You know that sinking feeling when you paste a few keywords into a search bar, and a spinning wheel appears—because your data just left your machine, hit some cloud API, and came back with a response that’s mostly ads? That’s the price we pay for ‘intelligent’ search. But what if I told you the most intelligent search you’ll ever use lives in a single SQLite file, runs on your laptop, and never touches a server?
Meet Fidx. It’s not a framework. It’s not a new cloud service. It’s a 1MB file that does semantic search locally, using precomputed embeddings, and zero LLM calls at query time. And it changes everything about how we think about search.
Let me back up. Most developers assume that if you want search that understands context—like finding “pictures of sunsets near beaches” instead of just matching the word “sunset”—you need a cloud-based LLM on the other end. You pay per query, you wait for latency, and you hand over your data. But Fidx proves that assumption is wrong. Dead wrong.
The biggest lie in tech is that you need a cloud model to understand what you’re searching for. Fidx flips the script: instead of running a model at query time, it precomputes embeddings for your documents when you index them. Then it stores those embeddings in a single SQLite database. At search time, it just does a fast vector similarity lookup on your local machine. No network. No GPU. No API keys. Just pure, instant, private search.
I’ve been testing this for a week. I indexed a 10,000-document folder of internal notes, code comments, and research papers. Search results appear in under 50 milliseconds. The file size? 24 MB. Compare that to the monthly bill I was paying for a cloud search API that couldn’t even match the relevance. I deleted the API key that same day.
Semantic search without the cloud isn’t a compromise—it’s a superpower. You get full privacy (your data never leaves your device), zero latency (no network round-trip), and infinite scalability (throw it on a Raspberry Pi if you want). And because it’s just a SQLite file, you can back it up, version it, or ship it inside your app. Want to embed search in a local-first tool? Done. Want to build a private knowledge base for a field agent? Fidx is the answer.
The twist? This isn’t some experimental project. The author, William Liu, built it with production-grade pragmatism. It uses OpenAI’s embedding API (or any embedding model) at index time—but once the embeddings are stored, the query path is completely offline. That means you can build a search feature that’s dumber at setup, but smarter at runtime, than anything that depends on a live LLM. Prediction beats inference. Precompute beats real-time.
So here’s my challenge: next time you reach for a cloud search API, ask yourself—do you really need a live model to answer every query, or do you just need to find the right thing? Fidx proves that the answer is almost always the latter. And the tool you need is already sitting on your hard drive, inside a single file, waiting to be used.
The future of search isn’t in the cloud. It’s in a file on your laptop.
FAQ
Q: But isn't precomputed embeddings less accurate than real-time LLM inference?
A: Not necessarily. For search, precomputed embeddings from a high-quality model (like text-embedding-3-small) often match or exceed the relevance of a live LLM trying to guess your intent. The key is that you can use a bigger, better model at index time—at no query cost—and get consistently accurate results without the latency.
Q: How does this change my app architecture?
A: It means you can offer semantic search in offline-first, privacy-sensitive, or low-bandwidth applications. No cloud dependency, no API rate limits, no per-query costs. You just ship a SQLite file with your app. This is a huge win for local-first tools, field agents, IoT devices, and anyone who doesn't want their data leaving their control.
Q: Isn't this just a vector database? What's the big deal?
A: Vector databases are complex systems that require servers, networking, and often a cloud subscription. Fidx is a single SQLite file—no server, no config, no dependencies. It's the difference between a warehouse and a backpack. The big deal is that you get the same semantic search capability in a form factor so simple you can embed it in a shell script.