Searchable workflow library & documentation API
The reference for 2,053 automation workflows was a single 71 MB HTML file. It took ten seconds to open, so the team had stopped using it.
- Role
- Architecture and build
- Stack
- Python, FastAPI, SQLite FTS5, Mermaid
- Corpus
- 2,053 workflows, 365 integrations, 29,445 nodes
- Result
- 71MB in 10s became <100KB in under 1s
The 71-megabyte reference
The collection was already valuable: 2,053 n8n workflows, 365 distinct integrations, 29,445 nodes between them. Years of accumulated automation patterns, every one of them a worked example of something somebody needed to do.
It shipped as one static HTML file. Seventy-one megabytes. Every workflow, every node, every parameter, rendered into a single document and handed to the browser in one piece.
- Ten seconds or more before anything appeared on screen.
- Roughly two gigabytes of memory once it had settled.
- Search was
Ctrl+Fover the whole DOM. - On a phone it simply gave up.
So people stopped opening it. The content was in good shape. The problem was access, and no one had treated that as a problem.
What I built
An indexer and a small API in front of it. The workflows stay where they are, as files on disk; everything expensive about reading them happens once, at index time, into SQLite.
Full-text search
SQLite’s FTS5 full-text index does the work. Names, descriptions and integration lists are tokenised once, and a query against two thousand workflows returns in under a tenth of a second on a laptop with no service running behind it except the one process. Filters sit on top of the same index: trigger type (webhook, scheduled, manual, triggered) and a complexity band derived from node count.
Names a human can scan
The filenames were machine-generated and hostile: 2051_Telegram_Webhook_Automation_Webhook.json. The indexer derives a readable title at parse time (“Telegram Webhook Automation”) and detects the integrations by walking the node types, since the filenames aren’t reliable. That is unglamorous string work and it is most of why the list became browsable.
Generated diagrams
Each workflow’s node graph is turned into a Mermaid diagram on request, from the connections already in the JSON. No one has to maintain those diagrams: if the workflow changes, the picture changes with it. The raw JSON can be viewed and downloaded from the same page, since most visitors end up importing a workflow into their own n8n.
The part you’d be right to argue with
SQLite is not a search engine, and if this corpus grows an order of magnitude the right answer becomes Postgres with proper full-text, or Meilisearch, or something else that was built for the job.
- It buys zero-operations. One file, no server to run, no cluster to keep alive, no monthly bill. Someone can clone the repo and have the whole thing up with
pip installand one command. A search cluster would lose that. - The index is disposable. It rebuilds from the files on disk in seconds, and there’s a
/api/reindexendpoint for when it drifts. Nothing of value lives in the database, so nothing of value is lost by throwing it away.
What it does
- Full-text search across 2,053 workflows and 365 integrations
- <100msFull-text search across 2,053 workflows and 365 integrations
- Smaller payload: 71MB of HTML became under 100KB
- 700×Smaller payload: 71MB of HTML became under 100KB
- Memory in use, against roughly 2GB for the page it replaced
- <50MBMemory in use, against roughly 2GB for the page it replaced
What I haven’t solved: search still matches on what a workflow is called and which services it touches. Someone looking for “notify me when a payment fails” has to already know it was built on Stripe and Slack. Semantic search over node graphs is the obvious next move and I haven’t found a version of it yet that beats good naming.