# Supported Features This page is the canonical human-readable inventory of features implemented in the pgContext 0.3.0 clean-install codebase. It is grounded in the installed SQL surface, the SQL lifecycle registry, the machine-checked capability contract, and its focused and lifecycle tests. It does not list roadmap-only work. Maturity is assigned per capability, not per release. The 0.3.0 version number acknowledges compatibility changes from 0.2; it does not promote every new retrieval or storage feature to Stable. The maturity labels mean: - **Supported:** a release platform covered by the current support policy. - **Stable:** covered by the public compatibility policy. - **Experimental:** implemented and testable, but its API, storage format, or operational contract may change before promotion. - **Internal:** a machine-checked implementation invariant rather than a public SQL compatibility promise. - **PostgreSQL-native:** supplied through PostgreSQL rather than a parallel pgContext subsystem. Each label applies to the named feature, not to every child feature it can compose. When one named surface spans different maturity levels, this page uses the least mature relevant label and explains the boundary. See the [SQL API contract](api_reference.md) for exact signatures and the [support policy](support_policy.md) for compatibility guarantees. ## Platform and data authority | Feature | Maturity | Description | |---|---|---| | PostgreSQL 17 and 18 | Supported | pgContext builds and runs against PostgreSQL 17 and 18. PostgreSQL 17 remains the primary benchmark and deep-lifecycle target. | | Linux release images | Supported | Release images target `linux/amd64` and `linux/arm64`; source builds remain available for supported PostgreSQL installations. | | PostgreSQL-native authority | PostgreSQL-native | Ordinary PostgreSQL rows remain authoritative. MVCC, transactions, ACLs, row-level security, WAL, replication, and PostgreSQL backup tools remain in force. | ## Vector representations | Feature | Maturity | Description | |---|---|---| | Dense `vector` SQL surface | Stable | `vector` supports dimensions, checked casts, L2, inner product, cosine, and L1 scoring, comparison operators, B-tree ordering, and sum/average aggregates. | | Half-precision vectors | Experimental | `halfvec` supports typmods, checked conversions, exact metrics, ordering, aggregates, and metric-specific HNSW opclasses. | | Sparse vectors | Experimental | `sparsevec` supports canonical sparse values, array and dense conversions, exact metrics, ordering, aggregates, and metric-specific HNSW opclasses. | | Bit vectors | Experimental | `bitvec` supports checked bit representations, Boolean and PostgreSQL bit conversions, Hamming and Jaccard distance, bitwise aggregates, and explicit HNSW opclasses. | | Provider-native integer vectors | Experimental | `int8vec` and `uint8vec` are authoritative signed and unsigned 8-bit source types with `1..=16000` dimensions, allocation-bounded text and PostgreSQL binary I/O, destination-typmod enforcement during text and binary COPY, profile-aware imports, array/dense conversions, exact wide-accumulator L2, raw and negative inner product, cosine, and L1 scoring, deterministic ordering, wide sum/average aggregates, and eight validated metric-specific HNSW opclasses. Runtime NEON or AVX2 kernels must match the scalar integer accumulator exactly; unsupported hardware uses the scalar oracle. | See [Dense vectors and exact search](vector_search.md) and the [metric and operator matrix](metric_operator_matrix.md) for exact score and ordering semantics. ## Collections and source management | Feature | Maturity | Description | |---|---|---| | Table-backed collections | Stable | A collection records metadata for an ordinary PostgreSQL source table without copying or taking ownership of its rows. | | Collection aliases | Stable | Aliases provide an atomic logical name that can be redirected to another registered collection. | | Collection resource limits | Stable | Per-collection limits bound dimensions, vectors, points, filter nodes, result size, candidate work, query time, and index memory. | | Named dense vectors | Stable | A collection can register multiple dense vector columns with independent names, dimensions, and metrics and select them by name during search. | | Per-vector index and quantization configuration | Experimental | pgContext validates and stores per-vector HNSW, quantization, and status metadata; full build-to-scan consumption is still being completed. | | Stable