VectrixDB 2.2.0

Added

• The server stays up under load.
• More MCP tools, resources and prompts.
• A skill for an assistant connected over MCP.
• The docs, grown.
• One release, every registry.
• The documentation's home page shows every way in:
• A client in four languages, one surface.
• Clients safe to point anywhere.
• A company's presets, its command's name, and its keychain.
• The vectrixdb command on a server.
• Run it inside a company's registry.
• vectrixdb doctor tries every part of an install.
• A collection keeps up with feeds and pages.
• A release is one button and one approval.
• Container images, for Intel and ARM, signed.
• vectrixdb extract-serve starts the extraction service
• The key for an extraction service can come from a file.
• vectrixdb check knows the variables Kubernetes sets.
• Searches, ingestion and evaluation runs can be traced, off until asked.
• MCP on the server, for a team or a company.
• vectrixdb mcp over HTTP serves the reading tools only
• A rechunk can be previewed.
• The docs have an index for coding agents and a prompt to give one.
• The test-question writer spreads its questions and judges harder.
• stale_evidence() says which questions' quotes have left the index.
• Moving pictures of the dashboard.
• VectrixSync.cdc() carries deletes to the target.
• scripts/dashboard_shots.py retakes the dashboard's pictures.
• Single sign-on, before its provider is set up.
• A security group is narrowed to a list of people.
• Emergency sign-in.
• Single sign-on and the People list, together.
• Single sign-on needs somebody named.
• An app's token needs no place on the list.
• Developer Access.
• A path for each endpoint behind a gateway.
• vectrixdb check says what reads each kind of file.
• The restricted note says why.
• The brand goes further.
• About, for admins.
• A NOTICE file travels with VectrixDB.
• A web page is read as its content.
• A PDF is read by PDFium, from where its text sits.
• A chart a PDF draws is described.
• The pages the rules are unsure of are read by sight.
• A file is known by its bytes.
• Spreadsheets read as they print.
• Word and PowerPoint read as they print.
• RTF and OpenDocument files are read.
• Who said what, in the languages it was said in.
• What a video shows.
• A picture's values come as rows.
• The walkthrough tries regions in the order you write them.
• The walkthrough makes a Translator.
• The steps have a requirements file.

Changed

• vectrixdb check says when a key leaves reads open.
• The mcp extra needs mcp 2 or later.
• A YouTube video with captions is read from them, not transcribed.
• The authenticator's QR code is drawn as the dashboard's own.
• The walkthrough sets how the query app scales.
• Three ways to set sign-in, and passkeys with one of them.
• Emergency sign-in goes with every way in.
• Developer Access is found by the spinner.
• GraphRAG on Bedrock asks with Converse.
• The examples are not in the repository.
• Types are checked as Python 3.12
• The walkthrough's chat model is gpt-5.4-mini.
• The documents extra brings pypdfium2 and xlrd,
• Extraction quality judges prose.
• A masked JSON reply masks everything in it,
• A busy service answers 503.

Fixed

• A signed link stays secret.
• A bot check is refused, not indexed.
• Three optional models are found where they were published.
• Two results that score alike come back in the same order every time.
• Years and figures are not phone numbers.
• A table's stacked column headings head their own columns.
• A footnote mark glued to its word is cut loose on every page,
• Keyword and hybrid search run in the store that holds the text.
• VECTRIXDB_OFFLINE holds for the speech model.
• The audit before release.
• The README's pictures showed as broken images.
• mypy vectrixdb failed in CI with numpy 2.4's stubs.
• The docs no longer describe what left with visibility.
• A guest opening a collection is not asked to sign in.
• Policied counts the same on both pages.
• A test of a killed writer miscounted.
• A fresh checkout passes its own tests.
• A collection's Overview drew nothing for somebody signed in.
• A tab the collection's policy refuses is no longer left blank.
• Counts on a deployment.
• The line under the single sign-on screens sits in the middle.
• The second vector installs with the Azure extra.
• The walkthrough's query app asks people to sign in.
• The walkthrough's query app finds its collections.
• The walkthrough's query app has guests.
• The walkthrough's Cosmos account has one database, data_db.
• The walkthrough's embedding model keeps up with an annual report.
• A collection kept in the cloud opens where nothing was written before.
• A Markdown table whose header leaves out the label column reads right.
• A web page's text has no empty lines of spaces and no empty items.
• An old page reads whole and in time.
• A web page is decoded in the charset it names.
• Step 03 no longer reads a working create as a refusal.
• The records' database is data_db.
• The walkthrough's app is deployed twice: reads and queries apart.
• Reading a file survives a bad hour.
• The walkthrough's dashboard shows the files that would not read.
• A run records the collections it ran over.
• A collection's knowledge graph is kept where every server reads it.
• The overview recomposed, and a guest's overview drawn from it.
• A collection's record is its policy, and nothing else.

