fix: Update GraphRAG hardware sizing for single-store vector indexes

[katarinasupe]

What

Rewrites the Hardware sizing section of Memgraph in GraphRAG use cases.

The section said Memgraph stores embeddings in two places (8 B in the property store + 4 B in USearch = 12 B per float) and listed reference-based indexing and float16 as upcoming. Both have shipped:

• Single-store vector indexes on nodes (v3.8.0) and edges (v3.10.0): the vector lives only in USearch, the property store keeps a reference.
• scalar_kind (f16, bf16, …) on vector indexes.

This page now contradicts the vector search page, which documents the single-store pattern.

Changes

• Says that each indexed vector is stored once, for node and edge indexes.
• New formula: count × dimension × bytes per component, with bytes taken from the scalar kind, plus a worked example.
• Lists the memory costs that remain: index structure, reserved capacity (not freed on delete, released on drop), overlapping indexes, and the property-store growth when an index is dropped.
• Points to vector_index_memory_tracked and the existing Monitor vector index memory section, which also covers the ai_platform license.
• Removes the Upcoming optimizations subsection and adds an f16 tip to Best practices.

The other best practices are left as they are.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
documentation	Ready	Preview	Oct 1, 2026 7:43am UTC