Skip to main content

Use the native RT-J backend

Goal: score predictions with the real RT-J model instead of the built-in history baseline.

1. Build the C++ engine​

cd cpp
cmake -B build -S . && cmake --build build -j

This produces cpp/build/librt_c.{dylib,so}. All bindings find it there automatically; elsewhere, set RELATIVEDB_RT_LIB=/path/to/librt_c.dylib.

2. Get the checkpoints​

Default routing resolves hf://stanford-star/rt-j/{classification,regression} against your local Hugging Face cache — nothing downloads implicitly. file:// and plain paths work via a custom ModelConfig.

3. Plug in the backend​

Python (needs pip install -e ".[rt]"):

backend = relativedb.RtNativeBackend(schema=schema)
engine = relativedb.Engine(schema, wiring, model_backend=backend)
result = engine.execute(relativedb.ExecutionInput(query=query, anchor_time=t0))

Java:

TextEncoder encoder = new PrecomputedEncoder(embeddingTable); // string -> float[384]
try (RtNativeBackend backend = new RtNativeBackend(ModelConfig.defaults(), encoder)) {
    RelativeDbEngine engine = RelativeDbEngine.newEngine(schema, wiring)
        .modelBackend(backend).build();
}

Rust:

let engine = Engine::new(schema, wiring)
    .model_backend(Box::new(RtNativeBackend::new(...)));

What to expect​

  • Classification returns probabilities (sigmoid over logits); regression returns denormalized values.
  • Text cells require MiniLM embeddings: Python computes them with sentence-transformers; Java and Rust take a TextEncoder (a precomputed table works for closed vocabularies).
  • Multiclass and ranking currently fall back to the history baseline (the C ABI exposes a single score head).
  • A missing library or checkpoint raises a clear, actionable error — nothing fails silently.