0.9.4 — 2026-08-13

Store

• Bound hot collection caches (P1-22)
• Preflight instance archive restores (P1-65)
• Reject incomplete archive collections (P1-65)
• Validate archive metadata databases (P1-65)
• Reject empty archive catalogs (P1-65)
• Reject uninitialized archive catalogs (P1-65)
• Keep archive export independent of current routing (P1-65)
• Validate archive metadata schema (P1-65)
• Validate archive metadata columns (P1-65)
• Preserve list metadata filter parity (P1-35)

Documentation

• Repair P3-18 roadmap row (B21)
• Reconcile P1-52 roadmap status (P1-52)
• Split hybrid search roadmap work
• Close Elixir public API parity (P1-67)
• Document archive runtime requirements (P1-65)
• Record archive restore integrity (P1-65)
• Add Elixir client quick start
• Clarify SDK and client repository roles
• Add v0.9.4 release announcement

Bug Fixes

• Preserve list-valued metadata filtering (B22)
• Reject SQLite archive sidecars (P1-65)
• Reject malformed archive restores (P1-65)
• Reject inconsistent archive indexes (P1-65)
• Reject wrong-metric archive indexes (P1-65)
• Reject archives with incomplete catalog schema (P1-65)
• Reject wrong-dimension archive indexes (P1-65)
• Reject wrong-dimension archive indexes for openai/ollama (P1-65)

Log

• Add product-signal metrics (P1-50)
• Cache product gauge refresh (P1-50)
• Bound search metric cardinality (P1-50)
• Restore bounded search metric labels (P1-50)
• Cache query-log gauge scans (P1-50)
• Persist worker metrics snapshots (P1-50)
• Normalize worker metrics persistence (P1-50)

Infrastructure

• Backfill and harden issue-key sync
• Add chaos failure issue triage (P1-52)
• [fix] Fail on malformed roadmap rows (B21)
• Pick the interpreter from the declared Python range
• Accept reconcile CLI mode (B23)

Performance

• Add chaos correctness harness (P1-52)
• Finish chaos nightly invariants (P1-52)
• Accept controlled ingest shedding in chaos
• Exercise both collection move routes
• Show concurrency in stress results

Testing

• Guard metrics endpoints from gauge refreshes (P1-50)
• Cover remaining product metric signals (P1-50)
• [infra] Chaos recipe var regression + xeon runner tag (P1-52)
• Cover incomplete archive rollback (P1-65)

API

• Remove unused slugify helper (P1-32)
• Forward shared raw-vector queries (B24)

Build

• Add HTTPX2 test client dependency

CLI

• Describe shared metrics reset (P1-50)

Plan

• Split ANN roadmap into flat/HNSW/IVF metrics (P3-54/55/56)

0.9.3.1 — 2026-08-02

Packaging

• Ship P1-46 distribution documentation

0.9.3 — 2026-07-30

Bug Fixes

• Return 413 for oversized ingest (R2-07)
• [core] Fix CSV column loss on blank/dup headers; collapse meta keys
• Reject invalid collection names on ingest (B2)
• Reject overlong ingest docids (B6)
• Reject non-object ingest metadata (B7)
• Fail fast on macOS direct SBERT with FAISS (B3)
• Keep catalog metrics off request path (B4)
• Accept case-insensitive Bearer scheme in static auth (B10)
• Fix dev loopback coercion in enforce_policy (B9)
• Harden auth/config startup validation (R4-70, R4-71)
• Break circular import via lazy LocalStore export (B11)
• Map encode outages to 503 (B12)
• Let filters import standalone (B11)
• Defer negated metadata filters (B14)
• Import embedder config from its owner (B11)
• Preserve BYOV vectors across restart (B13)
• Hold search slots through timeouts (B16)
• Bootstrap admin key in workers (B17)
• Preserve collection config on catalog recovery (B15)
• Guard query cache identity (R3-77)
• Keep cancelled search slots occupied (B16)
• Backfill legacy collection config sidecars (B15)

