Skip to main content
The VectorShift Python SDK targets the same backend as the visual editor — anything you build in the SDK is editable in the platform UI and vice versa. Pick a primitive and start shipping.

Pipeline

A graph of nodes (LLMs, loaders, transforms, knowledge readers) you build, save, and run.

Agent

Tools, typed I/O, and multi-turn conversations via Session.

Knowledge Base

Managed vector + document store you query directly or expose as a tool on an Agent.

Table

Typed schema, AI-generated columns (Pipeline / Agent), and SQL-grade filters.

Transformation

Reusable, typed Python functions you run directly or wire into a Pipeline.

Analytics

Query runs, tokens, costs, latency, errors, and traces across every primitive.

​
Get started in 60 seconds

1

Install

pip install vectorshift
2

Authenticate

export VECTORSHIFT_API_KEY="vs-..."
Generate a key under Account → API Keys.
3

Run your first pipeline

Follow the Quickstart for a five-line LLM pipeline you can paste and run.

​
What you can build

RAG pipelines

Wire a knowledge-base reader into an LLM in a few lines.

Conversational agents

Tools, approvals, and streamed tool events out of the box.

Background batch runs

Kick off long-running jobs and poll for results.

Streaming chatbots

Multi-turn sessions with token-level streaming.

​
Reference

Pipeline reference

Every public method, grouped by topic.

Agent reference

Lifecycle, tools, running, types.

Errors

Full exception hierarchy with status codes.
The SDK ships with type hints and a mypy plugin that catches pipeline-wiring and tool-definition mistakes at static-check time.

​
Writing SDK code

Follow these conventions — every example in these docs does, and they keep generated code correct and re-runnable:
  • Authenticate via the VECTORSHIFT_API_KEY environment variable. Never hardcode the key inline.
  • Build pipelines with the fluent builder — pipeline.add(name="...").llm(...). Don’t use standalone node constructors or pass nodes=[...] to the constructor.
  • Guard with fetch-or-create at the top, then build. Nodes dedup by name, so a stable name= makes re-runs idempotent — don’t pass hardcoded id= values.
  • Set temperature via sampling=SamplingConfig(temperature=...), not as a top-level .llm() argument (where it’s silently dropped).
  • For exact signatures use the primitive’s reference page; for full runnable scripts use its examples pages.