Contents

• Migrating from pgvector
  • • Upgrading a pgvector-first 0.1 install
  • Dense Vectors
  • Converting Column Ownership
  • Filters and Hybrid Retrieval
  • Indexes
  • Current Gaps

Migrating from pgvector

pgContext supports an incremental coexistence workflow for existing pgvector databases. The two main extensions can be installed in either order because pgvector owns public.* types while pgContext owns canonical pgcontext.* types. Keep an existing pgvector column in place and install the certified pgcontext_pgvector companion bridge before building a pgcontext_hnsw index over it. The bridge profile is PostgreSQL 17 with pgContext 0.2.0 and pgvector 0.8.x installed in public. Dense vector and halfvec layouts are byte-certified. Existing public.sparsevec columns can be indexed directly, and ownership conversion is available through a validated restricted-online rewrite because the sparse physical layouts differ. See Trying pgContext on an Existing pgvector Database for the live workflow and inventory tools.

Upgrading a pgvector-first 0.1 install

The 0.1→0.2 extension update deliberately refuses this legacy layout before mutation because public.vector belongs to pgvector. Export pgContext collection/vector/filter registrations and inventory every object depending on the old pgContext extension before any DROP EXTENSION ... CASCADE; CASCADE can remove application views/functions as well as indexes. Then install pgContext 0.2 and pgcontext_pgvector, recreate the registrations and dependent objects, and rebuild pgcontext_hnsw indexes over the original unchanged pgvector columns. The upgrade preflight never rewrites or retypes those columns.

Explicit pgContext HNSW opclasses cover half and sparse L2, inner product, cosine, and L1, plus bit Hamming and Jaccard. The names and metric bindings are stable, while the variant SQL types and HNSW on-disk format remain experimental; review the index-specific single-page dimension envelope before rebuilding large pgvector indexes.

Dense Vectors

Existing pgContext-owned vector columns can be registered as named collection vectors:

SELECT pgcontext.create_collection('docs', 'public.docs');
SELECT pgcontext.register_vector('docs', 'embedding', 'embedding', 1536, 'cosine');

Dense vector(n) typmods, text, casts from numeric arrays, distance functions, distance operators, comparison operators, and dense vector aggregates are implemented with pgvector-compatible behavior. Assignments to dimensioned columns reject mismatches with SQLSTATE 22023. Intentional differences are documented with tests.

An existing column owned by the pgvector extension can be indexed and registered through the companion bridge. Run pgcontext.migration_report() first: it verifies the type owner and reports defaults, arrays, generated columns, partitions, dependent views, and complex indexes that must be handled before an ownership cutover. Index adoption never changes the column type.

Converting Column Ownership

Install the certified bridge, inventory the target, and choose one of two fail-closed modes:

SELECT *
FROM pgcontext.start_pgvector_ownership_conversion(
    'public.items'::regclass,
    'embedding',
    'fast',
    'cosine',
    application_dependencies_reviewed => true
);

SELECT *
FROM pgcontext.run_pgvector_ownership_conversion(
    1,
    sessions_drained => true
);

Fast mode takes ACCESS EXCLUSIVE, refuses named prepared statements in the calling backend, changes a certified public.vector/public.halfvec column to the corresponding pgContext-owned type without rewriting the heap, and rebuilds certified pgvector HNSW or IVFFlat indexes as pgcontext_hnsw. Dimensioned sources become an unmodified canonical base type plus a validated dimension CHECK constraint; NOT NULL, values, and index options/tablespace are preserved when the target AM can represent them. Because pgcontext_hnsw does not currently expose pgvector’s per-index HNSW reloptions, a source HNSW index with nondefault options is refused rather than silently changed. IVFFlat lists is intentionally not translated when that access method is rebuilt as HNSW. Invalid source indexes and indexes with comments are also refused. The caller must retain CREATE on the table schema and on any preserved nondefault tablespace needed to rebuild an index. The operation is one transaction.

Fast mode deliberately rejects public.sparsevec: its packed pgvector layout is not binary-compatible with pgcontext.sparsevec, so a metadata-only type swap would corrupt values. Use restricted-online mode for sparsevec. The bridge decodes and validates pgvector’s packed indices and values during backfill and same-transaction dual writes; conversion fails closed for malformed data or dimensions above pgContext’s 16,000-dimension limit.

Restricted-online mode is for the narrow supported profile when the long lock is unacceptable:

SELECT *
FROM pgcontext.start_pgvector_ownership_conversion(
    'public.items'::regclass,
    'embedding',
    'restricted_online',
    'cosine',
    application_uses_column_lists => true,
    application_dependencies_reviewed => true
);

-- Repeat in separate transactions until status = 'index_pending'.
SELECT * FROM pgcontext.run_pgvector_ownership_conversion(1, 1000);

-- Execute the returned next_command as a top-level statement, then certify it.
CREATE INDEX CONCURRENTLY ...;
SELECT * FROM pgcontext.run_pgvector_ownership_conversion(1);

-- Drain/recycle application sessions before the short locked swap.
SELECT * FROM pgcontext.cutover_pgvector_ownership_conversion(
    1,
    sessions_drained => true
);

The shadow trigger runs in the same source DML transaction and overwrites direct shadow assignments from the authoritative column. Backfill calls persist a heap-TID range cursor and examine a bounded range; an authoritative full scan is reserved for the end of a pass and resets the cursor if concurrent locks or drift left mismatches behind. That scan runs without the cutover’s ACCESS EXCLUSIVE lock; the certified trigger preserves equality until the short lock upgrade freezes DML. Candidate index construction is deliberately emitted to the caller because PostgreSQL forbids CREATE INDEX CONCURRENTLY inside a function transaction. After cutover, the trigger maintains the old pgvector column for rollback. Use rollback_pgvector_ownership_conversion(1) to restore the original column and indexes, or finalize_pgvector_ownership_conversion(1) to validate once more and irreversibly remove the rollback column.