Documentation

• Document GitLab MR workflow
• Fix README auth key (R1-01)
• Refresh compose and env examples (R1-02, R2-08, R2-09, R2-10)
• Fix embedder_label example to colon form (R4-38)
• List all reference pages in README, prettify links
• Document embedder..* global fallback keys (R4-44)
• Mark release reproducibility complete (P3-40)
• Clarify legacy env mapping was temporary (R4-37)
• Regen OpenAPI snapshot: batch 413 response (B8)
• Fix embedder and CSV docstring drift (R4-76)
• Align embedder label examples (R4-38)
• Mark rebrand plan historical
• Remove stale bench-stress findings note
• Clarify benchmark artifact storage (R6-59)
• Explain Docker volume persistence (B20)
• Correct CLI selector order (R2-79)
• Forward Docker auth key (B20)

Core

• Register local client provider
• Share local client surface with client package
• Point local client surface at Python SDK
• Honor include_common search opt-in
• Commit schema migration explicitly (R3-22)
• Replace config in place so child views stay live (R3-15)
• Order multi-collection write locks by key, not id (R3-17)
• Close orphaned embedder worker on cache eviction (R3-16)
• Re-encode search query if embedder swapped mid-flight (R3-23)
• Validate native embedder device; fail fast on unavailable/invalid (B1)
• Fix instance path defaults (B5)
• Return not-found for vanished collection updates (R3-36)
• Support string not-equal filters (R2-74)
• Add identity (raw-vector/BYOV) embedder

Testing

• Align CLI query listing tests (R2-06)
• Drop unused fake embeddings helper (R4-48)
• Prune redundant API coverage
• Remove leftover test noise (R4-50)
• Release timed-out search worker instead of sleeping 5s
• Remove misnamed dead tests/init.py (R4-51)
• Cover close() reader-drain path (R4-52)
• Reproduce close() racing an in-flight read (R4-46)
• Drop duplicate CSV ingest API tests (R4-47)
• Trim duplicate auth status tests (R4-45)
• Cover batch metadata validation (B7)
• Cover single-doc ingest encode lock scope (R3-18)
• Assert canonical negation filtering (B14)
• Close pytest resources deterministically (R3-80)

API

• Share current_store dependency (R3-63)
• Map update_collection not_found structurally, not by message (R3-36)
• Validate name and report not_found on delete_collection (R3-28)
• Reject metadata-only multipart ingest (R1-67)
• Reject invalid POST search k values (R2-68)
• Gate health metrics behind admin auth (R4-69)
• Enforce batch ingest size limit (B8)
• Add tenant-scoped collection move route (R3-30)
• Classify invalid document ingest errors (R1-73)
• Add raw-vector fields to ingest and search schemas (BYOV chunk 2)
• Gate BYOV: identity vectors-only, embedding dim validation (BYOV chunk 3)
• Validate BYOV dimensions (P1-64)
• Validate raw-vector document metadata (P1-64)

Infrastructure

• Prepare release mirror state
• Add GitLab issue sync
• Remove config import fallbacks (R1-61)
• Source issue keys from commits only, not the MR description
• Drop phantom PAVEDB_CONFIG from Docker image (R4-39)
• Bench: cache HF model + 120s startup timeout on cold runners
• Sync B/F/G issue keys, not just R/P
• Auto-retry jobs on runner/system failures (max 2)
• Retry source checkout fetches
• Allow reserved review issue suffixes
• Warn on reserved review issue suffix gaps (R4-66)

Build

• Regenerate third-party licenses from runtime closure (P3-40)
• Publish runtime pip-freeze snapshot on release tags (P3-40)
• Build only the CPU image in CI; rename torch-wheel selector (P3-40)
• Tighten requirement pins after dependency due-diligence
• [cli] Drop source-tree shell wrappers (R4-65)
• Copy pyproject metadata in Dockerfile (R4-42)
• Remove invalid Torch floor (R6-58)
• Copy project license into Docker image (R3-78)

