RAG pipelines

Build RAG pipelines that answer from your data.

Embed documents with BGE-M3 on a GPU instance, keep vectors in Qdrant, and answer with DeepSeek V3.2 or Qwen3 through the API.

L4from $1.36/hr

What you can build

Three jobs people run here. Each on a product you can start today.

Instance

Run BGE-M3 or E5 on an L4 or L40S instance and embed your corpus in batches, at the speed the GPU allows.

Storage

Run Qdrant, Weaviate or Chroma on an instance and keep the index on a volume that outlives it.

Inference

Send retrieved passages to DeepSeek V3.2, Qwen3 or Llama 4 Scout in the OpenAI format, and switch models with one field.

Examples

Connected by the line, Claude Code or Codex calls the Runcrate tools for you: launch, copy files, run, delete.

Hand this to your agent

Add a Runcrate model to this project: run curl -fsSL"https://runcrate.ai/agents.md?intent=build"and follow it.

Your agent asks for an API key. Create a key

Works inClaude CodeCodexOpenCodeClineRoo CodeKilo CodeCursor

example

Models and tools

Name	Role	Runs on
BGE-M3	Multilingual dense embeddings	Instance
Qdrant	Vector database	Instance
LlamaIndex	Chunking and retrieval	Instance
DeepSeek V3.2	Answer generation	Inference API
Qwen3 235B Instruct	Long-context answers	Inference API
Llama 4 Scout	Open-weight generation	Instance or API

How it works

01
Embed your documents
Chunk your documents with LlamaIndex or LangChain and embed them with BGE-M3 or E5 on a GPU instance.
02
Store and retrieve
Index the vectors in Qdrant, Weaviate or Chroma on an instance, with the data on a volume. Query it for the closest passages.
03
Generate the answer
Pass the passages to DeepSeek V3.2, Qwen3 or Llama 4 Scout through the Inference API, billed per token.

Pricing

GPUs are billed for the time they run: by the second on most GPUs, by the minute on a few.
A multi-node cluster for a month or longer: reserve one, and see every offer side by side.

Name

As billed