v3.12.3

EvolvingLMMs-Lab/Aero-1v3.12.3Jun 17, 2026by ruvnet

AI Summary

Fixes a data quality regression where the MCP memory store was incorrectly emitting 128-dim mock embeddings instead of the expected 384-dim ONNX embeddings.

Key Highlights

Fixed dimensional sanity check to detect and reject mock embeddings
Backend labels now accurately reflect the actual embedding source
Silent mock embedding corruption in similarity recall is resolved

New Features

Mock embedding detection via hardcoded model name check
Route to real ONNX chain when dimensions do not match expected model

Full Release Notes

## Bundled fix

### #2395 — MCP `memory_store` emitted 128-dim mock embeddings (data quality regression)

**Symptom (per issue):** standalone CLI used real 384-dim ONNX embeddings, but the in-session MCP path persistently emitted 128-dim hash-fallback ("mock") embeddings — silently corrupting similarity recall and wasting any benefit of vector memory.

**Root cause:** `bridgeGenerateEmbedding` returned `embedder.embed()` results labeled `backend: 'onnx'` unconditionally, even when AgentDB's vectorBackend controller silently fell back to a 128-dim hash stub. The stub didn't expose `isMock=true`, so the existing isMock check let it through with a wrong label.

**Fix:** dimensional sanity check. The hardcoded model name `Xenova/all-MiniLM-L6-v2` always produces 384-dim; anything else is definitively a stub. Return `null` from the bridge wrapper in that case so the caller falls through to `generateLocalEmbedding`, which routes via the real ONNX chain (transformers.js / ruvector).

Net: backend labels now match actual semantics, no more silent mock embeddings.

## Install

```bash
npx ruflo@3.12.3
# or
npm i ruflo@latest
```

All 3 packages × 3 dist-tags published in lockstep.

🤖 Generated with [RuFlo](https://github.com/ruvnet/ruflo)