Store

• Drop dead bare-list handling from search output (R4-55)
• Remove unreachable async _flush_caches close path (R3-12)
• Snapshot embedder caches before orphan scan (R3-16)
• Use concrete FAISS backend in LocalStore (R4-53)
• Drop custom SearchOutput equality (R4-55)
• Release catalog guard before bootstrap (R3-13)
• Drain active readers before close (R4-52)
• Keep filtered metadata on one snapshot (R6-56)

Config

• Parse list-valued env overrides (R2-11)
• Drop .gitignore un-ignore for nonexistent example (R4-40)
• Package favicon.ico in pave.assets (R6-57)
• Drop dead scripts/*.sh include from MANIFEST.in (R4-41)
• Align pave.toml Python version with setup.py (R4-43)
• Ignore local tenants file (R4-40)
• Reject invalid text chunk overlap (R3-75)

Packaging

• Remove archived patchvec shim (R3-64)
• Remove qdrant stub (R1-60, R4-54)
• Replace BaseStore ABC with protocol (R2-62)
• Sanitize default docids (R2-05, R4-27)
• Move ingest size limit into service (R2-07)
• Drop unused imports stale from relocated archive logic (R4-34)
• Rename pave.toml to pyproject.toml so its tables are discovered (R4-42)

Log

• Harden metrics persistence (R1-03, R2-04)
• Assign ops-log handle before dest in configure (R3-19)
• Drain executors before flushing metrics on shutdown (R3-24)
• Drop redundant lock around stdout dest assignment
• Split query-log writer from document writes (R3-20)

SDK

• Scope SDK get_query/replay to tenant+collection (R3-33)
• Return the move envelope from SDK _rename (R3-32)
• Map no_text_extracted to InvalidRequest in the SDK (R3-31)
• Align local rename return contract (R3-32)
• Reach BYOV search/add through the public client (B18)

CLI

• Reject non-object –filters early (R3-26)
• Include vector_space_key in CLI collection listing (R3-29)
• Validate collection name on create, drop auto-slug (R3-35)
• Return JSON errors for malformed options (B19)

Performance

• Stabilize local vector stress tails
• Walk catalog metrics outside the metrics lock (R3-21)
• Remove historical benchmark dumps (R6-59)

Plan

• Record v0.9.3 final review

Refactor

• Unify tenant cap logic via tenant_limit_gate (R4-72)

UI

• Refresh app icon artwork and regenerate favicon

0.9.2 — 2026-06-20

Core

• Add batch document ingest (P1-16)
• Add JSON document ingest (P1-16)
• Single-source config schema in pave/config package (P1-53c)
• Centralize branding in version.py; promote CLI runtime flags to global
• [conf] Add embedder instance registry (P1-58)
• [fix] Stop spurious force-exit, especially under –reload
• Add embedder inventory endpoints (P1-59)

Log

• Create log parent dirs and disable cleanly on open failure
• Show instance name in startup banner and reflow config fields
• Log store SEARCH-IN at debug level
• [fix] Route uvicorn access log to file via log_config
• Add HTTP request metrics middleware (P1-49)
• Polish startup banner and warmup logging
• Preserve public v1 paths in HTTP metrics

Documentation

• Add generated CLI reference docs (P1-53b)
• [core] Document store protocol contract (P3-26)
• [core] Add module and class docstrings on public seams (P1-53d)(P3-26)
• [build] Add OpenAPI contract snapshot (P1-53e)

Config

• Tweak production-ready defaults (static auth, file logs)
• Rely on schema defaults for config reads
• Standardize config access calls

API

• [doc] Add OpenAPI reference metadata (P1-53a)
• Add embedded/library mode client (P1-15)

Build

• Wire dev serve flags for workers, reload, and instance name
• Recreate stale PaveDB venvs

Performance

• Refactor benchmark server config, restore bench model
• Cover batch document ingest endpoint (P2-48)

CLI

• Add parser metadata for CLI reference (P1-53b)

UI

• Migrate served template to centralized branding tokens

0.9.1 — 2026-06-15

Core

• Move embedder specs onto embedder modules (P1-32)
• Convention-based embedder dispatch (P3-49)
• Add OllamaEmbedder as a thin OpenAI-compatible wrapper for a local Ollama dae…
• Replace startup warmup with store.warm() behind a config knob (P3-48)
• Align typing policy with SDK boundaries (P1-57)

