Cut Vector Search Costs by 95% with Self-Hosted Qdrant on $6/Month

Five million vectors, 800,000 monthly searches, and a hard latency limit under 50 milliseconds. When a developer’s RAG pipeline for legal contract Q&A hit Pinecone’s pricing wall, the solution wasn’t a bigger cloud plan—it was taking control.

Six months ago, they migrated from Pinecone Serverless to a self-hosted Qdrant instance on a budget VPS. The bill dropped from $210 per month to under $7, with no compromise on speed or accuracy. Here’s the complete breakdown of the setup, migration, and trade-offs.

The Starting Point: High Costs and Predictable Workloads

The application processes legal contracts using OpenAI embeddings and vector search to answer questions in natural language. The dataset grew to 5.2 million 1536-dimensional vectors, and the system handles nearly 800,000 queries monthly. A strict P99 latency target of under 50 milliseconds ruled out slower object-store solutions.

On Pinecone Serverless, the monthly invoice consistently reached $210. That covered storage, read units for frequent queries, and write units for daily document ingestion. The cost scaled with usage, making it expensive to scale up or down.

The Shift to Self-Hosting: Hardware and Software Choices

Instead of relying on a managed vector database, the developer chose a single Hetzner CX32 cloud server:

• 4 virtual CPUs and 8 GB of RAM
• 80 GB SSD storage
• €8.50 per month (approximately $9.20)

They ran Qdrant in a Docker container and configured automated daily backups to S3-compatible storage, adding around $0.50 to the monthly bill. Total infrastructure cost: about $10 per month—a 95% reduction from Pinecone.

Migrating Vectors: A Straightforward Process

The migration took less than a day and required only a few scripts.

First, data was exported from Pinecone using their scroll API. A simple Python script handled the job:

export_pinecone.py --index legal-docs --output vectors.jsonl

Next, Qdrant was started with a single Docker command that mapped a local volume for persistent storage:

docker run -d -p 6333:6333 -v ./storage:/qdrant/storage qdrant/qdrant

Finally, the vectors were imported using another Python script:

import_qdrant.py --input vectors.jsonl --collection legal-docs

The API surface of Qdrant closely mirrored Pinecone’s, making the transition smooth. No rewrites to application code were necessary.

Performance Comparison: Faster and Cheaper

A controlled test compared 10,000 identical queries on both systems:

| Metric | Pinecone Serverless | Qdrant Self-Hosted | |----------------------------|---------------------|--------------------| | P50 Latency | 23 ms | 4 ms | | P99 Latency | 89 ms | 12 ms | | Recall@10 | 0.97 | 0.97 | | Monthly Cost | $210 | $10 |

The self-hosted setup was faster because the full vector index resided in RAM on the local machine. Pinecone Serverless, by contrast, loads data on demand from object storage, introducing cold-start delays.

When Self-Hosted Vector Search Makes Sense

Not every team should move away from managed services. Self-hosting is a strong fit when:

• The team has basic Docker and server management skills
• Cost sensitivity is high—saving $2,400 per year between $10 and $210 is meaningful
• Data control and indexing flexibility are priorities
• The vector count grows predictably over time

Self-hosting is less ideal when:

• The team lacks DevOps experience or support
• Uptime requirements exceed 99.9% SLA for enterprise clients
• Vector volume fluctuates unpredictably (e.g., from 10M to 100M in one month)
• Every engineering hour is critical to product development

Cost at Scale: A Clear Pattern

A cost comparison tool built by the developer shows how expenses diverge across major vector databases as data scales:

| Scale | Pinecone | Qdrant Cloud | Qdrant Self-Hosted | Supabase pgvector | |------------|----------------|--------------|--------------------|------------------| | 1M vectors | ~$22/month | ~$14/month | ~$7/month | ~$27/month | | 10M vectors| ~$210/month | ~$120/month | ~$72/month | ~$95/month | | 100M vectors| ~$1,900/month | ~$950/month | ~$480/month | N/A |

Self-hosting Qdrant consistently delivers the lowest total cost, especially at higher volumes.

One Trade-Off: Losing the Dashboard

The only feature the developer missed from Pinecone was the web dashboard. It allowed browsing vectors, running test queries, and viewing index statistics visually. Qdrant offers a basic web UI, but it lacks the polish of managed solutions. For now, curl and Python scripts handle monitoring.

Would they switch back? Not at this cost difference. For prototyping, though, Pinecone’s free tier (100,000 vectors) remains a practical starting point.

The takeaway is clear: if your vector workload is predictable and your team can manage a server, self-hosting with Qdrant can deliver dramatic savings without sacrificing performance.