Docs

SDK Reference

The Python package exports both a hosted API client and a projection-first workflow. The important distinction is which one you reach for first.

Choose the right entry point

Export	Role	What you call
`Schift`	Hosted API client	`catalog`, `embed`, `routing`, `bench`, `db`, `query`, `rerank`, `usage`
`Client`	Legacy projection client	`fit()`, `bench()`, `list_projections()`, `get_projection()`
`Projection`	Pure local transform object	`transform()`, `save()`, `load()`
`migrate()` / `client.migrate.run()`	Adapter-driven rewrite	Read vectors from one store and write projected vectors to another

Hosted API modules

pythonfrom schift import Schift

client = Schift()

client.routing.set(
    primary="openai/text-embedding-3-small",
    fallback="google/gemini-embedding-001",
    mode="failover",
)

documents = [
    {"id": "doc-1", "text": "Quarterly revenue report"},
    {"id": "doc-2", "text": "Board meeting summary"},
]

client.db.create_collection("finance-docs", dimension=3072)
client.db.upsert_text("finance-docs", documents, model="openai/text-embedding-3-large")
results = client.query("board summary", collection="finance-docs", rerank=True)
usage = client.usage.get(period="30d", granularity="day")

Module	Public calls	Notes
`catalog`	`list()`, `get(model_id)`	Model metadata lookup
`embed`	`client.embed(text, model, dimensions=None)` and `.batch(...)`	Returns `numpy.ndarray` values
`routing`	`get()`, `set(primary=None, fallback=None, mode=None)`	Routing policy stored server-side
`bench`	`run(source, target, data=None)`	Returns `BenchReport` from the API
`db`	`create_collection`, `list_collections`, `get_collection`, `collection_stats`, `delete_collection`, `upsert`, `upsert_text`	Hosted collection operations
`query`	`client.query(query, collection=None, db=None, model=None, top_k=10, rerank=False, rerank_top_k=None)`	Callable module
`rerank`	`client.rerank(query, documents, top_k=5)`	Callable module
`usage`	`get(period="30d", granularity=None)`	Aggregate account usage

Projection-first workflow

pythonimport numpy as np
from schift import Client

legacy = Client(api_key="sch_your_key_here")

projection = legacy.fit(
    source=np.load("sample-old.npy"),
    target=np.load("sample-new.npy"),
    source_model="openai/text-embedding-3-small",
    target_model="google/gemini-embedding-001",
)

converted = projection.transform(np.load("all-old.npy"))
projection.save("./projection-openai-to-gemini")

Adapters and store-to-store migration

Adapter	Install	Use it for
`NpyAdapter`	Included	Offline `.npy` arrays and batch exports
`PgVectorAdapter`	`pip install schift[postgres]`	PostgreSQL + pgvector tables
`QdrantAdapter`	`pip install schift[qdrant]`	Qdrant collections

pythonfrom schift import migrate
from schift.adapters.pgvector import PgVectorAdapter
from schift.adapters.qdrant import QdrantAdapter

result = migrate(
    source=PgVectorAdapter(
        conninfo="postgresql://user:pass@localhost/app",
        table="documents",
    ),
    sink=QdrantAdapter(
        url="http://localhost:6333",
        collection="documents-v2",
    ),
    projection=projection,
    dry_run=True,
)