Vector limitations

What vector search does not do in v0.1.0 — experimental types, fallbacks, and unsupported paths.

Version
Latest
v0.1.0 · latest 3 min read
On this page
  1. Experimental: sparse and multivector
  2. Limited: binary vectors
  3. Limited: other index methods
  4. Unsupported
  5. Durability notes

Honest boundaries for v0.1.0. Anything listed here as limited, experimental, or unsupported may improve in later versions; check the versioned section of a newer release rather than assuming.

Experimental: sparse and multivector#

  • SPARSE_VECTOR(element, dim) and MULTI_VECTOR(element, max_vectors, dim) parse in DDL but have no verified literal or insert path through SQL. Creating such a column succeeds; populating it through documented SQL does not. Status: Experimental — do not build on these types.
  • Sparse inverted and multivector index machinery exists, but without a SQL write path there is nothing to index from ordinary clients.

Limited: binary vectors#

  • VECTOR(BINARY, n) DDL, '0101…' string literals, and Hamming (<~>) / Jaccard (<%>) queries are supported.
  • Binary ANN indexes are declared but queries fall back to exact computation — correct results, no acceleration.
  • Numeric operators (<->, <=>, <#>) on binary columns are rejected, and bit operators on numeric columns are rejected. The errors name the mismatch.

Limited: other index methods#

  • ivf_flat, ivf_sq, ivf_pq, diskann, and flat index declarations are accepted. Small scans may still execute exactly (the planner says so in EXPLAIN), and ivf_sq/ivf_pq promote some quantized element types through float rather than natively. INT4 and binary elements are rejected by ivf_pq.
  • Hybrid ranking (reciprocal-rank fusion / weighted combination of vector + full-text scores) exists only as internal machinery with no SQL surface in v0.1.0. Implement hybrid retrieval in your application (two queries, one merge) instead.

Unsupported#

  • GPU execution. There is no working GPU backend in v0.1.0; compute settings requesting GPU degrade to CPU. Do not plan capacity around hardware acceleration.
  • CREATE INDEX ... USING hnsw inside an explicit transaction. Refused with a descriptive error; create indexes in autocommit.
  • Distributed / sharded vector search. Single-node only; multi-region and distributed search are roadmap items.
  • Native embedding generation. PLOMID stores and searches vectors; it does not produce them. Bring your own model.
  • Composite and UNIQUE vector indexes. Rejected; one indexed vector column per index, non-unique.
  • Vector index on an expression. Index a stored column, not an expression.

Durability notes#

Vector data participates in the same cold-backup / restore workflow as relational data, and post-restore vector queries have been verified to return correct results. The derived ANN cache is excluded from backups by design and rebuilt from rows on reopen. See Backup and restore.

Was this page helpful?