One filter language for Qdrant, pgvector, and Pinecone

VecStore SDK 0.1.0 is on npm. It gives you one TypeScript API for three vector databases, and one metadata filter that each adapter compiles to the provider’s native syntax.

What ships

• A filter AST with builders such as eq, gt, isIn, and, or, and not, plus a compiler for each provider: Qdrant must and should clauses, pgvector SQL over jsonb, and Pinecone’s MongoDB-style operators.
• Four record verbs (upsert, query, fetch, delete) and three index verbs (createIndex, deleteIndex, listIndexes) with the same signatures on every provider.
• Errors as a discriminated union you match on. The SDK never throws them.
• A raw property that returns the native client you passed in, for anything the unified API does not cover.

Embedding stays out of scope. Generate vectors with your model provider and pass the numbers in.

import { createQdrantStore } from "vecstore-sdk/qdrant";
import { and, eq, gt } from "vecstore-sdk";

const store = createQdrantStore({ client });

const result = await store.index("docs").query({
  vector,
  topK: 5,
  filter: and(eq("lang", "en"), gt("stars", 100)),
});

Read the getting started guide to run your first filtered query.