Store

• [fix] Treat concurrent-delete race in search as 404
• [fix] Stop persisting embedder defaults that bust the load cache
• Remove _system tenant carve-outs
• [fix] Keep health probes off the per-collection lock walk
• Hoist query embedding out of the collection read lock

API

• Add collection model specs and move/update APIs (P1-32)
• Return 503 for unavailable embedders (P1-32)

Performance

• Cover collection update and move APIs (P1-32)

0.9.0 — 2026-06-11

API

• Get document by ID endpoint (P1-17)
• Per-collection embedder params in create-collection (P1-32 prep)
• Add ok field to search response (P1-14)
• Add response models for CRUD endpoints (P1-14)
• Mount API routes under /v1/ prefix (P1-43)
• Expose timing breakdown in search response (P1-40)
• Add X-Request-ID middleware (P1-11)
• Echo request_id and latency_ms in all JSON response bodies (P1-11)
• List documents endpoint (P1-44)
• Enrich list collections response (P1-45)
• Get collection detail endpoint (P1-45)
• Log search queries + collection-scoped query log endpoints (P1-41)
• Reject search-body request_id; use X-Request-ID (P1-48)
• Query replay endpoint (P1-42)
• Admin query-home resolver + shortcut routes + CLI rework (P1-51)
• [fix] make global searches go through the ops_event annotation and generate o…
• Add actor field to ops_log events (P2-46)
• [cli] Add chunk inspector endpoints and commands (P2-23)
• [bench] Surface search query IDs for stress replay (P2-48)

Store

• Add CatalogDB catalog store (P1-33)
• Wire CatalogDB into LocalStore (P1-33)
• Return SearchOutput with per-phase timing (P1-40)
• Persist query logs in per-collection meta.db (P1-41)
• Enrich query_log with historical scope and actor (P2-46)
• Consolidate pre-1.0 SQLite schemas
• [fix] Avoid collection reopen races after cache flush
• [fix] Avoid CatalogDB reopen races after cache flush
• [fix] Wait for SQLite writers before closing DB handles
• [fix] Serialize archive restore with store ops (P1-52)
• [fix] Keep search hydration under collection lock (P1-52)
• [fix] Serialize SQLite read connections (P1-52)
• [fix] Replace documents atomically during ingest (P1-52)
• Allow concurrent same-collection reads (P1-52)
• [fix] Make catalog updates atomic (P1-52)
• [fix] Return 409 for rename target collisions (P1-52)
• [fix] Pin cached collection reads during races (P1-52)
• [fix] Reopen vector backend after collection rename (P1-52)

Performance

• [fix] Disambiguate bench-stress coverage-pass misses
• [fix] Preserve ephemeral server logs on benchmark failure
• [fix] Fail when managed benchmark server exits (P2-48)
• Add critical/full stress suites (P2-48)
• Auto-discover stress ops for full suite (P2-48)
• Warn on stress coverage gaps and race outcomes (P2-48)
• Add stress ops for remaining OpenAPI endpoints (P2-48)
• Add bench-stress-full target (P2-48)

Build

• Defer release commit until local validation passes
• Keep version in one file
• [fix] Avoid eager embedder loads in build-check (P3-52)
• [fix] Force BuildKit on docker build
• [fix] Use validated bash for make recipes
• Make release builds opt-in and name dev artifacts

UI

• Align docs UI with PaveDB theme
• Regroup API browser tabs and tag groups
• Add API browser tab deep links
• Add contextual hints to API browser tabs
• Simplify API browser wording and chrome

CLI

• Add query log inspection commands (P1-41)
• Add collection and document inspection commands (P1-44 P1-45)
• Add collection config args to create-collection (P1-45)
• Add explicit common-merge search flag

Core

• [fix] Normalize _do_search error results
• [fix] Cross-check scoped query lookup against query_home
• [fix] Isolate SBERT from server process on macOS (P3-26)
• [fix] Prefer current SBERT dimension API (P3-26)

