Qdrant Learning Syllabus — From Scratch to Production

Phase 0: Prerequisites & Conceptual Foundations (3–5 days)

Goal: Understand why vector databases exist before touching Qdrant itself.

[x] What embeddings are (text/image/audio → dense vectors) and how similarity search works (cosine, dot product, Euclidean distance)
[x] Difference between exact nearest neighbor (kNN) and approximate nearest neighbor (ANN)
[ ] HNSW (Hierarchical Navigable Small World) algorithm — conceptual understanding, not implementation
[ ] Where vector DBs fit in an AI stack: embedding models → vector DB → retrieval → LLM (RAG pattern)
[ ] Landscape context: how Qdrant compares to Pinecone, Weaviate, Milvus, pgvector, Chroma (know the trade-offs, not just Qdrant in isolation)

[ ] Install via Docker (docker run qdrant/qdrant), local binary, and Qdrant Cloud free tier
[ ] Qdrant Web UI dashboard — explore collections visually
[ ] REST API vs gRPC API — when to use which
[ ] Client SDKs overview (Python, JS/TS, Go, Rust, Java) — pick Python as primary if doing RAG/ML work

[ ] Collections, points, vectors, payloads (metadata)
[ ] Point IDs (UUID vs integer)
[ ] Distance metrics: Cosine, Dot, Euclidean, Manhattan — when each applies
[ ] Named vectors / multivectors (storing multiple embeddings per point — e.g., text + image)

[ ] Payload schema design — flat vs nested metadata
[ ] Filtering: must, should, must_not, range filters, geo filters, full-text match filters
[ ] Payload indexing for filter performance (keyword, integer, geo, text indexes)
[ ] Pagination: offset, scroll API for large result sets
[ ] Faceting — aggregating results by payload field
[ ] Grouping (query_groups) — e.g., top-N per category
[ ] Sparse vectors and hybrid search (dense + sparse via BM25/SPLADE-style scoring, RRF and DBSF fusion)
[ ] Recommendation API (positive/negative example points)
[ ] Discovery API (constraining search to a region of vector space using context pairs)
[ ] Relevance tuning: Maximal Marginal Relevance (MMR), Relevance Feedback queries (Qdrant 1.17+ feature)

[ ] Pick an embedding model (OpenAI, Cohere, or open-source via sentence-transformers/FastEmbed)
[ ] Qdrant's built-in FastEmbed integration for local, no-API-key embedding generation
[ ] Build an end-to-end RAG pipeline: document chunking → embed → upsert with payload (source, chunk index, doc id) → query-time retrieval → pass to LLM
[ ] Multitenancy patterns — isolating data per user/tenant within one collection (payload-based partitioning vs separate collections)
[ ] Update strategies: upserting changed documents, handling deletes, versioning

[ ] HNSW index tuning: m, ef_construct, ef (search-time) parameters and their trade-offs
[ ] Quantization: Scalar, Product, and Binary quantization — memory savings vs recall trade-offs (binary quantization can give up to ~40x memory reduction)
[ ] On-disk vs in-memory storage (on_disk payload/vector options) — memory-mapped storage for large collections
[ ] Segment architecture and optimizers (indexing thresholds, segment merging)
[ ] Benchmarking: measuring recall, latency (p50/p95/p99), and throughput under load
[ ] Choosing vector precision (float32 vs float16 vs uint8)

[ ] Single-node limits — when you need to scale out
[ ] Sharding: distributing a collection across nodes
[ ] Replication for high availability
[ ] Consistency levels for reads/writes
[ ] Cluster setup (self-hosted, Raft-based consensus for cluster state)
[ ] Snapshotting and backup/restore procedures
[ ] Zero-downtime collection migrations and re-indexing strategies
[ ] Multi-AZ deployment patterns (if using Qdrant Cloud) for failover without downtime

[ ] Security: API key auth, TLS, role-based access control (RBAC) if on Cloud/Enterprise
[ ] Observability: Prometheus metrics integration, per-collection metrics, audit logging, telemetry API
[ ] Resource management: memory sizing calculations (vectors × dimensions × bytes-per-value + overhead), CPU sizing for indexing vs querying
[ ] Rate limiting & quotas: protecting cluster from noisy-neighbor issues (global quota APIs)
[ ] Deployment strategies: Docker Compose for small setups, Kubernetes (Helm chart) for scale, Qdrant Cloud managed vs self-hosted vs hybrid vs private deployment trade-offs
[ ] CI/CD: automating collection schema migrations, testing embedding pipeline changes
[ ] Cost optimization: quantization + on-disk storage to reduce cluster size; GPU-accelerated indexing for large bulk-loading jobs

[ ] Building hybrid search systems combining keyword (BM25/sparse) + semantic (dense) retrieval well
[ ] Re-ranking pipelines (cross-encoders after Qdrant's initial retrieval)
[ ] Multi-modal search (text-to-image, image-to-image) using named vectors
[ ] Integrating Qdrant with LangChain, LlamaIndex, Haystack for RAG frameworks
[ ] A/B testing different embedding models on the same dataset
[ ] Disaster recovery drills and chaos-testing a cluster
[ ] Contributing to or reading Qdrant's Rust source for deep internals (optional)