Legal-tech startup
A contract-review tool needs RAG over client documents that must stay under its own control under its enterprise contracts, ruling out most hosted vector database services outright.
Vector Database Hosting for RAG
DevOps services
Vector database hosting for teams building retrieval-augmented generation on top of their own data. We run Qdrant and ChromaDB on VPS and GPU servers, so your embeddings and source documents stay inside your own pipeline. Optional GPU-hosted embedding models keep the whole pipeline self-contained, with DevOps hours included in the plan.
- Embeddings and documents stay inside your own pipeline
- Qdrant and ChromaDB, sized to your dataset
- Optional GPU servers for self-hosted embedding models
- VPS 1–8 h/month; dedicated 5–10 h/month depending on plan; GPU 10 h/monthIncluded hours
- NVIDIA RTX PRO 4000 / RTX PRO 6000 availableGPU hardware
- EUR; applicable taxes itemised on the invoiceBilling
01
What a working RAG pipeline needs
A retrieval-augmented generation system is only as good as its retrieval, and retrieval quality depends on decisions most teams make once and never revisit: chunking strategy, embedding model, index type, and whether you actually measure recall. We set up the pipeline — ingestion, embeddings, vector index, hybrid search combining vector and keyword matching — and, where regulation or a client contract requires it, keep every stage on infrastructure you control.
- Embeddings pipeline from your documents into Qdrant or ChromaDB
- Hybrid search combining vector similarity with keyword filtering
- Metadata filtering so retrieval respects tenant or permission boundaries
- Quantization applied where it cuts cost without hurting recall
02
Best practices we apply
Vector databases are new enough that most teams get the operational basics wrong the first time — public endpoints, no snapshots, no evaluation of whether retrieval is actually working. We apply the same checklist we'd want if it were our own production RAG system, and it's written into the handover documentation. Where documents can't go to a third-party API, we keep both the vector store and any self-hosted embedding step on your own server.
- Private-network access only, authenticated with API keys
- Scheduled snapshots, with restores actually verified
- Quantization tuning for cost, checked against retrieval accuracy
- Metadata filtering enforced at query time, not left to the application
- Retrieval quality evaluated with real queries, not assumed
03
Qdrant vs ChromaDB, and PII in your documents
Qdrant tends to suit larger, production-scale collections with tighter latency requirements; ChromaDB is often faster to prototype with and fine for smaller collections. We'll recommend based on your dataset size and query volume rather than a fixed opinion. If your source documents contain personal data, we also help you decide what should be redacted before embedding, since removing it from the vector index afterwards is far harder than filtering it up front.
04
How the included hours work
DevOps hours come with the plan: 1–8 hours a month on VPS tiers, 10 hours a month on GPU servers and 5–10 hours on dedicated servers depending on the plan — the tier most RAG workloads with self-hosted embedding models will need. Request set-up, a re-indexing pass, or a retrieval-quality review from the client panel under DevOps Services. Hours don't roll over, so most teams front-load the pipeline build in one month and use the next for tuning.
How we work
Request the work from the panel
You describe your document set, expected query volume and any data-handling requirements under DevOps Services, and propose a deadline.
We scope the pipeline and the index
We choose between Qdrant and ChromaDB, size the collection, and confirm whether GPU-hosted embeddings are worth it for your volume.
We build and evaluate retrieval quality
We ingest your documents, build the index, and run real queries against it to check recall before calling the pipeline done.
We hand over with documentation
You receive the pipeline diagram, the snapshot schedule, and the metadata filtering rules that keep results scoped correctly.
Use cases
Financial-services SaaS with strict access control
A fintech product needs metadata filtering enforced at query time, so a support-tier query never surfaces chunks from a document the user isn't entitled to see.
Agency building internal knowledge search
A consultancy indexing years of internal reports moves from an unmanaged ChromaDB instance on a laptop to a hardened, backed-up Qdrant deployment before rolling it out company-wide.
Technology stack
- Qdrant
- ChromaDB
- pgvector
- Ollama
- LangChain
- sentence-transformers
- NVIDIA RTX GPUs
Every subscription includes monthly DevOps, CyberSecurity and Programming hours: 1 h from the cPanel Advanced plan up, 1–8 h on VPS, 5–10 h on dedicated servers depending on the plan, and 10 h on GPU servers. Hours are requested from your client panel, don't roll over to the next month, and anything above the quota is accepted and billed separately.
Frequently asked questions
Qdrant vs Chroma — which vector database is better for RAG?
Qdrant generally suits larger, production-scale collections with stricter latency and filtering needs; ChromaDB is quicker to prototype with and fine for smaller collections. We size the recommendation against your actual document count and query volume rather than defaulting to one.
How do I self-host a RAG pipeline for GDPR?
Run the vector database, the embedding model and the document store all on infrastructure you control, so nothing leaves it during ingestion or querying. We set this up end to end, including GPU-hosted embedding models where needed.
How much RAM or GPU does Qdrant need for millions of vectors?
It depends on vector dimensionality, collection size and whether you quantize. We size this during scoping against your real dataset rather than a generic rule, and apply quantization where it cuts memory cost without a measurable hit to retrieval accuracy.
pgvector vs Qdrant vs ChromaDB — which should I use?
pgvector suits teams who want vector search inside an existing PostgreSQL database and don't need extreme scale. Qdrant and ChromaDB are purpose-built vector databases with more advanced indexing and filtering, better suited once your collection or query volume outgrows pgvector.
When should I migrate from ChromaDB to Qdrant?
Typically once collection size, query latency requirements or the need for advanced metadata filtering outgrow what ChromaDB comfortably handles. We can run a migration with minimal downtime, re-indexing your embeddings into Qdrant and verifying retrieval quality before cutting over.
Can the embedding model itself stay on your infrastructure?
Yes — we can host open-source embedding models on our GPU servers, so documents never need to be sent to a third-party API to be embedded. This matters when contracts or GDPR restrict where the pipeline's data can go.
Building a RAG pipeline that has to stay under your control?
Tell us your document volume and query pattern, and we'll design a Qdrant or ChromaDB deployment that keeps everything inside your own pipeline.