The caller that executes next_command must own the table and have CREATE on its schema. Final validation, like cutover validation, scans while the reverse trigger is active under ACCESS SHARE; only the final trigger/column DDL uses the upgraded exclusive lock.

The release gate exercises vector and halfvec conversions for L2, inner product, cosine, and L1, plus sparsevec restricted-online conversion and same-transaction writes on both sides of cutover. It compares exact distances before and after each conversion, terminates a backend between bounded online batches and resumes from the persisted cursor, validates rollback to untouched pgvector objects, drops both the bridge and pgvector after finalization, and restores a custom format dump into a clean database. A pgvector-derived pg_regress profile also keeps the pgvector-owned columns and query operators unchanged while replacing only the HNSW access method and opclass. Run the live gates with:

scripts/check-pgvector-ownership-conversion.sh
scripts/check-pgvector-regression-compat.sh

The regression profile is deliberately bounded to the supported PostgreSQL 17 HNSW migration contract; it does not claim IVFFlat implementation or pgvector’s HNSW-specific GUC surface.

Online mode adds a physical column, so applications must use explicit INSERT column lists throughout the migration. PostgreSQL cannot inventory prepared SQL in other backends; the sessions_drained value is an operator attestation, not automatic global detection. PostgreSQL also does not record column dependencies for application SQL or ordinary string-bodied SQL/PLpgSQL functions, so application_dependencies_reviewed => true is a required operator attestation that those call sites were inventoried and can accept the type-ownership change. The conversion refuses catalog-discoverable unsupported dependencies including RLS, comments, custom column statistics/storage, and unsupported index options rather than attempting partial rewrites. Arrays/domains, partitions, and composite-row dependencies remain unsupported.

Filters and Hybrid Retrieval

Register payload columns and JSONB paths that should be filterable:

SELECT pgcontext.register_filter_column('docs', 'tenant_id', 'tenant_id');
SELECT pgcontext.register_jsonb_path('docs', 'topic', 'metadata', ARRAY['topic']);

Filters are Qdrant-style JSON objects that render through typed SQL and SPI parameters. Full-text hybrid retrieval can combine a registered dense vector with a text column through reciprocal rank fusion.

Indexes

Exact search is the correctness baseline. Keep existing PostgreSQL indexes for high-cardinality filters, joins, and partitioning. Add pgContext index paths only after recall checks and operational diagnostics show that approximate retrieval is appropriate for the workload.

pgContext does not implement pgvector IVFFlat indexes for the first production surface. The production serving path is exact table-backed search first, with pgcontext_hnsw maturing behind explicit recall, visibility, filter, and restart gates. IVFFlat’s training/list maintenance model is not the selected artifact shape for pgContext’s PostgreSQL-native source-table ownership model. Applications that depend on IVFFlat during bind-mode evaluation should keep those pgvector indexes in place for that workload, and register the same source tables with pgContext for exact search, filters, hybrid retrieval, diagnostics, and HNSW evaluation. pgcontext.adopt_pgvector() and fast ownership conversion inventory IVFFlat and emit or execute a rebuild-as-HNSW plan; pgContext does not translate IVFFlat options or claim an IVFFlat implementation.

Current Gaps

Experimental SQL wrappers exist for halfvec, sparsevec, and pgContext’s bitvec bit-vector type. They support text input/output, dimension helpers, exact distance helpers, and distance operators, and they reject malformed values through the same core validators used by Rust code. halfvec also supports explicit-only numeric-array casts that round to half precision, halfvec(n) typmods, and sum/average aggregates. sparsevec also supports sparsevec(n) typmods, a structured constructor from aligned integer[] indexes and real[] values plus canonical index/value accessors, dense real[]/vector casts, and sum/average aggregates. Experimental pgcontext.search_sparse provides exact top-k over explicit sparse candidate arrays and registered sparse source columns. bitvec also supports bitvec(n) typmods, boolean[] casts for structured SQL construction and extraction, casts from PostgreSQL bit and bit varying, and casts back to PostgreSQL bit and bit varying. Pgvector-compatible built-in bit Hamming and Jaccard functions plus <~> and <%> operator overloads delegate through the same checked bitvec path. bitvec also supports bitwise OR/AND aggregates through pgcontext.bit_or(bitvec) and pgcontext.bit_and(bitvec). The variant types also install default btree ordering opclasses for deterministic comparison and ordinary PostgreSQL btree indexes.

PgContext installs first-class HNSW opclasses for halfvec and sparsevec L2, inner product, cosine, and L1, plus bitvec Hamming and Jaccard. These classes store dense graph payloads but bind traversal and SQL ordering to the selected metric. Bitvec remains explicit—choose pgcontext.bitvec_hnsw_hamming_ops or pgcontext.bitvec_hnsw_jaccard_ops; a default pgcontext_hnsw attempt still fails with SQLSTATE 42704 rather than guessing a bit metric. Quantized candidate generation, sparse exact array search, and exact reranking are available from SQL as experimental APIs while serving-path integration continues to mature.