point mappings | Stable | Source keys map to stable logical point IDs. Deleting a mapping makes that point unavailable without deleting the authoritative source row. | | Bulk point maintenance | Stable | Bulk upsert, delete, and source-table backfill APIs validate typed ordered rowsets before mutation, execute one set-oriented statement per bounded chunk, preserve stable point IDs and input order, reject duplicates before mutation, and report progress. | | Registered structured filters | Stable | A bounded Boolean JSON grammar targets registered columns and JSONB paths, renders typed SQL predicates, and binds values through parameters. | | Registered payload mutations | Stable | Set, delete, and clear helpers update only registered source-table columns or JSONB paths and preserve PostgreSQL permissions and transactions. | | Profile-backed embedding migrations | Experimental | Migration records reference immutable source and target embedding profiles, track bounded progress and typed lifecycle states, and never perform provider inference in PostgreSQL. The mutable model-version registry was removed without a compatibility shim. | | Immutable embedding profiles | Experimental | A collection owner binds each provider/model/revision contract to one typed source column, its fixed typmod, and one live metric-matched HNSW index. Profiles record representation, dimensions, normalization, metric, input/output templates, configuration hash, optional integer scale/zero point, and explicit packed-binary bit/byte order, and cannot be updated in place. Profile-aware constructors reject representation, dimension, layout, stale-index, and ownership drift. Security-definer list/explain functions use a pinned search path and filter by `SESSION_USER`; explain reports the durable source/index binding and whether it is still valid. | See [Collections](collections.md), [Filters](filters.md), and [Multi-tenancy](multi_tenancy.md) for the detailed contracts. ## Retrieval and query execution | Feature | Maturity | Description | |---|---|---| | Canonical retrieval contracts | Internal | Core owns score ordering, representations, index/source identities, lifecycle reasons, and typed revisions; query owns an exhaustive leaf/source registry plus occurrence, generation, configuration, profile, source-version, rank, and fusion-contribution provenance through final results. Codec compatibility is validated through one exhaustive registry. Crate-boundary checks reject duplicate score, branch, and codec definitions. | | Exact vector search | Stable | Exact search scores explicit arrays or visible rows from registered tables and provides the correctness oracle for approximate paths. | | Exact-first readiness | Experimental | Dense-vector source registration exposes complete invoker-authoritative exact search immediately, then freezes explicit exact-only, HNSW, or IVFFlat plans. Fenced jobs return reviewed top-level concurrent-index DDL; publication accepts only a live structurally matching index. Exact serving remains available through build, cancellation, restart, and operational failure, while source identity drift fails closed. Both frozen ten-million-row lanes preserve bit-exact membership/scores, 100% top-10 recall, and exceed the 5,000-row/s build floor (PG17: 34,682; PG18: 33,253), but miss the building-query p95, indexed-query p95, and temp ceilings, so Stable promotion is a measured no-go. | | Sparse search | Experimental | Explicit sparse candidate arrays use exact scoring; registered sparse vectors use exact search or a validated metric-matched HNSW candidate path followed by authoritative reranking. | | Scroll | Stable | Keyset-style scrolling returns stable point-ID pages over active, visible mappings and accepts the shared filter grammar. | | Count and facet | Stable | Count and facet operations reuse the visible point and filter plan and apply deterministic missing-value and ordering rules. | | Grouped search | Stable | Exact dense results can be capped per registered group field with deterministic group and result ordering. | | Recommendation search | Stable | Positive and negative points or raw vectors form an exact query while deleted or unauthorized examples are rejected. | | Discovery and explore search | Stable | Visible context points form a centroid used for deterministic diversity-oriented ranking. | | PostgreSQL-native lexical retrieval | Stable | Registered lexical sources bind ordered weighted text or JSON-path fields, or a stored/generated `tsvector` column, to