IRIS Vector Management

Overview

Turn vector data into something you can inspect, understand, and operate. IRIS Vector Management brings vector discovery, storage evidence, similarity search, guarded HNSW creation, and repeatable vector-generation tasks into one portal hosted inside InterSystems IRIS. Start with an existing table, select the text to encode, and create a vector column using an embedding model or your own Python function—without running a separate application server.

Built with IRIS Embedded Python, native SQL, dictionary metadata, and SysAdmin API v2. This release targets the pinned IRIS Community 2026.2 Build 221U image, not every IRIS release. Screenshots show the local demo; some newer task controls currently have Portuguese labels.

Features

• Discover VECTOR and managed EMBEDDING columns, dimensions, model bindings, and HNSW indexes.
• Browse records with keyset pagination and fetch bounded vector previews on demand.
• Search by vector or managed text, with SQL and actual query-plan evidence.
• Trace namespaces, storage globals, database mappings, and database directories.
• Validate and create eligible HNSW indexes through explicit confirmation.
• Add a vector column to an existing table and populate it through one IRIS task.
• Use a registered embedding configuration or trusted Python code with syntax validation and sample testing.
• Rerun tasks for all rows, changed rows, or missing vectors; inspect scheduler history and run counters.

Install and try it

1. Prerequisites

Install Git and Docker with Compose v2 and Linux-container support; Docker Desktop is suitable on Windows. The first build downloads IRIS and CPU machine-learning dependencies, so allow time and disk space for a large image. Running the portal requires no separate Python, Node.js, or front-end build.

2. Clone and configure

git clone https://github.com/Davi-Massaru/iris-vector-management.git
cd iris-vector-management

Copy .env.example to .env: use Copy-Item .env.example .env in PowerShell or cp .env.example .env in Bash. Keep WITH_FIXTURES=1 for this tutorial; 0 omits the demo fixtures and their SentenceTransformers dependencies.

3. Set the local demo password

Create .local/iris-password.txt containing a strong password on one line, saved as UTF-8 without a BOM. Compose mounts it as a secret; the startup script sets the local _SYSTEM password. The image contains no baked-in human password, and .local/ is ignored by Git—do not share this file.

# PowerShell: create the directory, then enter the password in your editor.
New-Item -ItemType Directory -Force .local
notepad .local/iris-password.txt

On Linux/macOS, use mkdir -p .local, create the same file with your editor, and restrict its permissions with chmod 600 .local/iris-password.txt.

4. Build and start

docker compose up -d --build
docker compose ps
docker compose logs -f iris

Wait for startup to finish and the service to become healthy. Press Ctrl+C to stop following logs without stopping the container. Open the local portal and sign in as _SYSTEM with your configured password. This account is convenient for the local demo; use scoped accounts for real deployments. Compose binds port 52773 to loopback only.

5. Explore the included data

The default build creates data.Document with 20 Faker-generated documents, sample vector tables, and the vector-fixture-tiny embedding configuration. Seeding happens during the build, not through a portal screen. The tiny 16-dimensional model is an offline, randomly initialized fixture: it tests the pipeline without an external model API key, but is not a semantic-quality benchmark.

6. Create your first vector column

Open Vector tasks and enter the example below. Check the sample-test checkbox, click Validate and preview, review the proposed operation, enter the exact target confirmation, and select Publish and create task. Publishing creates an on-demand task; its Run now action adds the column and populates matching rows.

Refresh inventory after completion to inspect data.Document.DescriptionVector. Run the same task again to test repeatability: with CHANGED and unchanged input/generator, rows should be skipped. A new recipe creates a new column; it is not how you rerun an existing column’s task. If the demo model cannot initialize in your environment, use the three-dimensional Python example below to exercise column creation independently.

Data lifecycle: this Compose configuration has no persistent database volume. docker compose stop and docker compose start preserve the same container; removing or recreating it can discard runtime data, recipes, tasks, and indexes. Back up important data before down, rebuilds, or recreation, and configure persistence before using this outside a disposable demo.

Screen guide

Vector inventory

Select a namespace and filter by type or index state, then click Inspect on a column. Counters apply to the current page; row totals are intentionally not queried. The Python catalog reads compiled IRIS dictionary metadata through /api/instances/local/namespaces/{namespace}/vectors, displaying dimensions, verified source/model bindings, and indexes. Unknown means provenance is not established: an ordinary vector column does not inherently remember its generator.

Records

Open an asset’s Records tab to browse 25 records per page and click Preview vector for a bounded numeric preview. Managed embeddings show verified source text; unbound vectors do not invent a source mapping. The Embedded Python explorer uses keyset-paged SQL and separate row/preview API requests, keeping full vectors out of the initial response.

Similarity search

Choose the input mode, distance, and Top K, then run the query; try [1, 0, 0] on VectorFixture.Raw, or text vector search on VectorFixture.Managed when its provider is available. Parameterized SQL uses TO_VECTOR or EMBEDDING with native VECTOR_COSINE/VECTOR_DOT_PRODUCT. Expand Query plan & SQL to inspect EXPLAIN: an index’s existence does not prove its use, so the diagnosis may be HNSW used, Full scan, or Undetermined.

