Docs

Core Concepts

Schift separates the default Schift Embed 1 path from migration-first vector conversion so each surface stays explicit.

Embedding vectors from different models are not natively compatible. Even when dimensions match, the geometry changes with the model.

Two product surfaces

Surface	Primary objects	Use it for
Schift Embed 1 default path	`Schift`, `catalog`, `embed`, `db`, `query`, `usage`	New embedding calls, hosted collections, and canonical-space retrieval
Migration-first runtime	`Client`, `Projection`, `migrate()`, adapters	Learning a projection matrix and rewriting stored vectors without raw text

Projection matrix

textsource vectors (N, src_dim)
        -> learned matrix W (src_dim, tgt_dim)
        -> projected vectors (N, tgt_dim)

Schift learns W from paired examples embedded by both models. At runtime, migration is just a matrix multiply on the existing vectors.

Adapter pipeline

textAdapter(source store)
        -> Projection.transform(batch.embeddings)
        -> Adapter(sink store)

Property	Typical range
Paired samples needed	500-2,000
Supported migration shapes	1536->768, 3072->1024, and more
Runtime cost	Sub-millisecond per vector
Bulk raw text required	No

Operational note

Projection is the brownfield path. It preserves retrieval quality during provider changes, while Schift Embed 1 is the greenfield default path into the same space.

Temporal Constraints

Every vector can carry an event_time metadata field (epoch milliseconds). The temporal query parameter filters results by time, enabling point-in-time retrieval and time-series search.

Mode	Behavior	Required Fields
`before`	event_time < temporal_start	temporal_start
`after`	event_time > temporal_start	temporal_start
`between`	temporal_start <= event_time <= temporal_end	temporal_start, temporal_end
`as_of`	event_time <= temporal_start (snapshot at a point in time)	temporal_start
`latest`	Sort by most recent event_time	None

Vectors without an event_time field are excluded from all temporal queries. The latest mode over-fetches candidates by 20x before sorting to ensure quality results.

pythonfrom schift import Schift

client = Schift(api_key="sch_xxx")

# Documents from the last 24 hours
import time
one_day_ago = int((time.time() - 86400) * 1000)
results = client.collections.search(
    "news", "latest AI developments",
    temporal="after", temporal_start=one_day_ago,
)

# Snapshot: what was indexed as of Jan 1, 2026
results = client.collections.search(
    "contracts", "termination clause",
    temporal="as_of", temporal_start=1735689600000,
)

Edges & Graph

Schift buckets support typed, weighted edges between vectors, turning a flat vector store into a knowledge graph. Edges encode domain relationships like supersedes, contradicts, or caused_by.

Relation	Use Case
`related_to`	General association (default)
`supersedes`	Newer version replaces older document
`contradicts`	Two documents conflict
`caused_by`	Causal chain between events or issues
`is_a`	Taxonomy / classification hierarchy
`has_child`	Parent-child tree structure
`follows`	Sequential ordering (e.g., legal articles)

Edges are stored in a compact CSR format with WAL-backed durability. Each edge carries a weight (0.0-1.0) and can be queried by direction (outgoing, incoming, or both) and relation type.

bash# Add citation edges between legal documents
curl -X POST https://api.schift.io/v1/buckets/{bucket_id}/edges \
  -H "Authorization: Bearer $SCHIFT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"edges": [
    {"source": "article-301", "target": "article-42", "relation": "supersedes", "weight": 1.0},
    {"source": "ruling-2024-001", "target": "article-301", "relation": "caused_by", "weight": 0.8}
  ]}'

# Query outgoing edges from a node
curl https://api.schift.io/v1/buckets/{bucket_id}/edges/article-301?direction=outgoing \
  -H "Authorization: Bearer $SCHIFT_API_KEY"