a resolved text-search configuration, ranker, normalization, rank weights, and optional per-row `tsquery`. Typed plain, structured, phrase, web-search, prefix, distance, Boolean, weight-restricted, and registered-`tsquery` forms compile to native constructors with bound values. Without an attached index the exact path evaluates the complete invoker-visible corpus or reports budget exhaustion; with a valid, live, non-partial GIN or GiST index it probes at most the remaining candidate allowance plus one, marks a crossed boundary incomplete, and rereads current source rows under MVCC and RLS before exact native reranking. An attachment must have exactly one key and no included columns, use PostgreSQL's `tsvector_ops`, and match either the registered stored column or the planner-proven canonical raw-field expression. Registration metadata is private and exposed only through membership-filtered security-barrier views; relation, column, type, collation, configuration, index-expression, opclass, and index-definition drift each fail closed. Bounded `ts_headline` hydration admits point count, option bytes, and source bytes before PostgreSQL builds markup, caps total returned output while reading the response, and requires caller-side output-context sanitization. | | Trigram fuzzy retrieval | Experimental | Registered fuzzy sources require a PostgreSQL `text` column and resolve optional `pg_trgm` through extension membership rather than `search_path`. Attached indexes must key that exact column directly and use the extension-owned `gin_trgm_ops` or `gist_trgm_ops`; column number, type, and collation drift fail closed. The path serves `similarity`, `word_similarity`, and `strict_word_similarity` with typed finite thresholds, restores each mode's own trigram threshold default through an RAII scope guard, with transaction-local `set_config` covering the error-unwind path, and recomputes the final score with the explicit similarity function. `pg_trgm` is not an install requirement. | | Automatic document chunking | Experimental | Immutable plain-text, Markdown, and HTML profiles use the deterministic `unicode_words_v1` tokenizer to produce source-linked original citation spans, separate retrieval/context text, bounded overlap, structure paths, parent/neighbor identities, content hashes, and deterministic fake embeddings. A transactional trigger or explicit bounded enqueue creates fenced lease/progress/failure/retry/cancel/supersede jobs. The external no-network worker emits a bounded versioned response that PostgreSQL reparses canonically before staging and rechecks under current source/projection OIDs, attnums, types, collation, MVCC, ACL, RLS, version, hash, and registration revision before one atomic current-generation flip. Raw staging remains private; current reads rehydrate the authoritative row; stale source identity hides old chunks immediately; and identical publication replay converges. The opt-in real-model adapter can embed published chunks with revision-pinned `all-MiniLM-L6-v2`, build a 384-dimension HNSW index, retrieve the expected occurrence, and preserve its citation object, but it is a promotion smoke rather than the production embedding-job integration. Both frozen one-million-row source-cardinality lanes miss the predeclared 1,000-chunks-per-second publication floor (PG17: 10.837; PG18: 33.325), so Stable promotion remains a measured no-go. | | Multi-model retrieval | Stable | `pgcontext.query_multi_model` validates one immutable, version-bound query branch per active or draining profile, applies the same typed filter and PostgreSQL ACL/RLS boundary to every branch, authoritatively rechecks current source/version/deletion state, and fuses only one-based ranks with configurable weighted RRF. Different representations, dimensions, and metrics may coexist on one stable point namespace; native scores remain diagnostics and are never compared across profiles. Candidate work, native scoring, and HNSW traversal use the canonical global executor budgets, including the sum across every attached child index of a partitioned source; a verified HNSW plan that reports no access-method work falls back only after the complete invoker-visible, version-current corpus fits the remaining comparison allowance, otherwise it fails closed. Extension-owned preparation, candidate identity, execution, and final-report allocations share one memory allowance, and the statement timeout remains armed through final JSON construction. The default elapsed allowance is 500 ms; an