Documentation

• Update roadmap and plan docs
• Document store concurrency model (P1-52)
• [fix] Default to non-cpu quick startup

Testing

• CatalogDB integration + upgrade tests (P1-33)
• Make LocalStore tests use temp_data_dir explicitly
• Consolidate collection API tests into one file

Log

• Improve server failure diagnostics (P1-52)
• Clarify concurrency cap diagnostics (P1-20)

Infrastructure

• Rename GitLab CI config file

Packaging

• Finalize PaveDB rebrand (P3-35)

0.9 series — Rebrand to PaveDB

Breaking changes

• PATCHVEC_* env vars are no longer recognized; use PAVEDB_*
• Default instance home is ~/pavedb/ instead of ~/patchvec/
• Docker image renamed to pavedb
• Published registry path is now registry.gitlab.com/flowlexi/pavedb/pavedb:*

0.5.9 — 2026-04-03

Store

• CollectionDB read-only open for fallback reads
• Close abandoned CollectionDBs on cache flush
• Extract vector backend seam (P1-29a)
• Clean VectorBackend protocol, add FaissBackend, activate embedder path (P1-29b)
• Adapt SbertEmbedder to Embedder protocol (P1-29b)
• Add QdrantVectorBackend stub (P1-29b)
• [fix] validate FAISS index dimension on load
• [fix] harden chunk/archive I/O against TOCTOU races
• Add chunk_meta CollectionDB storage (P1-29c)
• Add exact-negation CollectionDB pushdown (P1-29c)
• Sanitize metadata keys and values before inserting to chunk metadata table
• Split document/chunk metadata storage and pushdown (P1-29c)
• Remove dead txtai backend and legacy methods (P1-37)
• Simplify filter path — pass normed_filters directly to filter_by_meta (P1-37)
• Drop txtai dependency, promote SbertEmbedder to default (P1-37)
• Rename TxtaiStore → FaissStore, simplify file names, reshape provider-scoped…
• Fix filter_by_meta mixed-OR key narrowing incorrect results
• Reject invalid sanitized metadata keys (P1-36)
• Extract filters and sanitization to pave/filters.py (P1-31)
• LocalStore orchestrator replaces FaissStore (P1-31)

Build

• Fix make install-dev output under .ONESHELL
• Update make check to use faiss config
• Add make build-check target (P3-52)
• Include openai extra in Docker image
• Add make docker-check target (P3-51)
• Gate release on smoke checks
• Rename build/CI identifiers and benchmark labels to pavedb (P3-35)

Documentation

• Add third-party license audit and dependency policy
• Add Step 2 benchmark results to docs/PLAN-STORE.md (before/after FAISS cuto…
• Refresh docs for faiss/sbert naming (P1-37)
• Clarify collection logical-vs-physical mapping
• Remove unsupported qdrant runtime mode references
• Add inspectability/control layer to roadmap

Performance

• Implement variant filtering latency modes, with fresh model per variant run a…
• Latency results before and after metadata k,v table implementation, several f…
• Benchmarks after txtai removal
• Add –slo-p99-ms exit gate to search_latency (P2-30)
• Add –max-error-pct exit gate to stress (P2-30)
• Wire SLO flags into Makefile targets (P2-30)

Config

• Fix tenants sidecar precedence and dev config loading
• Temporarily map supported legacy PATCHVEC_* env vars to PAVEDB_* during the phase-1 migration; removed in the 0.9 rebrand series (P3-35)
• [fix] isolate default config state across instances

Service

• Add error logging at all ok:false service layer return sites (P2-40)
• [fix] wrap create_collection and health probe in collection_lock
• [fix] reorder archive restore and lock exclusion, rewrite data_export tests (…

Testing

• Drop unsafe skipped concurrent upsert race case
• Rename test_faiss_* → test_store_* (P1-31)
• Add opt-in public relevance regression checks (P2-29)

Tests

• FaissBackend and TxtaiEmbedder tests.
• Cover chunk_meta CollectionDB storage (P1-29c)
• Cover pushdown edge cases (P1-29c)

