Xns Ai Cookbooks

Working recipes for AI pipelines on XNS — S3-compatible storage with no per-read charge.

Every time a RAG pipeline re-reads a corpus, a training job pulls a checkpoint, or an agent fetches a shared artifact, the storage side of that read is free — compute and model API costs are yours as usual. These recipes target the workflows where repeated reads dominate the storage bill.

Each recipe states its limitations explicitly. These are starter recipes — single-process, happy-path — and each one says exactly where that stops being enough. The first four build their index in memory and lose it at exit; Persistent Index is the one that does not, and is where to go when re-embedding on every run stops being acceptable.

Recipes

Recipe	Pipeline	Status
Multimodal RAG	Video speech + frames → transcripts + vision captions cached in XNS → query	Ready
Agentic Document Parsing	PDFs/spreadsheets → local Docling parse → structured JSON, cached per document	Ready
Fine-Tune Checkpointing	Model weights ↔ GPU clusters via S3 multipart	Ready
Agent Workspace	CrewAI agents exchanging artifacts through a shared bucket	Ready
Persistent Index	Qdrant collection snapshotted to XNS; a cold container restores and queries without re-embedding	Ready

Each recipe includes three things:

1. Architecture blueprint — a text diagram showing where XNS sits in the pipeline and why reads being free matters at that point.
2. Runnable script — Python, under 60 seconds on a laptop once prerequisites are in place.
3. Config block — JSON to wire XNS into Claude Desktop, Cursor, or any MCP client.

Prerequisites

You need a running XNS Relayer (the S3 gateway). Two paths:

Docker Compose (if you have a Linux host with Docker):

git clone https://github.com/xns-cloud/relayer-quickstart
cd relayer-quickstart
docker compose up -d

AI-assisted setup (the MCP server walks you through it):

claude mcp add relayer -- npx @xns-cloud/relayer-mcp@latest

Then ask the agent to "set up XNS storage." It handles prerequisites, account registration, install, and credential provisioning. Either path writes ~/.xns/credentials, which every recipe reads automatically.

Links

• XNS — self-hosted S3-compatible storage, zero egress
• relayer-quickstart — Docker Compose setup
• relayer-mcp — MCP server for AI code assistants
• langchain-xns — LangChain loaders and stores
• S3 compatibility — conformance test results

License

Apache-2.0