When to use Qdrant

Use Qdrant when an agent or app needs durable semantic memory:

• Retrieval-augmented generation (RAG)
• Document and knowledge-base search
• Product or lead similarity matching
• Long-running agent memory across sessions
• Embeddings produced inside sandboxes or computers

Use object storage for the raw files and Qdrant for the vectors and payload metadata that make those files searchable.

What you get

Per project: a dedicated Qdrant instance. Credentials injected at boot:

QDRANT_URL=https://qdrant.proj_xxx.miosa.internal
QDRANT_API_KEY=...

Use any Qdrant-compatible client:

import { QdrantClient } from "@qdrant/js-client-rest"

const client = new QdrantClient({
  url: process.env.QDRANT_URL,
  apiKey: process.env.QDRANT_API_KEY,
})

// Create a collection
await client.createCollection("documents", {
  vectors: { size: 1536, distance: "Cosine" },
})

// Upsert vectors
await client.upsert("documents", {
  points: [
    { id: 1, vector: embedding, payload: { text: "Hello world" } },
  ],
})

// Semantic search
const results = await client.search("documents", {
  vector: queryEmbedding,
  limit: 5,
  with_payload: true,
})

from qdrant_client import QdrantClient
import os

client = QdrantClient(
    url=os.environ["QDRANT_URL"],
    api_key=os.environ["QDRANT_API_KEY"],
)

results = client.search(
    collection_name="documents",
    query_vector=query_embedding,
    limit=5,
)

Provision

Add Qdrant when creating a project:

const project = await miosa.projects.create({
  name: "my-rag-app",
  data: { qdrant: true },
})

Or add to an existing project:

await miosa.projects.dataServices.create(projectId, {
  type: "qdrant",
  size: "small",  // small | medium | large
})

cURL:

curl -X POST https://api.miosa.ai/v1/projects/$PROJECT_ID/data-services 
  -H "Authorization: Bearer $MIOSA_API_KEY" 
  -H "Content-Type: application/json" 
  -d '{"type":"qdrant","size":"small"}'

Data persists past VM destroy

Collections and vectors survive sandbox destruction and deployment rollbacks. The Qdrant instance keeps running between VM lifecycles. You do not need to re-index on every boot.

Typical pattern: RAG pipeline

import OpenAI from "openai"
import { QdrantClient } from "@qdrant/js-client-rest"

const openai = new OpenAI()
const qdrant = new QdrantClient({
  url: process.env.QDRANT_URL,
  apiKey: process.env.QDRANT_API_KEY,
})

// Index a document
async function index(id: number, text: string) {
  const { data } = await openai.embeddings.create({
    model: "text-embedding-3-small",
    input: text,
  })
  await qdrant.upsert("docs", {
    points: [{ id, vector: data[0].embedding, payload: { text } }],
  })
}

// Query
async function query(question: string, topK = 5) {
  const { data } = await openai.embeddings.create({
    model: "text-embedding-3-small",
    input: question,
  })
  return qdrant.search("docs", {
    vector: data[0].embedding,
    limit: topK,
    with_payload: true,
  })
}

Limits

Vector count per collection is bounded by memory. At 1536 dimensions (OpenAI text-embedding-3-small), small fits ~100K vectors; medium fits ~400K.

Bring your own Qdrant?

Yes. Set QDRANT_URL and QDRANT_API_KEY as env vars to point at an external instance (Qdrant Cloud, self-hosted):

await miosa.deployments.env.set(deploymentId, {
  environment: "production",
  vars: {
    QDRANT_URL: "https://your-cluster.qdrant.io",
    QDRANT_API_KEY: "your-api-key",
  },
})

Python SDK

See also