API

• Split main.py routes into APIRouter modules (P3-50)

Bug Fixes

• Harden store lifecycle races found by bench-stress

CLI

• [conf] add explicit instance bootstrap and runtime paths (P1-34)

Core

• Rename operator-visible identifiers to pavedb (P3-35)

Deps

• Package sbert runtime and relevance deps

Infrastructure

• Add benchmark CI job with SLO gate (P2-30)

0.5.8.1 — 2026-03-07

API

• [cli] add list-tenants and list-collections endpoints and commands
• Ensure latency_ms in all /search returns (incl. common disabled)
• Normalize error envelope and make document delete idempotent
• Reject uploads exceeding configurable size limit (default 500 MB)
• Search timeout + concurrency cap (P1-20)
• Ingest concurrency cap (P1-20 follow-up)
• Per-tenant concurrent request cap (P1-08)
• Expose server HW info in /health/metrics

Documentation

• Add coding style, commit message standards to developer docs
• Revamp ROADMAP structure, priorities, and release ordering
• Config.yml.example full docs pass
• [bench] baseline benchmark results before SQLite metadata store
• [perf] Phase 1 benchmark results + impl comparison (winner:impl2)
• Plan large scale store refactor

Performance

• Rebalance stress weights; error tracking in both scripts
• Improve benchmark scripts, make targets, and docs
• Improve benchmark resilience: retry on failed seeds and add –debug mode
• Run header, unified table format, sample results; tune stress duration/concur…
• Stress: coverage pass for ops not picked during timed phase
• Fix op_delete_collection: remove from world only after success

Store

• Migrate legacy txtai indexes missing documents/objects/sections tables
• Phase 1 SQLite per-collection meta.db — impl2 (read/write split conn)
• [tests] fix has_doc cache race and add regression test
• [search] scope metadata fetch to top-k when no post-filter
• Store-backed catalog metrics (P1-21) + validation hardening
• Accept optional doc_meta and pass it explicitly to SQLite upsert

Core

• Add collection rename functionality across all layers
• Standardize naming across API, CLI, and service layers
• Add SearchResult dataclass for type-safe search results
• Lazy build_app(): app only initialised on first access
• Replace key with character in TXT preprocessor (P2-41)

Log

• Dev stream cleanup (P2-28)
• Ops stream: pave/log.py, ops_event decorator, 8 routes (P2-28)
• Merge diverging log level configs, set defaults, surface log.level as source…
• Migrate dev stream from config.py to log.py; ~ expansion for log paths
• Add top result excerpt to search log line

Build

• Make release re-run safety, docker-build/push USE_CPU parity
• Fix docker-build/push error propagation on sub-make failure
• [docs] overhaul Makefile release flow and project docs

Config

• Ingest timeout guidance: expose timeout_keep_alive, nginx proxy hints
• Complete _DEFAULTS, ~ path expansion, fix Makefile env vars
• Default path ~/patchvec/config.yml; expand PATCHVEC_CONFIG too

Testing

• Speed up suite 5x: inject FakeEmbeddings for non-slow tests
• [store] fix Python 3.12 sqlite3 DeprecationWarning; silence SwigPy noise

Infrastructure

• Remove deploy jobs, add docker RC build, fix tag patterns

Packaging

• Changelog generation goes back to last tag in changelog itself not in git his…

Service

• Fix _flush_store_caches: drop refs, do not close() connections

0.5.7 — 2026-02-21

Store

• Ensure CRLF-rich documents round-trip intact
• Sanitize txtai metadata persistence (closes #3)
• Prevent infinite recursion with deeply nested collections
• Push !-prefixed filters into SQL pre-filter instead of post-filter, reducing…
• Change default embedding model to multilingual and add multilingual cross lan…
• Build a match_reason as part of the search return contract
• Fixed race condition in get_lock() with double-checked locking pattern using…
• Replace eval() with operator module in filter matching
• Replace assert with runtime check in index_records
• Fix collection_lock usage across all TxtaiStore methods
• [fix] Disable meta-device loading so that Pooling.to() works on PyTorch>=2.6…