explicit collection `query_timeout_ms` may widen only this mixed-model path up to the global 60-second ceiling. Ties are deterministic, profile/index/configuration drift fails closed, and `require_all_profiles = false` is the only way to obtain an explicitly named degraded result. Coverage reports current, stale, and active counts only for invoker-visible source rows under current ACL/RLS. The frozen held-out contract declares equal `1:1` weights, eight queries over independent spaces, two 50-candidate fused branches, `k = 60`, and a 102-candidate allowance. It records raw A-only, B-only, fused, filtered-partial, and degraded quality/cost evidence plus an explicit pass/no-go decision. At one million rows, the unchanged contract passes on PostgreSQL 17 and 18: fused recall is no worse than the stronger single-profile baseline at the same declared global budget. | | Provider-neutral semantic reranking | Experimental | `prepare_semantic_rerank` releases at most 512 currently authorized candidate texts through a versioned 4 MiB envelope carrying request/model identity, expiry, logical occurrence and point IDs, source version and SHA-256 digest, fused rank/score, branch contributions, and allow-listed metadata. `finalize_semantic_rerank` accepts only finite unique in-set scores for the exact request/model revision, then rereads each row under current MVCC, point deletion, typed filter, ACL, and RLS and verifies source version and content digest before returning identity/provenance without text. Operational failure may use stored fused order only under `allow_fused_fallback` and remains visibly `degraded_reranker`; cancellation, malformed or injected output, permission loss, catalog drift, and source change always fail closed. Identical replay still rechecks current authority. The separate no-network `pgcontext-worker` uses a current-thread runtime, tracked blocking scorer, bounded retries/circuit breaker, immutable license/platform/token/time manifests, and digest-verified operator-provided artifacts; no weights are bundled. The private `linear_pair_v1` fixture passes the frozen PG17/PG18 in-server, backup/restore, and one-million-row contract locally. Separately, an opt-in source adapter downloads a revision-pinned Apache-2.0 `ms-marco-MiniLM-L6-v2` ONNX artifact outside the release payload and has passed a real PostgreSQL prepare-score-finalize smoke. It is not yet a general model backend inside `pgcontext-worker`. Stable promotion remains pending retained held-out quality, timeout/cancellation and authority-churn coverage, plus hosted Darwin and Linux arm64/x86_64 evidence. | | Adaptive-dimension retrieval | Experimental | An embedding profile may certify 1..=8 strictly ascending Matryoshka prefix dimensions below its full dimension for dense or half representations under l2, inner_product, or cosine. The stored vector is never truncated or rewritten. A bounded invoker-visible-corpus preflight starts prefix scoring only when candidate, comparison, recheck, transient-memory, and expansion budgets can fund exhaustive coverage. The first probe admits 4x the requested limit; if that is not exhaustive, one wider step admits the complete visible corpus using the next declared prefix when available. Every admitted row is authoritatively reread and reranked at full dimensions under current MVCC/RLS. If exhaustive coverage cannot be funded, pgContext selects full-vector exact search before prefix work; that exact scan may itself fail closed at the comparison or elapsed limit. Elapsed time is enforced during execution rather than predicted by preflight. `pgcontext.adaptive_prefix_dimensions` selects automatically (`0`), pins a declared prefix, or disables the path (`-1`); telemetry reports prefix, expansion count, and termination. Applies only to a single exact dense leaf; HNSW ignores certified prefixes and composite branches keep their bounded per-leaf allocation. **Performance no-go:** the scan-based prefix schedule performs more work than full-vector exact search in the PG17/PG18 gate, while corpora beyond the 10,000-candidate ceiling safely select exact fallback. No latency benefit is claimed. | | Dense plus lexical fusion | Stable | One dense branch and one registered lexical branch combine through deterministic reciprocal-rank fusion with final source visibility checks. | | Composite query execution | Stable | One typed executor composes the bundled dense, sparse, lexical, fuzzy, quantized, late-interaction, recommendation, discovery, lookup, prefetch, threshold, formula, and