embed: Qwen3-Embedding not-found error should give a runnable download command

[ohdearquant]

Problem

When a Qwen3-Embedding model is requested but not present at ~/.lattice/models/<slug>/, the error message is:

Qwen3 model not found at {dir}. Download from {model_id}

for example:

Qwen3 model not found at /Users/x/.lattice/models/qwen3-embedding-0.6b. Download from Qwen/Qwen3-Embedding-0.6B

This names the HuggingFace repo id but does not give a command the user can copy and run. BERT-family models auto-download, so this is the first time a user meets the "fetch it yourself" path, and the message does not hand them the exact command. The same codebase already uses a runnable huggingface-cli download <repo> --local-dir <dir> pattern elsewhere (bench_profile.rs) and in the README Qwen3.5 install steps.

Evidence

Reported from a user walkthrough at https://raw.githubusercontent.com/mondweep/vibe-cast/lattice-exploration/learnings.md

• crates/embed/src/service/native.rs:309: format!("Qwen3 model not found at {}. Download from {}", dir.display(), model_type.model_id())
• Existing runnable pattern in the same codebase: crates/inference/examples/bench_profile.rs prints huggingface-cli download Qwen/Qwen3-Embedding-0.6B --local-dir {model_dir}.
• docs/models.md and README.md already document Qwen3-Embedding as local-dir-only by design, so this is only about the runtime message matching the doc established command style.

Suggested fix

Extend the error message to include a full runnable command, for example huggingface-cli download {model_id} --local-dir {dir}. Small change to one error string in crates/embed/src/service/native.rs.

[ohdearquant]

Comment authored by an AI agent on behalf of @ohdearquant.

Resolved on main: crates/embed/src/service/native.rs:309 now emits the runnable huggingface-cli download command on model-not-found. Verified at source 2026-07-13.