ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
pgvector — PostgreSQL vector search for projects
Official pgvector 0.8.6 on PostgreSQL 17 Bookworm, pinned by multiarch image digest. PostgreSQL license. Upstream: https://github.com/pgvector/pgvector/tree/v0.8.6.
This extension provides relational storage plus vector types, distance operators and indexes. It does not generate embeddings, select a language model or automatically ingest Portal conversations. The project supplies vectors with dimensions matching its chosen embedding model.
Credentials and connection
Configure PGVECTOR_ADMIN_PASSWORD and PGVECTOR_APP_PASSWORD as two distinct random 64-character hexadecimal secrets. The initialization creates database vectors, installs vector, and creates the restricted ods_app login owning schema project. Use this application account in projects; the ods_admin bootstrap account is a PostgreSQL superuser and is for administration only.
Host clients: localhost:11098; other ODS containers: pgvector:5432. Database vectors, username ods_app. Supply the password through the client's secret mechanism, not a committed connection string. Loopback is the only published listener. No HTTP page is provided; connect with a PostgreSQL client or explicitly configured database UI.
An application can create project.documents (id bigint GENERATED ALWAYS AS IDENTITY PRIMARY KEY, embedding vector(384)) and query cosine similarity using embedding <=> $1. Choose dimensions from the actual embedding model, not this example. Tables, indexes and documents are never seeded automatically. The shared application account is not tenant isolation; distinct projects needing isolation require separate roles/schemas and grants provisioned by the administrator.
Persistence and upgrades
pgvector-data persists PostgreSQL's /var/lib/postgresql/data. Initialization SQL and both passwords apply only on the first empty volume. Later environment edits do not rotate existing database passwords: change the role password through an authenticated administrator connection, then update configured clients/health secret. Never delete the volume to resolve a credential mismatch.
Back up with PostgreSQL-aware tools (pg_dump plus roles, or consistent physical backups). Do not copy a running data directory as an ordinary folder backup. A PostgreSQL major-version upgrade requires a supported dump/restore or pg_upgrade procedure, not just changing the image tag. Removing the extension does not authorize erasing its data.
Upstream startup is retained to initialize/chown the volume and drop to the postgres user. Host SQL uses SCRAM authentication; no trust override. Resources: 1 GiB, two CPUs, 128 MiB shared buffers, 50 connections and 128 MiB shared memory. Index builds may require deliberate resource adjustments for larger datasets.
The native health probe authenticates over TCP as ods_app, checks extension version and evaluates a vector distance without creating rows. Failed credentials or missing initialization therefore cannot report healthy solely because a port is open.
Compatibility and validation
Official image includes Linux amd64 and arm64. Runs through Docker Engine/Desktop on Windows, Linux and macOS; no GPU or inference dependency. Configuration/schema checks do not prove deployment: image build, initialization, actual SQL persistence, restore and host runtime validation remain pending. No database or model was started during recipe preparation.