Build with NeuroPipe.

Getting Started

NeuroPipe is a Python-first orchestration framework that lets you define AI pipelines as code. This guide walks you through the fundamentals — from installation to deploying your first pipeline.

Prerequisites: Python 3.10+, an active NeuroPipe account with a valid invite code, and API credentials (provided upon activation).

Installation

Install the NeuroPipe SDK via pip:

# Install the core SDK
pip install neuropipe

# Or with all optional dependencies
pip install neuropipe[all]

# Verify installation
neuropipe --version
# >>> neuropipe v0.4.2

Configure your credentials:

# Set your API key
export NEUROPIPE_API_KEY="np_live_xxxxxxxxxxxx"

# Or use the CLI to configure
neuropipe auth login

Pipelines

A Pipeline is the core abstraction in NeuroPipe. It represents a directed acyclic graph (DAG) of processing stages, where each stage can be an LLM call, a retrieval operation, a transformation, or an agent decision.

Creating a Pipeline

from neuropipe import Pipeline, Stage

# Initialize a new pipeline
pipe = Pipeline(
    name="document-qa",
    mode="deterministic",
    version="1.0.0"
)

# Add processing stages
pipe.add_stage(Stage(
    name="extract",
    handler=extract_text,
    retry_policy={"max_retries": 3, "backoff": "exponential"}
))

pipe.add_stage(Stage(
    name="embed",
    handler=generate_embeddings,
    depends_on=["extract"]
))

pipe.add_stage(Stage(
    name="respond",
    handler=llm_generate,
    depends_on=["embed"],
    circuit_breaker={"threshold": 0.85}
))

# Deploy the pipeline
pipe.deploy(replicas=2)
# >>> Pipeline "document-qa" deployed successfully

Execution Modes

• deterministic — Enforces reproducible execution order and caches intermediate results.
• streaming — Processes data in real-time with backpressure control.
• batch — Optimized for high-volume offline processing with checkpoint/resume.

RAG Integration

NeuroPipe provides first-class support for Retrieval-Augmented Generation workflows. The VectorIngestion stage handles document processing, chunking, embedding, and index management.

from neuropipe.rag import VectorIngestion, Retriever

# Configure the ingestion pipeline
ingestion = VectorIngestion(
    source="s3://docs-bucket/enterprise/",
    chunk_size=512,
    chunk_overlap=64,
    embedding="text-embedding-3-large",
    index="enterprise-knowledge",
    incremental=True
)

# Configure the retriever
retriever = Retriever(
    index="enterprise-knowledge",
    top_k=10,
    rerank=True,
    min_relevance=0.72
)

Supported Sources

• S3-compatible object storage
• PostgreSQL with pgvector
• Elasticsearch / OpenSearch
• Local filesystem
• Web crawlers (configurable depth and domain scoping)

Agents

The AgentCluster abstraction lets you deploy multiple autonomous agents with shared state, policy guardrails, and deterministic decision routing.

from neuropipe.agents import AgentCluster, Policy

cluster = AgentCluster(
    name="research-team",
    agents=4,
    reasoning="chain-of-thought",
    fallback="human-in-loop",
    policy=Policy(
        max_tool_calls=20,
        allowed_actions=["search", "retrieve", "summarize"],
        blocked_actions=["delete", "execute_code"]
    )
)

# Run the cluster on a task
result = cluster.run(
    task="Analyze Q4 revenue trends across all regions",
    context=retriever.query("Q4 financial reports")
)

API Reference

Full API documentation is available to authenticated users in the NeuroPipe Dashboard. Below is a summary of the core modules:

Need more detail? Full interactive API docs are available after activation at dashboard.neuropipe.co/docs