Skip to main content
Syft Space is a privacy-preserving SaaS platform that enables you to create secure, AI-powered endpoints by combining datasets, models, and policies. The platform follows a modular architecture where each component serves a specific purpose in the RAG (Retrieval-Augmented Generation) pipeline.

​
Core components

Syft Space is built around four primary concepts that work together to create intelligent, policy-controlled endpoints:

Datasets

Store and search your private data using vector databases like Weaviate or ChromaDB

Models

Connect to AI models (OpenAI, local vLLM) to generate intelligent responses

Endpoints

Combine datasets and models into queryable API endpoints

Policies

Apply rate limiting and access controls to protect your endpoints

​
How components relate

Each endpoint can connect to:
  • One dataset (optional) - for retrieving relevant context
  • One model (optional) - for generating responses
  • Multiple policies - for access control and rate limiting
At least one of dataset or model must be configured for an endpoint to be valid.

​
Response types

Endpoints support three response configurations based on which components are connected:
Returns search results from the dataset without model processing. Useful for document retrieval systems.Configuration: response_type: "raw"
Returns generated text from the model without dataset context. Useful for standalone chat interfaces.Configuration: response_type: "summary"
Combines dataset search with model generation - the full RAG pipeline. Search results are injected as context for the model.Configuration: response_type: "both"

​
Multi-tenancy

All components are tenant-isolated:
  • Each tenant has their own datasets, models, endpoints, and policies
  • Resources cannot be shared across tenants
  • All API operations are scoped to the authenticated tenant

​
Type system

Each component uses a type registry pattern:
  • Dataset types: weaviate, chromadb_local - define how data is stored and searched
  • Model types: openai, vllm_local - define how AI models are accessed
  • Policy types: rate_limit, accounting_guard - define access control rules
Types are registered at startup and provide:
  • Configuration schemas (what fields are required)
  • Validation logic (ensuring configurations are valid)
  • Runtime behavior (how to execute searches, chats, or policy checks)

​
Data flow

When a client queries an endpoint:
  1. Pre-hook policies execute (e.g., rate limit check)
  2. Dataset search runs (if configured) to find relevant documents
  3. Model chat executes (if configured) with search results as context
  4. Post-hook policies execute (e.g., accounting transaction)
  5. Response returned to client
Policy hooks can block requests at any stage by raising a PolicyViolationError.

​
Next steps

Datasets

Learn about dataset types, provisioners, and data ingestion

Models

Understand model types and chat interfaces

Build docs developers (and LLMs) love

Get started for freeTalk to us