finalization stages. Prefetch supports parameterized RRF and weighted RRF without mixing incomparable raw profile scores. Candidate, comparison, filter, depth, node, stage, expansion, extension-owned transient-memory, hydrated-source-key, elapsed-time, and result budgets are global across the tree and never return silent partial points. PostgreSQL executor-internal SPI/sort memory remains PostgreSQL-governed; adapters bound admitted row sets before materializing Rust-owned responses. PostgreSQL applies the elapsed cap to SPI work with a current-statement timeout guard. Scan adapters admit and score the same invoker-visible row set, while lookup charges the bounded requested identities before returning only ACL/RLS-visible rows so hidden identities cannot be inferred from work accounting. SQL hybrid overloads delegate to this executor; adapters no longer own fusion or ranking policy. The transport-neutral topology-expansion and external-rerank IR/port contracts are stable, bounded, and fail closed. The rerank adapter compares the complete authorized source snapshot again after provider scoring, so source-version, text, metadata, identity, deletion, filter, ACL, or RLS drift cannot return a stale reranked row. `execute_query` does not yet bundle either provider; SQL execution of those stages remains unavailable until their later graph/external-provider phases. | | Multi-vector late interaction | Experimental | A registered `vector[]` source maintains collection-owned token rows and HNSW candidates, then returns exact MaxSim reranking against the current source row. | See [Retrieval methods](retrieval_methods.md), [Lexical retrieval](lexical_retrieval.md), [Automatic document chunking](automatic_chunking.md), [Hybrid retrieval](hybrid_retrieval.md), and [Dense vectors and exact search](vector_search.md). ## Indexing and accelerated serving | Feature | Maturity | Description | |---|---|---| | Page-native HNSW | Experimental | `pgcontext_hnsw` stores metric-bound graph pages for dense, half, sparse, bit, signed 8-bit, and unsigned 8-bit vectors and performs bounded candidate traversal without a silent exact fallback. Integer graph navigation uses a lossless dense view, while PostgreSQL rechecks every candidate with the authoritative integer source operator before final ordering. | | Lazy HNSW cursor | Internal | Graph-read HNSW uses one statement-local bounded cursor for eager and resumable traversal. It preserves deterministic order, masks, work accounting, cancellation, mapped/segmented/delta composition, and PostgreSQL exact source recheck without adding a public SQL or planner surface. | | Virtual beam engine | Internal | `context-query` owns a provider-neutral vector-only beam with a bounded parent arena, batched expansion, deterministic dominance, duplicate/cycle/width/hop pruning, separated score components, exact-result ordering, evidence-path reconstruction, cancellation, and typed incomplete outcomes. Topology identities are represented but rejected until the later topology phases; no SQL or planner surface is exposed. | | Filter-aware ANN | Experimental | Filtered search builds a bounded candidate mask, permits excluded graph nodes to remain routing connectors, and exact-rechecks visible source rows and predicates. | | Transactional HNSW maintenance | Experimental | Source inserts, updates, deletes, and VACUUM maintain graph, delta, and tombstone state while final queries recheck the current visible source row. | | Segmented HNSW serving and compaction | Experimental | HNSW indexes publish at most 16 immutable graph segments plus one exact start/end/generation/count-bounded active delta capped at 10,000 records. Page items and the active cursor commit in one Generic-WAL record; readers hold the metapage shared while copying mutable delta pages, preventing false count-corruption during concurrent appends. Full deltas rotate into generation-stamped segments; live rows move into ANN pages and only unresolved tombstones remain in frozen extents whose persisted record counts are verified. Scans read the overlay first, apply segment-specific retirement masks during traversal, and replay `segment graph → segment mutations` chronologically so both top-k refill and heap-TID reuse are correct. Directory saturation preflights memory and compacts only the smallest adjacent pair while relocating the active delta contiguously. VACUUM uses the same table→append lock order and rotates tombstones in configured bounded chunks. `enqueue_hnsw_compaction` captures a