nebius/nebius-partner-cookbook

nebius/nebius-partner-cookbook

Language: Python

License: MIT

Stars: 4

Forks: 0

Open issues: 1

Created: 2026-06-09T15:34:05Z

Pushed: 2026-06-11T15:45:05Z

Default branch: main

Fork: no

Archived: no

README:

Agent Blueprint Recipes

Production-grade recipes for building AI agents on Nebius.

---

Agent Blueprint Recipes is a curated collection of runnable, production-shaped code for building AI agents on Nebius. It comes in two tiers. Recipes are small, sequenced FastAPI applications — typed, observable, containerized, and tested — that you can clone, configure, and deploy in five minutes. Blueprints are complete, deployable reference applications that show what a real agent looks like at full scale.

It is for engineers who have called an LLM API from a script and want to know what the gap looks like between that script and something they would actually ship.

What "production-shaped" means here:

• Type hints on every signature and Pydantic models on every boundary
• Structured JSON logging with request IDs
• Prometheus metrics and healthcheck endpoints out of the box
• Rate limiting, CORS, security headers, graceful shutdown
• A multi-stage Dockerfile and a one-command deployment recipe
• Tests that pass with no network access

Each recipe is autonomous — cloning a single cookbook directory is enough to run it, with no shared base packages and no implicit dependencies on the rest of the repo.

The recipes are also a sequence. They form a narrative arc — Foundation → Knowledge → Grounding → Orchestration → Thread Memory → User Memory → Observability → Guardrails → Actions → Simulation — and each one assumes the concepts of the one before it: #2 builds on #1, #3 on #2, and so on. You *can* run any cookbook on its own, but the documentation is written for a reader following them in order.

Blueprints sit outside that sequence. Where a recipe teaches one concept, a blueprint is a finished application — larger, opinionated, and deployable as-is — that combines many of those concepts into something you could put in front of users. They may diverge from the recipe stack (a React frontend, a LangGraph deployment, a bundled data corpus) and live under [blueprints/](./blueprints/).

Recipes

| # | Recipe | Stack | Difficulty | Reading | | --- | --- | --- | --- | --- | | 01 | [Foundation — Your First Agent on Nebius](./cookbooks/01-first-agent-on-nebius/) | nebius-agentkit fastapi structlog prometheus docker | beginner | 8 min | | 02 | [Knowledge — Domain Knowledge with Pinecone Nexus](./cookbooks/02-domain-knowledge-pinecone-nexus/) | nebius-agentkit pinecone-nexus fastapi structlog prometheus docker | intermediate | 13 min | | 03 | [Grounding — Real-Time Data with Tavily](./cookbooks/03-real-time-data-tavily/) | nebius-agentkit pinecone tavily fastapi openai-sdk sse docker | intermediate | 12 min | | 04 | [Orchestration — Stronger Agents with LangChain and LangGraph](./cookbooks/04-stronger-agents-langchain-langgraph/) | nebius-agentkit langchain langgraph fastapi openai-sdk structlog prometheus docker | intermediate | 13 min | | 05 | [Thread Memory — Short-Term Memory with LangChain](./cookbooks/05-short-term-memory-langchain/) | nebius-agentkit langchain langgraph fastapi openai-sdk structlog prometheus docker | intermediate | 12 min | | 06 | [User Memory — Long-Term Memory with LangChain and Postgres](./cookbooks/06-long-term-memory-langchain-postgres/) | nebius-agentkit langchain postgres langgraph fastapi openai-sdk structlog prometheus docker | advanced | 13 min | | 07 | [Observability — Observability with LangSmith](./cookbooks/07-observability-langsmith/) | nebius-agentkit langsmith langchain fastapi openai-sdk structlog prometheus docker | advanced | 14 min | | 08 | [Guardrails — Adding Guardrails with LangChain](./cookbooks/08-adding-guardrails-langchain/) | nebius-agentkit langchain guardrails fastapi structlog prometheus docker | advanced | 14 min | | 09 | [Actions — Making Actions with MCP and Stripe](./cookbooks/09-actions-with-mcp-stripe/) | nebius-agentkit mcp stripe langchain langgraph langsmith fastapi structlog prometheus docker | advanced | 18 min | | 10 | [Simulation — Testing Before Production with Snowglobe](./cookbooks/10-testing-before-production-snowglobe/) | nebius-agentkit snowglobe langsmith mcp stripe fastapi structlog prometheus docker | advanced | 14 min |

Blueprints

Complete, deployable reference applications. Larger than recipes and not part of the sequence — clone one and run it end-to-end.

| Blueprint | Stack | Integrations | Run | | --- | --- | --- | --- | | [Compliance — Sentinel — Regulatory Compliance Auditor](./blueprints/sentinel-compliance-auditor/) | langgraph pinecone tavily langsmith jira fastapi react openai-sdk deepagents | pinecone tavily langsmith jira mcp | ~23 min |

Stack

The recipe stack is deliberately small and boring. Every recipe makes the same technical choices so that what you learn in one cookbook transfers directly to the next. Blueprints are not bound by this — a blueprint picks whatever a real application of its kind would use (e.g. a React frontend, LangGraph for orchestration and deployment, Pinecone for knowledge), so its stack is documented in its own README.

Recipes (Python)

| Layer | Choice | Why | |---|---|---| | Language | Python 3.12 | Modern type system, broad ecosystem | | Package manager | uv | Fast, reproducible, lockfile-first | | HTTP framework | FastAPI | Async, Pydantic-native, OpenAPI for free | | ASGI server | Uvicorn | The boring standard | | LLM SDK | openai-python pointed at Nebius | Familiar API, no vendor lock-in to switch later | | Config | pydantic-settings | Type-safe env validation at boot | | Logging | structlog (JSON in prod) | Structured logs you can actually query | | Metrics | prometheus-client | /metrics endpoint, standard format | | Rate limit | FastAPI middleware + Redis fallback | Per-IP daily quotas, Redis when configured, in-memory locally | | HTTP client | httpx...