For decades software folks have talked about the triangle: Cost. Time. Quality. Pick two! That era is over... not because the constraint went away, but because a fourth dimension showed up beside it. Leverage. IMO, leverage is NOT agent quality. That's variable and improving fast, exactly the way human quality improved over decades. It's a curve, not a dimension. Leverage is whether your organization can convert what an agent produces into software that customers accept... if not love. This is what it looks like when you have no leverage: The agent writes a change. There are no tests that run, so it cannot tell whether the change worked. There is no environment to deploy to, so nobody else can tell either. It writes correct, idiomatic, modern code against library APIs that might have moved a year ago, then loops endlessly when nothing compiles. And every change it does produce queues behind one human who has to read all of it... that team did not get faster. They got SLOWER. Anyone who has run a factory knows this in their bones. So... four dimensions. And one can STILL only pick three. Leverage gets bought with time and money. Refuse to spend either and you do not escape the trade, you just pay in the fourth currency... quality. That is how you end up shipping agent-written code nobody verified, into a product customers depend on. Eek. The good news: leverage is the cheapest of the four to buy. Tests that actually run. A place to deploy them. Dependencies, standards, queues of tickets... none of that is exotic!!

Article: dev.to/sidswirl/build-your-o… Code: github.com/swirlai/private-l…

We published SWIRL for Backstage, then installed it the way a stranger would. It failed in under thirty minutes. Not the search. The install. Backstage's create-app template quarantines any npm package younger than 72 hours, so the documented yarn add refused our own package. And a search with no matches returned an error page instead of "no results", because a stock app has no TechDocs and our engine treated that as fatal. Both fixed in 0.1.1 the same day. The install docs now carry the one-line yarnrc change, and the engine treats a missing index as empty, not broken. I wrote up the whole thing, gauntlet numbers and all, because the fix is more interesting than the launch. Link in the comments... If you ship anything for Backstage, run the install from the published package... not your dev tree. It is the only test that counts...

Every enterprise I talk to has the same conversation twice a year. Legal will not let the contract data leave the building. Security will not approve a vendor that trains on prompts. Finance has seen the API bill and would like a word. And somebody in the room says: how hard would it be to just run our own? Read more: dev.to/sidswirl/build-your-o…

That is the SWIRL for Backstage index on a laptop, and it is why we did not ship a cluster. The Backstage default engine keeps its index in memory and rebuilds it on every restart. The Postgres engine cannot match part of a word (issue 8835). Elasticsearch support is still on version 7, which reached end of life in January (issue 32795). Tantivy, embedded in one container, with a stemmed field and an n-gram field on every title, passes all nine cases we built from the maintainers' own bug reports. The index lives on disk, so a restart costs nothing. The full table, with the cases the default engine gets wrong and why, is in the comments.

Copy-everything search makes two copies: your content and your permissions. The content copy is a new thing to audit, purge, breach. The permission copy must stay in sync forever. Every minute it's stale is an access-control bug. Zero-copy makes neither. swirlaiconnect.com

TL;DR: "The path to AI value is a management challenge, not a technical one." -- @tdav

You can use your favorite LLM/agent + SWIRL Enterprise to search iManage, M365 + Google Workspace privately (self-host), for much less $ per seat... swirlaiconnect.com/

Backstage portal search returns "team" when you type "tech". That is issue #27339, and it is the default engine doing exactly what it was built to do. We built SWIRL for Backstage to fix it. Real BM25 relevance for the catalog and TechDocs, and GitHub and Confluence results in the same list. One container. No cluster, no JVM. Free, Apache 2.0. Thirty seconds of it working is in the video. The install is three steps; link to documentation below...

claude -p "search SWIRL for our data retention policy and cite your sources" --mcp-config swirl.json 3 tool calls later: 32 sources listed, 288M hits federated, and a cited answer with deep links. No UI. Real session, unedited: youtube.com/watch?v=a6QNmvLZ… #MCP #Claude #AI

How we replaced HubSpot + SalesForce with Claude + PostgreSQL...!! "We run SWIRL for federated search, so the obvious next move was a SearchProvider that folds the CRM into the same search box as everything else..." dev.to/sidswirl/replaced-sal…

One query. 17 live sources. 148 ranked results in 20 seconds. Zero indexing. And when you hit the end: "Fetch more" goes back to every source that has more, live, and re-ranks 31 new results... in 9 seconds. youtube.com/watch?v=HPIsHtYe… #FederatedSearch #EnterpriseSearch

RAG tip: the best retrieval algorithm is sometimes the human. Search returns the web, the news, and your own policy docs. Flip on Select Items, pick the 2 documents that matter, and the AI answers from exactly those. Nothing else sneaks in. youtube.com/watch?v=fCV-fqBv… #RAG #AI

How to Get Every Email to Your Domain in One Gmail Inbox ... Free! You own a domain. Your website lives on it. With Cloudflare's free plan you can make `[email protected]` - any address, invented on the spot - land in your regular Gmail inbox. No Google Workspace subscription, no mail server, no changing registrars...

Enterprise search without enterprise sign-in is a demo, not a product. Register SWIRL in your IdP (Keycloak shown; same flow for Entra ID or Okta), add one Authenticator row, and every search runs with the user's own tokens. youtu.be/BtigrB06ztk #SSO #OIDC

"Just copy the data into a vector DB" is the advice that ends most enterprise search projects before they start. The moment you copy a document out of SharePoint or a case management system, you've also copied a permissions problem you now have to keep in sync forever. Miss one ACL change and someone sees a doc they shouldn't. SWIRL's answer has always been: don't move the data. Search it where it lives, at query time, and enforce each system's own permissions as you go. The implementation cost of that promise is real. SWIRL passes the user's identity through to each connector and let the source system filter, rather than filtering after the fact. Every connector has to resolve *this user's* access against *that system's* model, live, on every query - and still come back fast enough to feel like search. It's more work than a nightly sync. It's also the only version where "who can see what" is never stale. #enterpriseai #rag #knowledgemanagement #security

The catch: a hand-tuned vanilla RAG can match the per-call count. SWIRL gives you that by default, plus version de-dup most stacks lack. Full breakdown, source-verified defaults, and the math: swirlaiconnect.com/token-eff…

Over a versioned enterprise corpus: • one canonical answer, not 5 copies to the model • ~40% fewer input tokens than typical stuff RAG, 2-3× fewer than refine/tree modes • a cost per answer that stays predictable as the corpus grows

The model is only as good as the context you hand it. Most RAG stacks hand it duplicates, superseded drafts, and chunks that don't answer the question. Worse answers, and a bill that creeps as your data grows. We measured what each stack actually sends. Honestly. 🧵

That's what SWIRL 5 does. swirlaiconnect.com