sandx-embed

Shared embedding and vector similarity infrastructure for the SandX platform.

Part of the SandX Lab computational infrastructure ecosystem.

What It Does

sandx-embed is the shared latent representation layer used by all SandX engines. It provides:

• Pluggable encoders — sentence-transformers models out of the box; register any custom encoder
• High-performance ANN indexing — HNSW (production) and exact search (baseline), with save/load
• Cross-domain similarity — cosine, L2, inner product; normalized and unnormalized vectors

Not a standalone product — consumed by sandx-er, sandx-graph, and sandx-compute as a shared dependency.

Status

v0.1 — Phase 2 active development

Installation

pip install sandx-embed

Or from source:

git clone https://github.com/sandxlab/sandx-embed
cd sandx-embed
pip install -e ".[dev]"

Demo

pip install "sandx-embed[hnsw]"
python -m examples.encode_and_search

Encodes 15 company name variants into dense vectors, builds an ANN index, runs 4 semantic queries, and prints a pairwise similarity matrix — no external data required.

Quick Start

from sandx_embed import Encoder, VectorIndex

# Encode records into dense vectors
enc = Encoder(model="sentence-bert")   # downloads all-MiniLM-L6-v2 on first use
vectors = enc.encode(["John Smith, Boston", "Jon Smyth, Boston", "Alice Brown, NYC"])
# → np.ndarray shape (3, 384), L2-normalized

# Build an ANN index
idx = VectorIndex(method="hnsw", metric="cosine")
idx.build(vectors, ids=["r0", "r1", "r2"])

# Query nearest neighbors
result = idx.query(vectors[0], k=2)
print(result.ids)        # ["r0", "r1"]
print(result.distances)  # [0.0, 0.12]  (cosine distance)

# Persist and reload
idx.save("/tmp/my_index")
idx2 = VectorIndex.load("/tmp/my_index")

Built-in Models

Custom Encoders

from sandx_embed.encoder import BaseEncoder, Encoder
import numpy as np

class MyEncoder(BaseEncoder):
    def encode(self, inputs, *, batch_size=64, normalize=True):
        # your model here
        return np.random.rand(len(inputs), 128).astype(np.float32)
    @property
    def dim(self): return 128

Encoder.register("my-model", lambda: MyEncoder())
enc = Encoder("my-model")

Index Methods

Design Principles

• Pluggable — any encoder model or index backend can be registered
• Portable — indexes serialize to disk and reload without rebuilding
• Deterministic — same model version + input → same output
• No vendor lock-in — no hard dependency on any hosted vector service

Related

• sandx-er — entity resolution engine (uses sandx-embed for blocking + matching)
• sandx-graph — graph intelligence over resolved entities
• sandx.io — project home

License

Apache 2.0 — see LICENSE