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.eval...

reranker_en model

ONNX INT8 reranker_en for vectrixdb download-models

dense_en model

ONNX INT8 dense_en for vectrixdb download-models

colbert model

ONNX INT8 colbert for vectrixdb download-models

v2.1.7 - OpenSearch Serverless Compatibility

OpenSearch Serverless Compatibility Fixes

This release fixes critical compatibility issues with AWS OpenSearch Serverless:

Bug Fixes

• Silent error handling - Exceptions from storage backend were being swallowed; now properly re-raised
• Key mapping - Fixed _embedding not mapping to dense_embedding in insert_batch
• Custom document IDs - Serverless doesn't support custom _id; now using id field in document body
• Refresh policy - Removed refresh=True from all operations (not supported by Serverless)
• Document lookup - get, update, delete methods now search by id field instead of _id

Upgrade Notes

If you're using OpenSearch Serverless, upgrade to this version for proper functionality.

v2.1.5

What's Changed

Proper Hybrid Search with Reciprocal Rank Fusion (RRF)

Implemented industry-standard hybrid retrieval using RRF algorithm:

How it works:

1. Stage 1: Run k-NN search (semantic similarity) and BM25 search (lexical matching) independently
2. Stage 2: Fuse results using RRF: score = dense_weight/(k+rank_dense) + sparse_weight/(k+rank_sparse)
3. Stage 3: Return top results sorted by fused score

Why RRF?

• Robust to score distribution differences between retrieval methods
• No need for score normalization (uses ranks, not raw scores)
• Industry standard (used by Elasticsearch, Cohere, etc.)
• Works with OpenSearch Serverless (no scripting required)

New methods:

• text_search() - BM25 lexical search on text_content
• hybrid_search() - RRF-fused k-NN + BM25

Parameters:

• rrf_k=60 - Standard RRF constant from literature
• dense_weight=0.7 - Weight for semantic results
• sparse_weight=0.3 - Weight for lexical results

Full Changelog: v2.1.4...v2.1.5

v2.1.3

What's Changed

Fix Hybrid Search Integration

• Fix _hybrid_search in Vectrix wrapper to properly detect OpenSearchStorage
• Call storage.hybrid_search() with correct parameters (query_vector, query_text)
• Properly handle result format conversion from storage to Vectrix format

Full Changelog: v2.1.2...v2.1.3

v2.1.2

What's Changed

Hybrid Search for OpenSearch

• Add hybrid_search() method combining k-NN dense vectors with BM25 text matching
• Update vector_search() with optional hybrid mode
• Uses OpenSearch script_score to combine semantic + lexical matching
• Configurable weights: dense (default 0.7) and sparse/BM25 (default 0.3)

This enables true hybrid search on OpenSearch Serverless - combining the best of semantic similarity and keyword matching.

Full Changelog: v2.1.1...v2.1.2

v2.1.1

What's Changed

• Fix OpenSearchStorage missing abstract methods (flush, get_collection_config, scan)
• This fixes the TypeError when using VectrixDB with AWS OpenSearch Serverless

Full Changelog: v2.1.0...v2.1.1

v2.1.0: AWS OpenSearch and Aurora PostgreSQL

What's New

AWS Storage Backends

• OpenSearch Storage - AWS OpenSearch Serverless with native k-NN vector search

  • Supports dense and hybrid modes
  • Factory method: VectrixDB.with_opensearch()

• Aurora PostgreSQL Storage - AWS Aurora with pgvector extension

  • Supports ALL modes including ultimate (ColBERT) and graph
  • Factory method: VectrixDB.with_aurora_postgresql()

Mode Validation

• OpenSearch now validates mode compatibility and raises a friendly error if you try to use ultimate or graph modes

Installation

pip install vectrixdb[aws]  # For AWS backends

Quick Start

from vectrixdb import VectrixDB, Vectrix

# OpenSearch (dense/hybrid only)
opensearch = VectrixDB.with_opensearch(
    endpoint="https://xxx.us-east-1.aoss.amazonaws.com",
    region="us-east-1",
)
db = Vectrix("products", mode="hybrid", storage_backend=opensearch)

# Aurora PostgreSQL (all modes)
aurora = VectrixDB.with_aurora_postgresql(
    host="cluster.xxx.rds.amazonaws.com",
    user="admin",
    password="password",
)
db = Vectrix("products", mode="ultimate", storage_backend=aurora)