Embex — Universal Vector Database Client

7+Adaptersone interface across providers

4xFaster Hot PathsRust + SIMD accelerated core

1 lineProvider Switchconfiguration, not rewrites

Stop Rewriting Vector Infrastructure

Every provider ships different client semantics. Embex gives your app one stable surface, then maps it to each backend adapter.

Provider APIs

# Pinecone
index.upsert(vectors=[(id, values, metadata)])

# Qdrant
client.upsert(collection_name=name, points=points)

# Weaviate
client.data_object.create(data_object, class_name)

Embex API

# Same call pattern for all adapters
await client.collection("products").insert(points)
results = await client.collection("products").search(vector=query, top_k=5)

Switch Provider

client = await EmbexClient.new_async("lancedb", "./data")
client = await EmbexClient.new_async("qdrant", "http://localhost:6333")

---

Quick Start

Python

pip install embex lancedb

import asyncio
from embex import EmbexClient, Point

async def main():
    client = await EmbexClient.new_async("lancedb", "./data")
    collection = client.collection("docs")
    await collection.create(dimension=768, distance="cosine")
    await collection.insert([
        Point(id="1", vector=[0.1] * 768, metadata={"topic": "intro"})
    ])
    print(await collection.search(vector=[0.1] * 768, top_k=3))

asyncio.run(main())

Node.js

npm install @bridgerust/embex lancedb

import { EmbexClient } from "@bridgerust/embex";

const client = await EmbexClient.newAsync("lancedb", "./data");
const collection = client.collection("docs");
await collection.create(768, "cosine");
await collection.insert([
  { id: "1", vector: Array(768).fill(0.1), metadata: { topic: "intro" } },
]);
console.log(await collection.search(Array(768).fill(0.1), 3));

---

Why Teams Pick Embex

Unified Adapter Surface LanceDB, Qdrant, Pinecone, Chroma, Milvus, PgVector, Weaviate and upcoming adapters.

Runtime Stability Language boundary errors are mapped instead of crashing host apps.

Migration Path Start embedded locally, move to managed infrastructure with minimal change.

Production Readiness Connection pooling, retries, typed models, and pragmatic defaults.

---

Performance Snapshot

Provider	Client	Insert (ops/s)	Speedup	Search Latency
Qdrant	Embex	24,825	4.3x	1.95 ms
Qdrant	Native	5,754	-	4.69 ms
Weaviate	Embex	5,163	4.1x	1.77 ms
Weaviate	Native	1,256	-	4.03 ms

See full methodology in Benchmarks.

---

Embex vs Common Approaches

Capability	Raw Provider SDKs	LLM Framework Wrappers	Embex
Stable provider-agnostic API	No	Partial	Yes
Swap providers with minimal code changes	No	Usually no	Yes
Rust-level performance path	Rare	Rare	Yes
Embedded local start	Varies	Varies	Yes
Vector-first, no forced LLM coupling	Yes	No	Yes

Build Against One Vector Interface

Use Embex for application-level consistency and keep provider choice flexible as your scale changes.

Get StartedExplore ProvidersBuilt with BridgeRust