Documentation

• Refresh workflows and roadmap
• Revise roadmap based on technical evaluation
• [test] Document and test the live-data-update path (purge + ingest)
• Added benchmark suite documentation
• Update short description (ABOUT.md) to reflect recent workflow changes

Build

• Push torch cpu requirement to 2.8+
• Add cpu/gpu extras with proper torch wheel selection and python version. Clos…
• Fix assets package and license definitions in setup.py
• Consolidate deps into setup.py, single source of truth

Core

• Add feature to dump entire datastore as .zip file (api and cli)
• Add feature to push/restore data archive from zip backup (api and cli)
• [metrics] Persist metric across app resets and add metrics reset cli utility…
• Pretty terminal logging with colors and level-based filtering

API

• Add latency histograms (p50/p95/p99)** on for search and ingest
• Structured request logging with request_id, added latency and request_id to s…
• [cli] Add delete document by ID endpoint and command

Performance

• Add concurrent stress test exercising all supported API operations
• Eagerly load embedding model at server startup

Bug Fixes

• Atomic writes, coalesced metrics flush, Makefile benchmark targets

Config

• Implement initial/decent multilevel logging support

Infrastructure

• Add benchmarks/ directory with search latency load test

Packaging

• Revamp the automated changelog generation script

Chores

• Update ROADMAP.md
• Update to Python 3.10+ type hint syntax
• Update copyright notices to 2026, add missing copyright headers
• Use Python 3.10+ typing syntax; warn on corrupt index
• Small fix in README.md to make docs consistent with docker image naming scheme
• Update project urls

0.5.6 — 2025-10-29

Core

• Added ingestion timestamps to document metadata and improved CSV ingestion controls (headers, meta columns and include lists).
• Hardened API boot by forcing string-based Uvicorn startup and gating document purges behind has_doc checks.
• Normalized service entrypoint configuration, including stricter binding and authentication safeguards.
• Standardized request metrics emission and activated service-level telemetry across the API.

Store

• Prevented FAISS index overwrites on multi-document ingests and ensured index directories are created eagerly.
• Improved FAISS store concurrency through SQL filtering hooks, stronger locking, and thread-safe helpers.
• Guaranteed text chunks are persisted and hydrated when vector content retrieval falls back to storage.
• Ensured txtai_store consistently returns search text results.

Build & Packaging

• Added dedicated Makefile targets for local deployment, e2e checks (still needs work) and dependency cleanup.
• Enabled the Docker build pipeline with split GPU/CPU flows, refined image tagging, and updated startup scripts.
• Extended release automation with PyPI publishing support, GitLab pipeline steps, and tuned Makefile/setup.py metadata.
• Updated dependency sets, including explicit faiss-cpu support and auxiliary tooling definitions in pave.toml.

Config

• Introduced multilevel logging defaults and refreshed the example embedding model to a multilingual preset.
• Expanded configuration backend coverage with additional tests.

UI

• Added a lightweight Swagger/OpenAPI UI with branding, authorization helpers, and contextual headers/footers.

Testing

• Simplified store mocks by pinning default embedding models and cleaning up legacy FAISS test shims.

Misc

• Updated project metadata, copyright headers, and ignore lists.
• Advanced version markers for intermediate dev builds and release tags (0.5.5 → 0.5.6devN).

0.5.5 — 2025-09-02

Core

• Added CSV ingestion configuration knobs (headers, meta columns, include filters).
• Implemented default document ID handling to overwrite vectors deterministically on re- ingest.
• Fixed authentication edge cases and expanded accompanying tests.
• Ensured request metrics are emitted consistently across the API surface.

Build & Packaging

• Added an end-to-end Makefile target and improved startup scripts with dependency cleanup steps.
• Extended release automation with Docker targets and a PyPI publishing flow.
• Refined dependency management by bundling faiss-cpu for CPU builds and sqlite4 for testing.

Store

• Made FAISS-backed collections initialize their index structure on creation.
• Corrected the txtai store path so search responses always include original text payloads.

