Technical case study / workflow systems

n8n Automation Atlas

A public workflow collection and ML dataset built from handcrafted, community-curated, and synthetically generated n8n workflows for import, training, and retrieval.

36,985 import-ready workflows Curated JSON workflows with the n8n nodes and connections structure needed for canvas import.
97,000 ML-oriented variants Synthetic node-swap variations generated from 100 workflow archetypes for training and research.
131,648 largest dataset rows Rows in the n8n Mega Workflows Hugging Face release.
69 30-day downloads Combined live Hugging Face API count captured on 2026-07-29: 38 master corpus and 31 mega workflows.
01 / Executive summary

One corpus, three practical surfaces.

The Atlas can be used as an import library, an ML dataset, or a retrieval evidence base for workflow-generating agents.

The project combines directly usable n8n JSON workflows with larger training-oriented variations. Most import-ready files contain the full n8n skeleton, including nodes, connections, parameters, and canvas positions, so they can be copied or imported into an n8n workspace after credentials and environment-specific values are configured.

The broader working corpus is maintainer-reported at more than 2 GB and over 1 million lines of JSON and JSONL across repository data, derived exports, and Hugging Face releases. The two published Hugging Face dataset files currently account for about 1.62 GB of that surface.

02 / Corpus

36,985 workflows intended for direct import.

The import-ready collection is deliberately reported separately from the 97,000 training-oriented node-swap variations. Hover or focus a row to inspect the proportional composition of the directly usable corpus.

Community refined 24,696 / 66.8%
Synthetic v1 8,744 / 23.6%
Initial megapack 2,057 / 5.6%
AI Automation Lab 1,464 / 4.0%
Specialized kits 24 / 0.1%

Counts are taken from the public repository inventory. Total: 36,985 import-ready workflows.

03 / Generation pipeline

Scale came from structure, not random JSON mutation.

The generation pipeline uses functional archetypes and node-equivalency groups. Compatible services can be substituted within a known workflow pattern, then the result can be normalized, checked, deduplicated, and packaged for its intended surface.

  1. 01

    Collect and handcraft

    Combine maintained community templates, handcrafted AI workflows, and specialized integration kits.

  2. 02

    Normalize and verify

    Parse nested JSON, require nodes and connections, preserve license notes, and extract searchable metadata.

  3. 03

    Extract archetypes

    Use Python scripts to derive reusable workflow skeletons and compatible node groups from working examples.

  4. 04

    Generate variations

    Use Python and .mjs generators to expand compatible node combinations while retaining the intended workflow pattern.

  5. 05

    Deduplicate and package

    Remove repeated structures, generate manifests, and export import-ready JSON plus ML-oriented JSONL and Parquet surfaces.

  6. 06

    Explore or retrieve

    Browse through the Vue explorer or retrieve similar workflows from Qdrant to ground an agent with relevant examples.

The repository preserves both Python and .mjs generation tooling, including archetype extraction, synthetic expansion, manifest generation, dataset preparation, embedding, and Qdrant ingestion scripts.

04 / Published datasets

Two Hugging Face releases for different workloads.

The smaller master corpus favors structured browsing and analysis. The mega dataset packages 131,648 instruction-output rows for training and large-scale workflow-generation research.

Dataset Rows Published size Format 30-day downloads
n8n Master Corpus 36,405 50 MB Parquet / structured workflow JSON 38
n8n Mega Workflows 131,648 1.57 GB JSONL instruction-output pairs 31

The combined 69 downloads are a live Hugging Face API snapshot captured on 2026-07-29, not a lifetime total or an estimated average.

05 / Retrieval experiments

Real workflow context for agents, including a useful negative result.

A retrieval layer indexes 36,166 workflows in Qdrant using 768-dimensional Together AI M2-BERT embeddings. The repository reports roughly 50 ms semantic-query latency. The intended use is to give a workflow-generating agent relevant node combinations and complete JSON structures at generation time.

This can be especially useful for smaller or faster models that have less reliable n8n schema recall. Retrieved examples are intended to reduce unsupported node, parameter, and connection guesses, but the effect must be measured per model rather than presented as a guaranteed hallucination reduction.

Baseline Retrieve semantically similar workflows and place the most relevant examples in the agent context.
Corrective RAG experiment Add a retrieval-quality correction step before generation and attempt a second retrieval path when context appears weak.
Observed decision Corrective RAG did not improve the observed workflow-generation performance enough to justify its added latency and complexity.

Evidence boundary: the negative Corrective RAG result is a documented project finding, but the detailed run bundle and scoring table are not yet published. It should be treated as an engineering decision from an early experiment, not a general claim about Corrective RAG.

06 / Results

A reusable data asset, not only a template dump.

36,985 JSON workflows available for exploration and direct import after configuration.
131,648 Rows in the largest public training release.
36,166 Workflow vectors indexed for semantic retrieval.
Apache 2.0 Repository license, with upstream source notes retained separately.
  • For builders: search, filter, download, inspect, and import complete workflow structures.
  • For ML work: train or evaluate workflow-generation models on instruction-output and structured JSON data.
  • For agent systems: retrieve similar workflows to ground a generator in real node and connection patterns.
07 / Tradeoffs

Scale does not remove the need for review.

  • Credentials and environment: imported workflows still require the user's own credentials, webhook URLs, resource IDs, and production testing.
  • Import-ready versus ML-ready: the 97,000 node-swap variants are optimized for training research and are not all presented as production-ready canvas imports.
  • Community sources: upstream source and license notes must remain visible even when workflows are normalized.
  • Retrieval quality: semantic similarity can retrieve structurally relevant but operationally wrong examples, so validators and task-specific evaluation still matter.
  • Next evidence: publish reproducible model comparisons for plain RAG, Corrective RAG, and no-retrieval workflow generation.
08 / Interview explanation

The concise technical version.

I built a workflow corpus that can serve humans importing automations and agents learning how valid n8n systems are structured.

The difficult part was not producing a large directory. It was separating directly usable workflows from ML variations, normalizing inconsistent JSON, preserving licensing context, extracting reusable archetypes, and building retrieval and explorer surfaces around the data. An early Corrective RAG experiment was also valuable because it did not improve the result and prevented extra complexity from becoming permanent architecture.