ContextQMD
Back to Dify

VDB providers

api/providers/vdb/README.md

1.14.03.5 KB
Original Source

VDB providers

This directory contains all VDB providers.

Architecture

  1. Core (api/core/rag/datasource/vdb/) defines the contracts and loads plugins.
  2. Each provider (api/providers/vdb/<backend>/) implements those contracts and registers an entry point.
  3. At runtime, importlib.metadata.entry_points resolves the backend name (e.g. pgvector) to a factory class. The registry caches loaded classes (see vector_backend_registry.py).

Interfaces

PieceRole
AbstractVectorFactoryYou subclass this. Implement init_vector(dataset, attributes, embeddings) -> BaseVector. Optionally use gen_index_struct_dict() for new datasets.
BaseVectorYour store class subclasses this: create, add_texts, search_by_vector, delete, etc.
VectorTypeStrEnum of supported backend string ids. Add a member when you introduce a new backend that should be selectable like existing ones.
DiscoveryLoads dify.vector_backends entry points and caches get_vector_factory_class(vector_type).

The high-level caller is Vector in vector_factory.py: it reads the configured or dataset-specific vector type, calls get_vector_factory_class, instantiates the factory, and uses the returned BaseVector implementation.

Entry point name must match the vector type string

Entry points are registered under the group dify.vector_backends. The entry point name (left-hand side) must be exactly the string used as vector_type everywhere else—typically the VectorType enum value (e.g. PGVECTOR = "pgvector" → entry point name pgvector; TIDB_ON_QDRANT = "tidb_on_qdrant" → tidb_on_qdrant).

In pyproject.toml:

toml
[project.entry-points."dify.vector_backends"]
pgvector = "dify_vdb_pgvector.pgvector:PGVectorFactory"

The value is module:attribute: a importable module path and the class implementing AbstractVectorFactory.

How registration works

  1. On first use, get_vector_factory_class(vector_type) looks up vector_type in a process cache.
  2. If missing, it scans entry_points().select(group="dify.vector_backends") for an entry whose name equals vector_type.
  3. It loads that entry (ep.load()), which must return the factory class (not an instance).
  4. There is an optional internal map _BUILTIN_VECTOR_FACTORY_TARGETS for non-distribution builtins; normal VDB plugins use entry points only.

After you change a provider’s pyproject.toml (entry points or dependencies), run uv sync in api/ so the installed environment’s dist-info matches the project metadata.

Package layout (VDB)

Each backend usually follows:

  • api/providers/vdb/<backend>/pyproject.toml — project name dify-vdb-<backend>, dependencies, entry points.
  • api/providers/vdb/<backend>/src/dify_vdb_<python_package>/ — implementation (e.g. PGVector, PGVectorFactory).

See vdb/pgvector/ as a reference implementation.

Wiring a new backend into the API workspace

The API uses a uv workspace (api/pyproject.toml):

  1. [tool.uv.workspace] — members = ["providers/vdb/*"] already includes every subdirectory under vdb/; new folders there are workspace members.
  2. [tool.uv.sources] — add a line for your package: dify-vdb-mine = { workspace = true }.
  3. [project.optional-dependencies] — add a group such as vdb-mine = ["dify-vdb-mine"], and list dify-vdb-mine under vdb-all if it should install with the default bundle.