Storage & placement

Use Storage & placement to answer “where does this vector live?” through namespace defaults, class storage globals, mappings, database names, and directories; expand Full evidence for details. This read-only view combines IRIS dictionary storage metadata with allowlisted SysAdmin namespace, mapping, and database requests. Unresolved information remains explicit; the screen does not move data or edit database configuration.

Indexes

Open Indexes, review eligibility, enter a name such as DocumentHNSW, select a distance, and click Validate & preview DDL before confirming the exact target. The backend checks fixed-dimension floating-point vector types, storage/ID constraints, and existing indexes; short-lived single-use proposals, catalog revalidation, asynchronous DDL, and auditing guard creation. The validated profile is M=24, efConstruction=100; DotProduct requires normalized vectors, and drop/rebuild are not enabled.

Vector tasks: create a column

Choose an existing destination table, a new column name, a relationship column, and a two-column SQL query: the first value identifies the row, and the second is the text to encode. Select a model or Python generator, test a sample, and publish with exact-target confirmation, then click Run now. The /api/vector-recipes endpoints manage versioned recipes and preflight; VectorAdmin.SeedTask calls Embedded Python to create VECTOR(DOUBLE, n) when needed and update matching destination rows.

The accepted source shape is SELECT Key, Text FROM schema.table, without aliases, joins, expressions, filters, or a trailing semicolon. Source and destination may differ, but relationship values must match; choose a stable unique destination key such as ID. Missing destination keys are counted and skipped: this populates/updates vectors, not new document records.

Vector tasks: Python editor

Select Python function, set dimensions, and edit generate(row, context); Validar sintaxe checks code and Testar em uma amostra executes one sample without saving its vector. The front end calls server-side AST/compile validation with line/column feedback; /api/python/test verifies output shape. The function receives row["key"], row["text"], and context["dimensions"], must return that many finite numbers, and cannot import modules. These restrictions are not a security sandbox: run trusted administrator code only.

Example for 3 dimensions (numeric features for testing, not semantic embeddings):

def generate(row, context):
    text = row["text"] or ""
    return [float(len(text)), float(row["key"]), 1.0]

Vector tasks: monitoring and reruns

Below the creation form, run an existing task and open its detail/history panel to inspect scheduler status plus selected, created, updated, skipped, missing-target, and failed counts. Polling combines SysAdmin task-manager/info/history responses with VectorAdmin.SeedRun evidence. Each run reads the SQL again using the published generator and policy; per-target locks prevent overlapping writers, and checkpoints record progress. Commits occur per row, so earlier writes may survive a failed run.

“Created” counts newly populated vectors, not inserted documents. The current UI creates on-demand tasks and supports repeated Run now executions; it has no recurring-schedule editor, deletion, suspension, or cancellation controls.

Embedding configurations

Open Embedding configs to inspect configuration names, provider classes, dimensions, and allowed public settings before choosing a task’s model. The read-only /api/instances/local/embedding-configs endpoint queries %Embedding.Config and exposes approved settings rather than credentials. Provider creation/editing is outside this screen; a task using a model still writes an ordinary vector column, not an automatically source-bound managed embedding property.

Instance & capabilities

Open Instance & capabilities to check the installed build, operator, permissions, supported actions, and limits when a control is unavailable. The capability endpoint combines installed-version evidence with role checks. Disabled functions are intentional gates: generic record editing, configuration editing, and index drop/rebuild remain unavailable, even though authorized tasks can populate their own vector columns.

How it is built

Plain HTML, CSS, and JavaScript call a same-origin Python WSGI application hosted by IRIS (%SYS.Python.WSGI). Embedded Python reads dictionary metadata and executes SQL; administrative evidence and task operations use the local SysAdmin API. VectorAdmin.SeedTask, an IRIS task class, invokes the Python runner. There is no Flask service, separate application server, or frontend build toolchain.

Python module names in this table are relative to vector_admin/ unless stated otherwise.

SysAdmin API integration

The integration follows the SysAdmin API v2 specification, pinned by https://github.com/Davi-Massaru/iris-vector-management/blob/main/vendor/sysadmin-revision.txt; see the generated API map. Monitoring uses /v2/tasks, /v2/task, /v2/task/info, /v2/task/manager, and /v2/task/history; creation/execution use POST /v2/task and POST /v2/task/run. Vector SQL and custom recipe/run records are application responsibilities, not SysAdmin vector endpoints.

Access and safety

Portal roles supplement native IRIS SQL/database and SysAdmin privileges. Grant only the permissions needed for the chosen tables, embedding providers, and task operations. Mutation flows include authorization, CSRF protection, explicit confirmation, and auditing; syntax validation does not make arbitrary administrator code safe. Review TLS, authentication, persistence, backups, and least privilege before deployment.

Troubleshooting

Verification and further reading

Run the Python tests inside the image:

docker compose exec -T -w /usr/irissys/csp/vector-admin iris python3 -m unittest discover -s tests -v
docker compose config --quiet

With Node.js installed locally, node --check vector_admin/static/app.js checks frontend syntax. Integration/acceptance scripts under tools/ exercise the installed catalog, embeddings, indexes, and tasks; some modify demo data, so inspect them before running against a valued instance. The optional scale fixture is capped at 500,000 rows and is not created by the normal build.

See https://github.com/Davi-Massaru/iris-vector-management/blob/main/spec.md for design and acceptance goals; this README describes the implemented release, not every future specification item.