Scale Yazılım Training & Consulting · Qdrant
Training & consulting · Vector search

Vector search that holds up after the demo.

Consulting and training for teams building semantic search, RAG and recommendations on Qdrant. We work on production Qdrant clusters every week — sizing them, tuning recall against latency, migrating embedding models and measuring relevance — and we teach what we see there.

Fig. 01 · HNSW graph
Click — or press space — to run a new query How a Qdrant query finds its neighbours. Text is embedded, enters the sparse top layer of the HNSW graph and hops greedily down; the lit points on the bottom layer are the top-3 results.
Made with ISO-GLOW ↗
  • WeeklyHands-on work with production Qdrant clusters
  • 3 daysFrom the first embedding to a production cluster
  • ≤ 15People per class, with a cluster for each of them
  • EN · TRDelivered in English or Turkish, remote or on-site
Consulting

Get it right before it is big. Sizing, quantization, filtering and scaling decisions made on your data, with the trade-offs measured.

Training

Understand what the index is doing. Embeddings, HNSW, hybrid search and operations — with labs on real datasets.

Support

Keep it fast as it grows. Reviews before launches, version upgrades, model changes and incident help when you need it.

Scale Yazılım Ltd. Şti. · Ankara, Türkiye scaleyazilim.com/services/qdrant · info@scaleyazilim.com · 1 / 3
Scale Yazılım Training & Consulting · Qdrant
02 · Consulting

Recall, latency and memory, traded on purpose.

Most vector search problems are choices made early and never measured: the wrong model, filters without indexes, quantization without rescoring, a shard layout that cannot grow. We find them, measure them and fix them with you.

01

Architecture & sizing

Collections, named vectors, multitenancy, memory vs. on-disk storage and node count — for Cloud, Hybrid Cloud or self-hosted.

You getA sizing sheet (RAM and disk per million vectors) and a deployment plan.

02

Search quality

Embedding model choice, hybrid dense + sparse retrieval with RRF, rerankers and filters, judged on your own queries.

You getA judged query set with recall and nDCG, before and after.

03

Performance tuning

HNSW m and ef, payload indexes created before data, scalar, binary and product quantization with rescoring.

You getLatency vs. recall curves for your data and the settings to ship.

04

Scaling & high availability

Sharding, replication factor, resharding, write consistency, snapshots and disaster recovery.

You getA tested failover and restore runbook.

05

Model migration

Change or upgrade embedding models with zero downtime: parallel named vectors, backfill, shadow reads, alias switch.

You getA step-by-step migration plan with rollback.

06

Production readiness

Prometheus and Grafana, optimizer health, memory growth, API keys and JWT access, version upgrades.

You getDashboards, alerts and a reviewed upgrade path.

Signals it is time to call

  • Recall dropped after turning on quantization
  • Filtered searches are slow or quietly miss results
  • Memory keeps growing and nobody is sure why
  • Segments stay unindexed and the collection is yellow
  • You need to change the embedding model in production
  • Thousands of tenants share one collection

How an engagement runs

  1. Scoping call 45 min · free

    We hear the problem, agree goals and the numbers success is judged by.

  2. Assessment 1–2 weeks

    Read-only access, diagnostics, logs and short interviews. Findings in writing.

  3. Fix & tune 2–4 weeks

    Changes made with your team in small steps, each measured before and after.

  4. Hand-over ongoing

    Runbooks, a focused workshop for your engineers and optional retained support.

Field experience E-commerce search RAG & assistants Recommendations Media SaaS Public sector Banking & finance
Scale Yazılım Ltd. Şti. · Ankara, Türkiye scaleyazilim.com/services/qdrant · info@scaleyazilim.com · 2 / 3
Scale Yazılım Training & Consulting · Qdrant
03 · Training

From the first embedding to a production cluster.

Three days, one idea at a time, each followed by a lab on your own cluster. A shorter relevance workshop is available for teams that already run vector search and want to measure it properly.

Programme 01 · Core

Vector Databases with Qdrant

3 days · 18 hours
Level
Intermediate
Prereq.
Basic Linux, Docker and Python
  • Day 1 · Foundations. Embeddings and distance metrics; vector vs. traditional databases; Qdrant architecture; Docker setup, Python client, REST and gRPC; a first RAG prototype
  • Day 2 · Search. Batch upload, payloads and nested fields, named vectors; filters (must, should, must_not, range, geo); hybrid search, score thresholds, pagination
  • Day 2 · Index. HNSW parameters, scalar and product quantization, indexing strategies, sharding and replication; Qdrant or Elasticsearch — when to use which
  • Labs. An e-commerce product search and a semantic search engine, built end to end
  • Day 3 · Production. API keys and JWT, snapshots and backups, monitoring and logging, clusters, resource limits, horizontal scaling, replicas, load balancing and disaster recovery
  • Day 3 · Practice. Choosing an embedding model, filtering strategies, error handling, retries and rate limiting
Programme 02 · Workshop

Measuring Vector Search Relevance

1 day · 6 hours
Level
Intermediate – advanced
Prereq.
Programme 01 or equivalent
  • Building a judged query set quickly, with AI-assisted drafts and human review
  • Recall@k, precision, nDCG and MRR — what each one hides
  • Exact vs. approximate search: measuring what HNSW misses
  • Dense, sparse (BM25, SPLADE) and hybrid with RRF and DBSF
  • Filters, payload indexes and the cost of creating them late
  • Quantization with rescoring: memory saved vs. recall lost
  • Rerankers and when they pay for themselves
  • A weekly scorecard your team can keep running

Who it is for

  • Engineers building search, RAG or recommendations
  • Architects comparing Qdrant, Elasticsearch, pgvector

What your team leaves with

  • Hybrid search on their own data, tuned on purpose
  • A production checklist they can apply on day one
  • DeliveryRemote or on-site; daytime, evening or weekend sessions
  • LabsA cloud lab for every participant, on realistic data
  • Class sizeUp to 15 people, so everyone gets hands-on time
  • LanguageEnglish or Turkish, with all materials provided
  • TailoringContent, length and examples fitted to your stack
Next step

Start with a 45-minute scoping call.

Tell us what is slow, wrong or about to grow. We reply within one working day with a suggested next step — consulting, training or both.

Scale Yazılım Ltd. Şti. · Ankara, Türkiye scaleyazilim.com/services/qdrant · info@scaleyazilim.com · 3 / 3