Phase 5: Voyage-4-Nano ONNX (Opt-in, Deferred) #361

Phase 5: Voyage-4-Nano ONNX (Opt-in)

Scope

• ort crate dependency (ONNX Runtime)
• First-use model download (~1.4GB, auto-cache like fastembed)
• src/embed/voyage.rs — Voyage-4-Nano wrapper (340M params, 1024d, Matryoshka)
• Config: embedding.tier: local activates Voyage-4-Nano
• Re-index with new model: cora index --rebuild

Specs

• Model: voyageai/voyage-4-nano (Apache-2.0)
• ONNX: onnx-community/voyage-4-nano-ONNX (~1.41GB)
• Architecture: Qwen3ForCausalLM, bidirectional attention
• Context: 32K tokens, Dimensions: 1024 (Matryoshka: 256/512/1024/2048)

Trade-offs

• Quality: +15-25% retrieval accuracy over static tokens (full contextual understanding)
• Cost: +1.5-2GB RAM, +3-7min index time (1695 symbols × 200ms)
• Binary: +~5MB (ort crate), model downloaded on first use (not in binary)
• Feature flag: --features voyage-nano to keep default build light

Implementation Checklist

• Add ort crate (feature-gated: voyage-nano)
• Create src/embed/voyage.rs — model download, load, embed
• Wire embedding.tier: local to dispatch voyage.rs
• Handle dim mismatch (768→1024) on tier switch — reindex prompt
• Model download UX: progress indication, cache dir (~/.codecora/cora-code/models/)
• Tests: embed single, embed batch, dim validation

Effort

~400 lines

Depends On

• Phase 3 (usearch index, embed wiring) — this issue
• Phase 4 (optional, no hard dep)

Status

DEFERRED — build when there is user demand or cora review --brain needs better precision. Static tokens + hybrid search already provide 80% of value.

Depends on: #359 (Phase 3)

Status: DEFERRED