Schift›Docs›Quickstart

Quickstart

Get from zero to your first embedding migration in under 5 minutes.

1. Install

bash

pip install schift

2. Get an API key

Create a free account at [app.schift.io](/app) and generate an API key from the Dashboard. Keys use the format sch_xxx.

3. First embed call

python

from schift import Schift

client = Schift(api_key="sch_your_key_here")

# Embed a single text
vec = client.embed("quarterly revenue report",
    model="openai/text-embedding-3-large")

print(vec.shape)  # (3072,)

4. First migration (5-minute tutorial)

The key insight: Schift learns a projection matrix between two embedding spaces. You only need ~1,000 paired samples — no raw text required for the bulk migration.

python

import numpy as np
from schift import Schift
from schift.adapters import PgVectorAdapter

client = Schift(api_key="sch_your_key_here")

# Step 1: Embed a small sample with both models
# (you need ~1,000 representative documents)
sample_texts = load_sample_docs()  # your 1,000 docs
source_emb = embed_with_ada002(sample_texts)   # np.ndarray (1000, 1536)
target_emb = embed_with_3_large(sample_texts)  # np.ndarray (1000, 3072)

# Step 2: Fit the projection matrix
proj = client.migrate.fit(
    source="openai/text-embedding-ada-002",
    target="openai/text-embedding-3-large",
    sample_ratio=0.1
)
print(proj["id"])   # proj_a1b2c3d4e5f6

# Step 3: Benchmark quality before migrating
report = client.bench.run(
    source="openai/text-embedding-ada-002",
    target="openai/text-embedding-3-large",
    data="./eval_queries.jsonl"
)
print(report.verdict)    # "SAFE"
print(report.recovery)   # 0.962

# Step 4: Migrate all vectors
source = PgVectorAdapter(dsn="postgresql://...", table="embeddings_old")
sink   = PgVectorAdapter(dsn="postgresql://...", table="embeddings_new")

result = client.migrate.run(
    projection=proj["id"],
    db="pgvector://localhost/mydb",
    on_progress=lambda r: print(f"{r['progress']:.0%}")
)
print(f"Migrated {result['total_vectors']:,} vectors")
# Migrated 1,234,567 vectors

✓

Typical migration: 1M vectors in ~5 minutes. No downtime. No re-embedding cost.