Removed

• Visibility, the masking switch and per-document policies on records.
• The dashboard's charts, redrawn, and every list ten a page.
• The error catalogue.
• Masking with an engine that reads meaning, before a document is indexed.
• Who may retrieve from a collection: its policy, checked before anything is searched.
• Visibility is private or public, and a guest reads excerpts.
• The dashboard shows and sets who may retrieve.
• The Azure walkthrough's app makes a collection the way its steps do.
• Each collection's rules in one record every server reads.
• The audit trail and the access log in append blobs nothing can rewrite.
• Readable Cosmos DB items.
• The chunking techniques compared, and the Evaluate page's Chunking tab.
• A cut and a way of searching a host can follow the runs for, or pin.
• Runs are kept by kind: retrieval/runs/<id>/ and chunking/runs/<id>/.
• Four more ways to cut and embed.
• A golden row keeps its evidence: the words that answer it.
• The collection pages count every instance's chunks, not one.
• It runs behind a gateway.
• VECTRIXDB_TRUSTED_PROXIES says whose X-Forwarded-For to believe.
• The OpenAPI document ships with the release.
• Running heads and feet are taken out, and named.
• A chunk knows whether its own page was read by OCR.
• An HTML table in an extractor's reply becomes rows
• compare_extractors
• A document fetched from a bucket keeps its pictures.
• A walkthrough that deploys the whole thing on Azure
• Where to stop answering is measured, not borrowed.
• evaluation.sweep() compares how documents are cut.
• vectrixdb golden label
• Two homes.
• A figure found by what it looks like.
• The rest of the dashboard's trends
• The Evaluate page has five bands: first, top 3, top 5, top 10, missed.
• A reference extraction server
• A queue between the storage event and the worker on Azure
• The server records how long a search took, and the Overview draws it.
• The Overview and Audit pages have charts.
• Every refusal is one shape.
• An app can act as the person using it: VECTRIXDB_OIDC_API_AUDIENCE.
• A rate limit several servers share, and one for keys.
• vectrixdb keys add | list | revoke.
• vectrixdb check --url
• A key can be narrower than a role: collections and expires_in_days.
• A key may arrive as Authorization: Bearer.
• Build an app on it
• Put it behind a gateway
• Masking of email addresses, phone numbers and card numbers.
• Reference pages read off the code.
• Two pages for knowing it all.
• A test extra
• A collection's health says whether it has a knowledge graph
• Evaluate every setup, and pick one of three.
• Golden files from anywhere, and a template to fill.
• A golden file is checked before a run, against one published schema.
• A golden answer can be one page of a document.
• The walkthrough's files are a mirror of the storage account.
• vectrixdb golden template and vectrixdb evaluate.
• The Evaluate pages.
• A Function App that evaluates when the golden file changes.
• Azure's semantic ranker, chosen search by search.
• Expected documents are looked up before an evaluation searches.
• Older runs, one menu away.
• A note when a run is missing documents.
• The golden file a run used, for an admin to download.
• vectrixdb check, an env file and a template.
• The access log on the server's output.
• relevance: how good a match is, from 0 to 1, the same on every engine.
• The same number from Azure AI Search and OpenSearch.
• A collection's card says what it can do and what it is built with.
• A chunk's text is sent when somebody asks for that chunk.
• Opening a whole document is its own permission, document.read.
• Search excerpts, cut on the server.
• The quality report says why.
• People sign in to the server and its dashboard.
• Passkeys.
• How you sign in.
• Confirm it's you.
• Passwords, only if you want them.
• A lock that grows.
• Single sign-on by security group and an address list.
• Guests and shared collections.
• API keys for scripts.
• Sign-in kept where several servers can share it.
• Passkeys only, if you want that.
• A certificate in place of a client secret.
• A company's own name, logo and colour.
• The dashboard is light by default,
• Roles.
• A signed-in person can search a collection that carries a policy.
• An access log.
• Policy and filter pushdown on OpenSearch.
• embeddings= on OpenSearch, with Bedrock.
• reranker= on Vectrix.
• A burst of events.
• vectrixdb.evaluation.
• A reference AWS stack