directory epoch and runs one retry-idempotent pair operation through the supervised lifecycle. `hnsw_segment_stats` reports fan-out, active records/blocks, immutable rows, frozen mutations, smallest-pair work, debt, parallel eligibility, and epoch. `hnsw_segment_parallel_workers` uses a reusable backend-local pure-Rust pool with cluster advisory admission, backend-side interrupt polling, cancellation broadcast, panic draining, and visible serial degradation; it conservatively preflights all segment extents against the serving budget before materialization and rechecks exact packed bytes before retaining each graph. PostgreSQL page, mapped-file, and shared-memory access remains backend-affine. Build repartitioning consumes owned graph state without cloning the full graph, while pair and full compaction enforce `maintenance_work_mem` before dangerous allocation. Mapped generation retirement is relfilenode-aware. The clean metapage and mapped identities reject older formats with a `REINDEX` requirement. | | Quantized HNSW serving | Stable | `pgcontext_hnsw` binds one codec-spec revision to the index, then trains a deterministic binary, scalar/SQ8, or product artifact for each immutable segment from at most 4,096 evenly distributed authoritative segment rows. Each versioned, checksummed segment artifact binds its trained revision, exact-source-rerank policy, codebook, and genuinely 16-byte-aligned fixed-stride code rows; a traversal never mixes artifacts inside one segment adapter. This segment-local lifecycle lets rotation and compaction replace bounded generations independently under the same index spec. Traversal prepares one static query scorer per artifact and reads codes without per-node allocation; PostgreSQL marks every quantized order-by result for exact source-operator rerank using the certified SQL result type. Unsupported bitvec Hamming/Jaccard quantization and incompatible PQ dimensions fail at build/first insert with SQLSTATE `22023`. Unknown, corrupt, mixed-dimension, mixed-row-count, or old quantized formats fail closed and require rebuilding. First-use packing conservatively preflights decoded containers, final arrays, publication scratch, codec state, and training memory before segment allocation; external publication occurs only from a committed scan. Caller-supplied PQ codebooks are no longer accepted. The frozen release-mode 1M workload passed its recall, warm-latency, aggregate packed-memory, size, publication, VACUUM/REINDEX, and restart-recovery gates for every codec on PG17.10 and PG18.4; see [the retained benchmark](../benchmarks/quantized_hnsw_1m.md). | | Native IVFFlat | Experimental | `pgcontext_ivfflat` provides page-native, WAL-logged IVF generations for dense `vector`, `halfvec`, signed `int8vec`, unsigned `uint8vec`, and `bitvec`. Metric-bound opclasses cover L2, negative inner product, cosine, and L1 for continuous vectors plus Hamming and Jaccard for bit vectors. `lists`, bounded probes, scan-global candidate budgets, strict bounded-frontier ordering, lazy relaxed widening with explicitly approximate within/across-batch order, deterministic memory-bounded external construction and spill fan-in, native PostgreSQL parallel assignment workers, DML deltas, VACUUM tombstones, synchronous or supervised compaction, automatic compaction debt at 10,000 delta records, generation retirement, REINDEX, CIC, partitions, dump/restore, crash replay, physical replication, and standby promotion are exercised on PG17/18. Manual compactors require index ownership or `MAINTAIN` on the source table, decode `regclass` without opening the index, establish table-before-index locks, and serialize directly on the index generation lock; concurrent DML and concurrent-compactor regressions run with a 200 ms deadlock detector. SQ8 and PQ postings reuse the shared codec artifact and always request exact source-operator reranking; binary metrics reject those dense codecs. `ivfflat_index_info` verifies every page/checksum, codebook binding, posting, and generation and reports the build worker count, while `ivfflat_last_scan_work` reports actual visited, reranked, and widening work. The clean v4 format intentionally rejects earlier experimental pages and requires `REINDEX`. Release-scale 1M/10M comparative certification remains open. | | Rebuildable mapped graph generations | Experimental | Checksummed immutable graph artifacts support validation, bounded read-only mapping, reader pins, publication, retirement, cleanup, and source-row exact rechecks. | | Shared