IronFlow

RAG / WITHOUT THE BLACK BOX

RAG is a workflow. Own every step.

Extract, shape, embed, index and retrieve content in one explicit Lua graph—then run it on a Rust engine you control.

Read the full flowExplore semantic chunking

One graph. Every stage visible.

  1. 01source
  2. 02extract
  3. 03chunk
  4. 04embed
  5. 05index
  6. 06query
document + media nodes
20
AI primitives
6
S3 Vector lifecycle nodes
9
built-in nodes overall
101

THE WHOLE LIFECYCLE

Build the ingestion system, not just the answer step.

RAG quality is decided long before a prompt sees retrieved context. IronFlow keeps those decisions in the graph.

01

Extract

Turn PDF, Word, slides, HTML, captions and images into useful material.

02

Shape

Normalize metadata, filter noise and choose fixed or semantic chunk boundaries.

03

Embed

Generate vectors inside the graph, with the provider and model visible in code.

04

Index

Provision, populate and inspect the vector index as explicit workflow steps.

05

Retrieve

Embed queries, apply metadata filters and return ranked context.

06

Evaluate

Branch, score, log and tear down resources using the same runtime.

FLOW, NOT FRAMEWORK MAGIC

The dependencies are the architecture.

The graph says exactly what runs, what it reads, and what must finish first. Retries, branches and cleanup remain workflow concerns.

Current native vector boundary

IronFlow’s built-in vector lifecycle targets Amazon S3 Vectors today. Other stores can be reached through HTTP, code or custom nodes, but they are not presented as equivalent native integrations.

flow:step("extract", nodes.extract_pdf({
    path = input_path,
    output_key = "document_text"
}))

flow:step("chunk", nodes.ai_chunk_semantic({
    source_key = "document_text",
    output_key = "chunks"
})):depends_on("extract")

flow:step("embed", nodes.ai_embed({
    input_key = "chunks",
    output_key = "vectors"
})):depends_on("chunk")

RUN THE PROOF

Real flows. Not a capability collage.

Every example below is pinned to IronFlow v1.15.0, the same documented release shown across this site.

01

Ingest, query, tear down

A complete S3 Vectors flow from transcript extraction to cleanup.

Open Lua flow ↗02

Semantic chunks to embeddings

Extract a PDF, find semantic boundaries, normalize and embed.

Open Lua flow ↗03

Query expansion

Generate multiple retrieval queries and combine the results.

Open Lua flow ↗04

Retrieval evaluation

Evaluate query results as part of the executable workflow.

Open Lua flow ↗

THE HONEST FIT

Use IronFlow when your RAG pipeline has become software.

It is built for teams that need explicit processing, dependencies and operational control—not a hosted knowledge-base UI.

IRONFLOW FITS

The pipeline is specialized.

You have unusual documents, custom normalization, metadata rules, multiple retrieval passes or cleanup obligations.

LOOK ELSEWHERE

You need a turnkey AI workspace.

If end users need to upload files, manage a knowledge base and publish a chatbot from a visual UI, use a platform built around that surface.

START WITH THE GRAPH

Make the retrieval pipeline inspectable.

Read the executable example, then replace each stage with the providers and rules your workload actually needs.

Open the RAG exampleRead the engine manual