• search(vectors=) and vector_weights.
• More than one dense model locally.
• Documents over REST.
• The dashboard reads documents through the server.
• keep_source: the Markdown a document was indexed from is kept.
• Front matter.
• The worker reads an original once.
• The figure block.
• describe_figures=.
• Extraction is a slot.
• Engines for what is not text yet.
• Extras named for what the file becomes.
• embed_heading=True on add_document().
• Tables are rows wherever they come from.
• XLSX, PPTX and CSV loaders.
• Rerank on Azure AI Search means the service's semantic ranker.
• Sentences survive the page break, and figures are chunks of their own.
• An ingestion worker for every event source.
• Citations.
• Engine pushdown on Azure AI Search.
• The dashboard, rebuilt.
• Inspection routes
• "rerank": true on /text-search
• Every REST write mints an index build id
• Entitlement policies: a filter a collection cannot be queried without.
• A slimmer wheel: e5-small-v2 is fetched once instead of shipped.
• An audit sink under an object lock.
• An extraction quality signal, measured before it was shipped.
• A decision can be restated.
• Provenance on every chunk, and on the ingestion record.
• The evidence pack.
• A timing floor, for the one channel a filter cannot close.
• The write path's half of the policy: a value no rule can decide on.
• A live row level security suite, gated.
• A test kit for your own policies, shipped rather than kept in ours.
• A document the policy could never show anybody is refused at the write.
• The write path is audited too.
• A policied read can run as a database role, so PostgreSQL enforces it.
• Translating a policy into SQL is off the plan.
• revoke(): change entitlements across many chunks in one call.
• The wheel job runs the README's own first example.
• The id lookups are gated too, and a closed collection answers nothing.
• A policied collection searched through Collection says so.
• An append-only PostgreSQL sink for decision records.
• A decision record for every policied read, and somewhere to put it.
• The README covers conversation memory and the MCP server.
• Lakebase and Aurora PostgreSQL can be run against the real service.
• Every documentation page with a Python example now runs in CI.
• Nine filter operators gained tests.
• Thirteen regression tests, each with a replay.
• Every storage backend now passes the contract suite on every push.
• A collection can be written past RAM.
• Graph quality block.
• Examples gallery and governance.
• Developer experience block.
• Memory block: memory is finite and facts conflict.
• Azure AI Search storage backend.
• Ingestion block: where RAG quality is decided.
• Trust block: measured, reproducible, published.
• search(rerank=False)
• DenseEmbedder.embed(pad_to_longest=True)
• Testing block.
• tokenizer.encode_batch(pad_to_longest=True)
• search(explain=True)
• Fusion control.
• Query expansion and HyDE.
• Field boosts for keyword search.
• Snapshots.
• rebuild_index()
• Read-only, memory-mapped opens.
• AsyncVectrix
• add_many()
• The filter grammar is documented
• scripts/compare_modes.py
• import vectrixdb costs 4 ms, down from 585 ms warm and 1.2 s cold.
• Nothing downloads unless asked.
• Downloads are checksummed.
• Vectrix is a real context manager.
• numpy in, numpy out.
• Notebook reprs and progress.
• An explicit public surface.
• Concurrency is documented
• Memory per vector is measured
• Python 3.14
• ProductQuantizer.fit runs in seconds instead of minutes.
• Token budgets on search.
• score_gap
• Conversation memory
• An MCP server
• tier="graph" now builds a knowledge graph.
• Entity resolution that understands names
• config.entity_similarity_threshold is finally read.
• Confidence labels on relationships
• Typing marker (py.typed).
• A public exception hierarchy
• CI that runs before release.

Security

• The audit before release.
• The dashboard runs no inline script, and its policy refuses all of it.
• The chart library is pinned by its hash.
• Any-origin CORS stops the start while sign-in is on.
• Every reply carries security headers.
• A collection or document name could run script in the dashboard.
• Session cookies only this host can set.
• A collection name was used as a filesystem path with nothing checked.
• An empty API key was replaced by the environment's real one.
• A user in no groups saw every document.
• Cosmos DB queries now use bind parameters.
• Databricks SQL values are bound, not escaped.