generation-build primitives | Internal | One pure contract owns typed jobs and artifacts, monotonic checkpoints, leases and takeover, validation, publication aliases, versioned payload-checksummed manifests, reader pins, retirement, bounded spill runs, deterministic sampling/k-means, and redacted structural findings for later HNSW, IVF, projection, graph, and certification features. Certification fixtures retain authoritative vector seeds, true 100-dimensional counter expansion, independently balanced tenant filters, and deterministic exact top-k IDs and scores for the 100k, 1M, and 10M tiers. | The non-dense HNSW opclass names are stable SQL contracts, but their vector types and the HNSW access-method storage lifecycle remain experimental. See [Indexes](indexes.md) and [Rebuildable storage artifacts](storage.md). ## Operations and migration | Feature | Maturity | Description | |---|---|---| | Index diagnostics and advice | Stable | Typed functions report index readiness, corruption, memory estimates, optimization state, vacuum advice, recall checks, build/serving counters, segmented-HNSW fan-out and compaction debt, parallel admission/degradation, and index recommendations. | | Automatic query telemetry | Stable | Executor-backed queries offer bounded events to a PostgreSQL background worker that stores membership-filtered strategy, work, lifecycle, completion, and latency rollups without query contents. | | Supervised generation jobs | Experimental | Owner-scoped durable jobs support planned/running/validating/publishing states, attempt-and-canonical-backend fencing, bounded leases, crash takeover, idempotent checkpoints, single-transition retries, serialized cooperative cancellation through validation, post-commit wake-up, claim-boundary and active source-delta reconciliation, payload-integrity-checked certification evidence, timezone-independent reader pins, and serialized atomic publication. Registered executors cover certification evidence and one-pair HNSW compaction bound to an owned collection's authoritative source-table index; the worker revalidates the index binding before execution and validation. A worker keeps its fenced ownership for the whole transaction even when validation or alias locking outlives the lease TTL; the lease governs takeover only between transactions. Unsupported job kinds fail closed. Disabled or saturated workers leave retryable planned work and never remove exact retrieval. Native `CREATE INDEX` is not made resumable. | | Artifact diagnostics and cleanup | Experimental | Operators can validate, inspect, retire, and clean root-confined generated artifact files without treating them as authoritative backups. | | pgvector migration and compatibility | Stable | The main extension publishes an executable compatibility matrix and can create dump-visible, extension-owner-owned pgvector 0.8.x casts/opclasses plus conflict-safe `hnsw`/`ivfflat` name facades for canonical pgContext types when those access-method names are free. HNSW and IVFFlat ownership conversion is resumable, preserves certified options such as `lists`, and supports exact validation and rollback. Unsupported aliases and expression shapes fail explicitly. | | pgvector coexistence and index comparison | Experimental | Read-only reports and comparisons inspect pgvector columns and ANN indexes on PostgreSQL 17 and 18 while both extensions remain installed. The retired companion extension is no longer packaged. | | pgvector adoption and ownership conversion | Experimental | Fail-closed dry runs and resumable fast or restricted-online workflows convert certified dense and sparse columns. HNSW rebuilds on `pgcontext_hnsw`; IVFFlat rebuilds on `pgcontext_ivfflat` with `lists` preserved. Exact/ANN validation, rollback to untouched pgvector objects, finalization, session-drain attestations, and dump/restore are exercised on PG17/18. | | PostgreSQL backup, recovery, and maintenance | PostgreSQL-native | Source tables use normal `VACUUM`, `ANALYZE`, `REINDEX`, `pg_dump`, physical backups, WAL recovery, and replication procedures. Derived artifacts can be rebuilt. | See [Operations and support](operations.md), [pgvector coexistence](pgvector_coexist.md), and [Migrating from pgvector](pgvector_migration.md). ## What this page excludes This inventory excludes planned-only features. A feature moves here only when the code, installed surface, focused test, lifecycle evidence, maturity classification, and detailed user documentation agree. Future work remains in the [roadmap](roadmap.md).