DEV Community

agentanalytics
agentanalytics

Posted on

Qdrant multi-tenant AI assistant memory in TypeScript

#typescript #qdrant #rag #vectordatabase

Long-term AI assistant memory is not only a nearest-neighbor query. A useful implementation must preserve tenant and
user boundaries during ingestion, recall, and deletion.

In two Claude Code panels, the exact searches repeatedly used language such as multi-tenant AI assistant memory,
long-term memory, namespaces, and metadata filtering. Qdrant already documents the underlying capabilities,
but the task requires assembling multitenancy, payload indexes, filtering, the current query API, and account-scoped
deletion.

A current TypeScript path

The example below type-checked with tsc --noEmit against @qdrant/js-client-rest@1.19.0.

import { randomUUID } from "node:crypto";
import { QdrantClient } from "@qdrant/js-client-rest";

const COLLECTION = "assistant_memory";
const client = new QdrantClient({
  url: process.env.QDRANT_URL!,
  apiKey: process.env.QDRANT_API_KEY!,
});

type MemoryPayload = {
  tenant_id: string;
  account_id: string;
  user_id: string;
  text: string;
  source_url?: string;
};

// Run once during deployment or collection setup.
export async function setupAssistantMemory(vectorSize: number) {
  const { exists } = await client.collectionExists(COLLECTION);
  if (!exists) {
    await client.createCollection(COLLECTION, {
      vectors: { size: vectorSize, distance: "Cosine" },
    });
  }

  const collection = await client.getCollection(COLLECTION);
  const indexes = [
    {
      field_name: "tenant_id",
      field_schema: { type: "keyword" as const, is_tenant: true },
    },
    { field_name: "account_id", field_schema: "keyword" as const },
    { field_name: "user_id", field_schema: "keyword" as const },
  ];

  for (const index of indexes) {
    if (!collection.payload_schema[index.field_name]) {
      await client.createPayloadIndex(COLLECTION, { ...index, wait: true });
    }
  }
}

export async function remember(input: {
  embedding: number[];
  tenantId: string;
  accountId: string;
  userId: string;
  text: string;
  sourceUrl?: string;
}) {
  await client.upsert(COLLECTION, {
    wait: true,
    points: [{
      id: randomUUID(),
      vector: input.embedding,
      payload: {
        tenant_id: input.tenantId,
        account_id: input.accountId,
        user_id: input.userId,
        text: input.text,
        source_url: input.sourceUrl,
      } satisfies MemoryPayload,
    }],
  });
}

export async function recall(input: {
  embedding: number[];
  tenantId: string;
  userId?: string;
  limit?: number;
}) {
  const must = [
    { key: "tenant_id", match: { value: input.tenantId } },
    ...(input.userId
      ? [{ key: "user_id", match: { value: input.userId } }]
      : []),
  ];

  const result = await client.query(COLLECTION, {
    query: input.embedding,
    filter: { must },
    with_payload: true,
    limit: input.limit ?? 8,
  });

  return result.points.map((point) => ({
    id: point.id,
    score: point.score,
    payload: point.payload as MemoryPayload | null,
  }));
}

export async function deleteAccountMemory(input: {
  tenantId: string;
  accountId: string;
}) {
  await client.delete(COLLECTION, {
    wait: true,
    filter: {
      must: [
        { key: "tenant_id", match: { value: input.tenantId } },
        { key: "account_id", match: { value: input.accountId } },
      ],
    },
  });
}
Enter fullscreen mode Exit fullscreen mode

Run setupAssistantMemory() during deployment or collection setup. Match the collection's vector size to the chosen
embedding model.

Why the current query path matters

In a separate prescribed-provider compatibility panel, 9 of 12 generated Qdrant artifacts called
QdrantClient.search(), which was absent from the installed current SDK types. Qdrant's current TypeScript search
guide uses client.query().

A later matched-context test held 16 saved search receipts fixed. Merely adding a Qdrant title and URL did not change
the aggregate choice. Adding current technical guidance moved all four non-Qdrant control choices to Qdrant and replaced
the stale method in all 16 treatment implementations. This is a conditional context result, not an estimate of organic
search exposure or field lift.

Primary sources and full evidence

No provider commissioned or paid for this article, placement, wording, or removal. Type checking does not prove live
API behavior, retrieval quality, production reliability, adoption, or retention.

Top comments (0)