Changed

• The Azure walkthrough pushes each collection's record, and records every decision.
• Graph routes refuse in words a person can act on.
• CI tests what it claims to.
• The benchmarks were run again
• A missing OpenAI key is a ConfigurationError,
• The sidebar folds from its own header.
• Search results lead with relevance.
• A search mode the chosen collection cannot do is greyed out and says why.
• A card no longer says "hybrid" beside a tag that says "dense".
• The Quality tab's "Worst chunks" is "Below the quality line".
• The start-up banner says the API key is set and does not show it.
• The count of things that need attention is red, at zero too.
• Settings is a menu behind a gear in the top bar, not a page.
• The dashboard's look.
• A build id is shown as an id.
• Settings no longer lists the models on the machine.
• The dashboard works at three widths.
• Audit record schema version 2.
• The library no longer writes to stdout.
• Coverage is 85.1%, from 50.4%,
• mypy is clean and CI gates on it.
• The README led nowhere.
• PyPI's Documentation link pointed back at the README.
• uv is documented, because it already worked.
• The bundled English dense model is now bge-small-en-v1.5 (INT8).
• Graph search says when it fell back.
• Community detection and summarisation run only when the graph changed.
• The regex extractor sees lowercase prose.
• Type checking is a ratchet with teeth.
• The core install is five dependencies instead of thirteen.
• The document index was completely broken on SQLite.
• The one-line quick start failed on a clean install.

Deprecated

Removed

• setup.py and requirements.txt,
• filter_sql, ahead of its schedule.

