Cosmos DB DiskANN vs pgvector: When to Use Which for AI-200

The core difference in one sentence

Cosmos DB DiskANN is a globally distributed, schema-less NoSQL store with vector indexing built in — optimised for high-throughput key lookups and low-latency vector retrieval at internet scale. pgvector is a PostgreSQL extension that adds vector types and indexes to a relational database — optimised for combining vector similarity with rich SQL queries (joins, filters, full-text search).

Cosmos DB DiskANN — what you need to know

DiskANN is Microsoft Research's disk-based Approximate Nearest Neighbour algorithm, integrated natively into the Cosmos DB NoSQL API. Key facts for the exam:

• Vector policy is defined at container creation and cannot be changed retroactively. It specifies the path (e.g., /embedding), dimensions (e.g., 1536), distance metric, and index type.
• Distance metrics supported: cosine, dotProduct, euclidean. Cosine is standard for text embeddings (e.g., Ada-002, text-embedding-3-small).
• Index types: flat (exact, small datasets) and DiskANN (approximate, production scale).
• Partitioning matters: vector queries that include the partition key execute within a single partition (efficient). Without the partition key, the query fans out to all partitions (expensive RU cost).
• Scale: designed for billions of vectors globally with multi-region replication. Strong at low-latency point reads combined with vector search.
• Limitation: no joins, no relational model, no full-text search natively (though Change Feed + Azure AI Search can complement it).

pgvector — what you need to know

pgvector is a PostgreSQL extension that adds the vector type and ANN indexes. On Azure, it runs on Azure Database for PostgreSQL Flexible Server. Key facts:

• IVFFlat: inverted file index. Faster to build, lower memory. Tune with lists (typically sqrt of row count). Adjust recall at query time with SET ivfflat.probes.
• HNSW: Hierarchical Navigable Small World. Slower to build, higher memory, better recall at query time. Tune with m (max connections per layer) and ef_construction. Adjust at query time with SET hnsw.ef_search.
• Distance operators: <-> (L2/Euclidean), <=> (cosine), <#> (negative inner product).
• Hybrid search: combine pgvector similarity with tsvector full-text search in a single query — a major advantage over pure vector databases.
• HALFVEC: half-precision (FP16) storage — halves memory for embeddings with minimal quality loss.
• Pre-filtering: add a WHERE category = 'X' clause before ORDER BY — pgvector HNSW supports filtered ANN search.

Side-by-side comparison

How to choose in an exam scenario

The AI-200 exam gives you a scenario and asks which storage solution to recommend. Use these decision rules:

• Global distribution + low latency from multiple regions: Cosmos DB DiskANN
• Semantic search combined with joins, relational filters, or full-text: PostgreSQL + pgvector
• Existing PostgreSQL application adding vector capabilities: pgvector (no new service)
• Schema-less documents with metadata + embedding stored together: Cosmos DB
• Event-driven processing triggered by new embeddings: Cosmos DB (Change Feed → Azure Functions)
• Cost-sensitive, moderate scale, complex analytics: PostgreSQL + pgvector