pgvector

Type: pgvector

Uses PostgreSQL with the pgvector extension for cosine similarity search via the <=> operator.

---

Docker quickstart

docker run -d -p 5432:5432 \
  -e POSTGRES_PASSWORD=secret \
  pgvector/pgvector:pg16

The pgvector/pgvector image includes the extension pre-installed. For standard Postgres, install the extension manually:

CREATE EXTENSION IF NOT EXISTS vector;

---

Configuration

- name: embedder
  type: embedding/openai
  metadata:
    base_url: http://localhost:11434/v1
    model: nomic-embed-text
    dimensions: "768"

- name: pg-docs
  type: pgvector
  metadata:
    dsn: postgres://user:secret@localhost:5432/mydb
    table: daimon_documents
    embedder: embedder
    dimensions: "768"

Metadata keys

Key	Default	Description
dsn	—	Required. Postgres connection string
table	daimon_documents	Table name (auto-created if absent)
embedder	—	Name of a declared embedding/openai component
dimensions	"1536"	Vector dimensions (must match the embedder)

---

Auto-created table schema

CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE IF NOT EXISTS daimon_documents (
    id        TEXT PRIMARY KEY,
    content   TEXT NOT NULL,
    metadata  JSONB,
    embedding vector(768)
);

CREATE INDEX IF NOT EXISTS daimon_documents_embedding_idx
    ON daimon_documents USING ivfflat (embedding vector_cosine_ops);

The table and index are created on first startup. If the table already exists with a different schema, startup will fail.

---

Notes

• Scores are 1 - cosine_distance, so 1.0 is an exact match.
• The IVFFlat index requires at least a few hundred rows before it improves query performance; for small datasets the sequential scan is used automatically.
• dimensions must match the embedder's output exactly. Changing dimensions requires dropping and recreating the table.