query across repeated runs, not flaky. Longer, more natural-language queries are the ones that fail, which is the shape of query the tool is designed for.
Notes
vector_store.search(...)incrates/vera-core/src/retrieval/vector.rs:69, wrapped asVectorSearchError::StorageErrorwith context"vector search failed". The underlying store error is not included in the logged message, so I could not get further without patching.RUST_LOG=debugdid not add detail beyond the same WARN.Impact
Silent quality loss. The command exits 0 and returns plausible results, so a user has no reason to suspect the vector half of hybrid retrieval did not run. Combined with reranking being unavailable on this backend (see #33), a CoreML user can be served pure BM25 output while believing they are getting the full pipeline.
Suggestion
Propagate the underlying store error into the warning so the cause is diagnosable from the log. A hybrid search that loses a whole retrieval arm may also deserve louder treatment than a WARN.
Environment
Apple M4, macOS 26.6.1, ONNX Runtime 1.24.4, vera 1.0.0, backend
onnx-jina-coreml.