Commits

• [buid] Add e2e check target and cleanup Makefile
• [build] Enhance startup scripts, add dependency clean Makefile target
• [core] Add CSV ingestion options: headers (yes|no), meta_cols and include_cols
• [core] Add default docid behavior so that vectors are seamlessly overriden when same file is ingested even if no docid is provided
• [core] Fixed txtai_store to handle indexes correctly and always return search text.
• [core] Fixing auth and adding tests
• [core] Normalize entry point config and add binding and auth safeguards for prod envs
• [core] Standardize request metrics in API and enable service metrics in service pipeline
• [pkg] Add docker targets and make further adjustments do Makefile
• [pkg] Add pypi publish makefile target
• [pkg] Fix dependencies: add sqlite4 to testing and explicitly add faiss-cpu to cpu- only target
• [store] Make sure FAISS indexes and dir structure are initialized upon collection creation.
• Fix .gitlab-ci.yml file

0.5.4 — 2025-08-12

Feat

• initial public release of PatchVec — multi-tenant, pluggable vector search microservice

0.5.3 — 2025-08-12

Config

• Introduced minimal .env.example (required vars only; PATCHVEC_ scheme)
• Clarified tenants secrets via untracked tenants.yml

Packaging

• Added Makefile-based release flow (tests gate release; Docker/compose versions bumped automatically)
• Added Gitlab & Github CI/CD workflows (not tested)

Docs

• Split README into README.md (end-user) and CONTRIBUTING.md (dev)
• Added REST examples with curl
• Documented uvicorn server overrides via HOST, PORT, RELOAD, WORKERS, LOG_LEVEL
• Trimmed Quickstart and added PyPI install path

0.5.2 — 2025-08-12

Testing

• Added CSV and PDF ingestion/search
• Adjusted test cases for FastAPI’s stricter body/query validation.

Arch

• Added TxtaiEmbedder as default; added OpenAIEmbedder and SbertEmbedder.

API

• Fixed POST /search route to accept SearchBody via JSON body.

Other

• Added docker-composer stub

0.5.1 — 2025-08-11

Testing

• Added comprehensive pytest test suite covering:
• Collection creation/deletion
• Document ingestion & search (TXT)
• Re-ingestion with purge
• Expanded pytest coverage for TXT ingestion, re-ingestion, and search.
• Fixed relative import issues in tests.

Arch

• Refactored store and embedder factories to use Python 3.10+ match syntax.
• Standardized naming (*_store, *_emb).

Auth

• Refactored authentication/authorization into auth_ctx() and authorize_tenant() using FastAPI dependency injection.
• Authorization now automatically derives tenant from bearer token when applicable.

API

• Unified GET and POST search behavior.

0.5 — 2025-08-11

Testing

• Added DummyStore for testing.

Arch

• Isolated vector store interfaces via ABC (BaseStore) for plug-in stores (Qdrant, FAISS, etc.).
• Added StoreFactory and EmbedderFactory with runtime store/embedder selection (pluggable backends).

API

• Added /health endpoint with general metrics and alive status

0.4 — 2025-08-10

• Modularized codebase:
• stores/ for vector store backends.
• embedders/ for embedding backends.
• auth.py for authentication/authorization implementation.
• service.py for main logic implementation (abstracted from api and cli interfaces).
• cli.py for cli implementation.
• main.py for endpoint routing and default initialization.
• preprocess.py for file ingestion helpers.
• metrics.py for metrics implementation.

0.3 — 2025-08-09

• Added QdrantStore skeleton (methods unimplemented).
• Added OpenAIEmbedder proof of concept (not tested).
• Introduced CFG for unified cfg management.
• Added complete cli mode

0.2 — 2025-08-08

• Implemented multi-tenant routing (/{tenant}/{collection}).
• Added basic static authentication via global or per-tenant API keys.
• Added document ingestion and collection management endpoints.

0.1 — 2025-08-07

• First working prototype with:
• FastAPI service with search endpoint.
• FAISS store and Sbel embeddings.
• Command-line TXT ingestion and REST search endpoint.
• Single-tenant mode.
• Minimal auth stub.