Data and VCS Adapters
Install
Choose the provider extra instead of installing every client:
pip install "briefcase-ai[rag-pinecone,vcs-dvc]==4.4.0"Vector extras are rag-chroma, rag-pinecone, and rag-weaviate. VCS extras
include vcs-dvc, vcs-pachyderm, vcs-ducklake, and vcs-iceberg; adapters
whose clients are dependency-free use the base vcs surface.
Bind retrieval to a source version
from briefcase.rag.vector_stores import VersionedPineconeStore
print(VersionedPineconeStore.__name__)Construct the adapter with index_name, lakefs_repository, and
lakefs_commit; the Pinecone client reads PINECONE_API_KEY when api_key is
omitted.
Every vector adapter stamps indexed data with its lakeFS repository and commit, then filters retrieval to that commit. This prevents a replay from silently using vectors built from newer source files.
Provider coverage
| Surface | Providers |
|---|---|
| Vector stores | Chroma, Pinecone v3, Weaviate v3 |
| Versioned data | DVC, Nessie, Pachyderm, ArtiVC, DuckLake, Iceberg, git-LFS |
| lakeFS workflow | Branch management, staged writes, lineage, snapshots |
All VCS clients implement VcsClientBase. Provider imports are lazy, so one
missing optional client does not prevent other adapters from importing.
Limits
- The Pinecone adapter targets the v3 client API. The Weaviate adapter requires
weaviate-client<4. - Provider credentials, repository permissions, and lifecycle policies remain application responsibilities.
- A recorded commit proves identity, not availability. Retain the referenced data for the full replay window.
API reference
briefcase.rag has the version-aware vector-store
signatures; briefcase.integrations.lakefs has the VCS
client surface the adapters share.
Where this fits
Adapters bind the Store & Query act to systems you already run, so a retrieval records which version of a source it read.