Fixed

• A scanned page in two columns was read across the columns.
• Text that loops scored a perfect 1.0 on extraction quality.
• No collection card alone on a row.
• The keys table says what a key may do in full.
• The hand-written notebooks are out of tests/.
• The CI coverage job would have failed on its first run.
• Cosmos DB asked to create the document containers on every write.
• The dashboard said "1 chunks".
• A hybrid search spent most of its time on highlights nothing read.
• scripts/compare_modes.py --doc deleted part of the page it updates.
• The README's comparison table undersold two of the engines it compares.
• scripts/replay.py counted a replay that never ran as one that passed.
• Three scripts ignored --help and did their whole job.
• Benchmark results say what produced them.
• OpenSearch scores came back as one minus the engine's score,
• The MCP server did not start on a fresh install.
• A Markdown document of headings alone was indexed as nothing.
• requests was imported and never declared.
• The OpenAI embedder's errors were garbled.
• Tests skipped far more than they needed to without an optional package.
• Seven API test files could not be collected in CI
• The golden graph fixture depended on what else was installed.
• One test left the API modules re-imported for every test after it,
• vectrixdb check said there was no admin when single sign-on maps one,
• Rerank scored every result the same.
• Search results could come back without their text.
• The Overview counted no documents.
• A viewer was offered what a viewer cannot do.
• The sign-in card named ways in the server does not have.
• A guest's shared collections table split on a phone.
• Opening an Azure index could remove its semantic configuration.
• vectrixdb serve --path kept sign-in state in the wrong folder.
• The dashboard at tablet and phone widths.
• A folded sidebar showed a squeezed copy of its name.
• API keys were compared with ==.
• A selective filter could miss documents that match.
• An open sidebar drawer could not be closed.
• The menu button did nothing on a desktop.
• On a phone the dashboard's text overlapped.
• An upgraded server could show the old dashboard.
• A search repeated after a write returned the answer from before it.
• A delete never reached the storage backend.
• A revocation did not revoke anything a store served.
• Clicking a point in the dashboard did nothing.
• LocalFetcher could not read a real file URI on Windows.
• The Azure backend dropped its filter on a semantic query.
• The REST server never saved the ANN index after a write.
• The dashboard's Rerank and Late interaction modes were dense search.
• The dashboard said "Read Only" on a server with no key
• The sparse index only knew ASCII letters.
• A delete kept the build id.
• The decision record's duration_ms included the timing floor's sleep.
• BM25 counted documents the principal was refused.
• Four read paths on Collection were never gated.
• A wholesale denial returned instantly with time_ms=0.0.
• mode="sparse" refused a principal it had been given.
• A plain pip install vectrixdb could not create a collection.
• vectrixdb-mcp answered a missing extra with a stack trace.
• scroll() applied its filter after cutting the page.
• Seven filter shapes raised out of the middle of a search.
• A bounding box could not cross the antimeridian.
• ColBERTEncoder returned random vectors.
• The binary quantiser answered metrics it cannot measure.
• The product quantiser's euclidean returned squared L2
• A non-finite vector encoded to an ordinary-looking code.
• The one live-backend job could never run.
• A selective filter returned nothing at all.
• Reading from a store that never held a document raised.
• A missing model raised a bare FileNotFoundError.
• The scalar quantiser truncated where it should round.
• The product quantiser's cosine ranked by projection, not angle.
• The product quantiser crashed on any duplicate-heavy corpus.
• is_models_installed() said no on a correct install.
• vectrixdb serve --no-dashboard served the dashboard.
• index_text accepted a chunk size and discarded it.
• VectrixSync.incremental() was full() under another name.
• Deleting memories by an empty id list deleted everything.
• $in and $nin were wrong for every list-valued field.
• Absent, null and empty were the same thing to three operators.
• A failed search on Lakebase returned an empty result set.
• A collection that would not load was dropped in silence.
• The binary quantiser could fit to NaN and report success.
• Aurora PostgreSQL's hybrid search accepted a filter and ignored it.
• A sparse model that cannot run was swapped for BM25 in silence.
• VectrixSync.full(collections=[]) copied every collection.
• Lakebase lost the vector on a round trip and could not update at all.
• Delta Lake reported success whatever happened.
• A search limit of zero returned one result.
• The dashboard's file upload never worked.
• A heavily deleted collection got slower with no sign of why.
• One unbroken token could stop an ingest.
• An explicit empty filter returned everything instead of nothing.
• Cosine search could miss its own best hit.
• Graph extraction dropped or dangled relationships when entities merged.
• Merging two extraction results reset a fact to inferred.
• LLMExtractor.extract_batch() ignored batch_size.
• Three crashes and a silent miscount in the search helpers.
• The chunker could exceed the size it was given, and repeat an offset.
• VectrixSync.full() copied nothing and reported success.
• VectrixDB.with_auto_scaling() raised TypeError on every call.
• Graph extraction over more than one chunk crashed.
• The global graph searcher stashed a score on somebody else's dataclass.
• Enterprise search returned documents nobody was cleared to see.
• A collection reported HNSW parameters it was not built with.
• An installed ColBERT model was reported as missing.
• vectrixdb download-models --type colbert raised KeyError.
• A failed model download left its zip behind.
• TextAnalyzer(stopwords=set()) stripped English stopwords anyway.
• A new Point carried a naive timestamp.
• The missing-dependency hint dropped the extra it names.
• create_app(db_path=...) ignored its argument.
• POST /api/v1/documents raised on every call.
• The benchmarks page quoted e5-small-v2 numbers as bge-small-en-v1.5.
• A deleted vector could hide the documents that remained.
• Graph mode returned results in an order its scores did not explain.
• Vector search spent most of its time outside the index.
• AuroraPostgreSQLStorage could not be instantiated.
• Readers in other processes failed while a writer was busy.
• A process that only searched rewrote the index file on close.
• Malformed filters escaped the parser as AttributeError.
• Docs examples that could not run.
• Search did not work. Four separate defects, now all fixed.
• The reranker stage never ran for bundled models.
• BM25 uses BM25's IDF.
• The sparse index saw no CJK text.
• Vectors in the local document store are float32 blobs.
• NativeHNSWIndex is usable now.
• create_app() can be called more than once.
• The LLM and hybrid extractors could not be constructed.
• DeltaLakeStorage could not be instantiated.
• get() and similar() raised on every call.
• A graph with no entities raised on every query.
• A graph with no communities was indistinguishable from no graph.
• Storage backends no longer swallow failures.
• Adding the same text twice crashed.
• geo_radius silently returned too few documents.
• StringIndex contains behaved as equals.
• The SQLite backend is now actually thread-safe.
• SyncStatus.lag_seconds was wrong by the machine's UTC offset.
• The Delta Lake backend wrote local-time timestamps
• __version__ said 2.1.0 while the package was 2.1.7.

Known limitations

• The bundled model is the weaker one for conversation memory.

Every entry in full: CHANGELOG.md