Drop your documents into a group: manuals, papers, contracts, scanned files. Then ask questions in plain language. LEXIS finds the right passages and writes a grounded answer, with the sources one click away. Nothing leaves your machine: no API keys, no cloud, no cost per question.
Employees are eligible for 12 weeks of paid parental leave following birth or adoption, starting after 6 months of continuous employment [1]. Leave can be taken continuously or in two blocks within the first year [2].
Grab the disk image, drag LEXIS into Applications. That's the whole install. No Homebrew, no terminal, no database to configure.
Create a group and drop in manuals, papers, contracts, scanned files. Indexing is pure text processing. A thousand documents take minutes, not hours.
Ask in plain language. LEXIS searches, reranks, and writes a grounded answer from the passages it found, with every source one click away.
Most document chat tools embed everything into vectors and search those. LEXIS uses classic full text search, the same family of algorithms behind real search engines, and applies models only where they help.
A small local model refines your question and picks synonyms that fit the meaning before searching.
Full text lexical search over the index. Exact terms (part numbers, codes, names) always match.
A tiny embedding model reorders candidates by meaning, so the best passage leads.
A local chat model writes the answer from those passages and nothing else. Fully grounded.
The index never needs rebuilding when models change. Swap models freely; your index stays valid.
The standard local RAG recipe embeds every chunk of every document into vectors at ingest time. LEXIS searches by words instead and applies models only at the edges.
| LEXIS lexical + models at the edges | Embedding RAG vectors | |
|---|---|---|
| Ingest speed | Word counting: thousands of passages/sec | Neural net over every chunk: ~100× slower |
| Exact terms | Exact match, always found | Compressed into "meaning": rare identifiers blur |
| Pure paraphrase | Covered by query rewrite, synonyms & reranker | Strongest case |
| Changing models | Index untouched | Model specific vectors: embed everything again |
| Why this answer? | Visible search terms and passages | A similarity score |
| Index at scale | Postgres on disk | Vector index, typically in RAM |
Questions that are pure paraphrase are the one real advantage vectors hold. On our benchmarks, the questions where a full vector index would still have helped came to about 2%.
The installer is ~80 MB. On first launch, LEXIS downloads the models it needs (~5.1 GB) with a progress bar. Interrupted downloads resume. Your documents, config, and database live in ~/Library/Application Support/LEXIS/; the app runs its own private database that only your macOS user can reach.
After setup, LEXIS works fully offline: on a plane, in a vault, anywhere.
No. Everything (the search index, the models, the database) runs locally on your machine. There are no accounts, no API keys, and no network calls after the initial model download.
No. The models ship with the app's first launch download and run on your hardware. There is no cost per question and nothing to sign up for.
No. LEXIS is Apple Silicon only for now. The local models are built and tuned for Apple Silicon Macs.
The installer is about 80 MB. On first launch the app downloads ~5.1 GB of models (with progress and resume). Plan on ~11 GB of disk total for the app, models, and database.
Delete LEXIS.app from Applications, then delete ~/Library/Application Support/LEXIS/ to remove your models, documents index, and settings. Nothing else is touched.
Free to download. Yours to keep. Nothing leaves your Mac.
Download LEXIS for macOS