@molecule/api-ai-vector-store-memory

Auto-generated, AI-first package reference for the molecule.dev ecosystem. It is written to be read by coding agents as much as by people, and is generated from this package's source — edit src/index.ts JSDoc, not this file.

In-memory ai-vector-store provider for molecule.dev.

A brute-force-cosine vector store held entirely in process memory, with zero external dependencies — ideal for small corpora, tests, and local development (the pgvector / Pinecone / Chroma providers all need an external service). Bond it once at startup, then use the @molecule/api-ai-vector-store core.

Quick Start

import { setProvider, requireProvider } from '@molecule/api-ai-vector-store'
import { provider } from '@molecule/api-ai-vector-store-memory'

setProvider(provider) // at startup

const store = requireProvider()
await store.createCollection({ name: 'docs', dimension: 384, metric: 'cosine' })
await store.upsert({
  collection: 'docs',
  records: [{ id: 'a', embedding: vec, metadata: { topic: 'x' } }],
})
const hits = await store.query({ collection: 'docs', embedding: queryVec, topK: 5 })

Type

provider

Installation

npm install @molecule/api-ai-vector-store-memory @molecule/api-ai-vector-store

API

Constants

provider

In-memory vector store provider.

Implements the AIVectorStoreProvider interface with process-local state and a brute-force similarity scan. No persistence, no external services.

const provider: AIVectorStoreProvider

Core Interface

Implements @molecule/api-ai-vector-store interface.

Bond Wiring

Setup function to register this provider with the core interface:

import { setProvider } from '@molecule/api-ai-vector-store'
import { provider } from '@molecule/api-ai-vector-store-memory'

export function setupAiVectorStoreMemory(): void {
  setProvider(provider)
}

Injection Notes

Requirements

Peer dependencies:

• @molecule/api-ai-vector-store >=1.0.1

Runtime Dependencies

• @molecule/api-ai-vector-store

• Not persistent — the index lives in process memory and is gone on restart. Rebuild it at startup, or use a persistent provider (pgvector/Pinecone) for durable data.

• upsert throws if the collection doesn't exist, or if an embedding's length differs from the collection's dimension (validated before any write, so a bad batch leaves the collection unchanged).

• Query is O(n) per call (brute-force cosine) — great for thousands of vectors, not millions.

E2E Tests

Integration checklist — drive the real flow (no mocks): upsert real vectors, run real query calls, and verify through the FEATURE this store powers (semantic search / RAG retrieval / related-items) plus direct property checks on the returned VectorSearchResult[]. Adapt each item to this app's actual corpus/screens and check every box off. A box you can't check is an integration bug to fix — not a skip:

• [ ] upsert records (each with a stable id + metadata), then query with an embedding: results come back ranked by score (highest first), the semantically closest stored item is result #1 and unrelated items rank lower — the whole point. score is a sane similarity (bounded, ~0–1, higher = closer) and each hit's record.id / record.metadata come back intact.
• [ ] topK is honored: a query with topK: k returns AT MOST k results, best-first — never more, never unordered.
• [ ] Metadata filter works: a query carrying a MetadataFilter (e.g. { field: 'userId', operator: 'eq', value }) returns only records matching the filter and never leaks non-matching ones.
• [ ] Collection/namespace ISOLATION: a query scoped to one collection never returns another collection's vectors — the multi-tenant boundary that keeps one user's private docs out of another's results. Confirm with two collections (or two owner ids) that a scoped query returns only its own.
• [ ] delete removes a record: after delete({ collection, ids }) the vector stops appearing in query results (and fetch omits it).
• [ ] The feature built on the store returns MEANING-ranked results end-to-end in the UI — a semantic-search / RAG / related-items query surfaces the relevant items first, not a keyword or insertion-order match. This store does NOT embed text itself, so confirm it composes with @molecule/api-ai-embeddings (query text → embedding → query).
• [ ] Every upsert / query runs SERVER-SIDE — the provider/store key stays on the server and never ships in the browser bundle (the package is server-only; a client import throws by design).