LexisLocal

Privacy-first, 100% offline PDF intelligence for your desktop.

LexisLocal ingests your PDFs, extracts text and definitions, chunks and embeds everything locally, and lets you ask natural-language questions about your documents — all without a single byte leaving your machine. No API keys, no telemetry, no cloud. The only network call is to a llama-server process on your own localhost.

Features

• Local RAG Chat — Ask questions about your documents. Answers are grounded in the actual text via embedding similarity search (M-TREE vector index) with BM25 full-text fallback.
• Pixel-Perfect PDF Viewer — A canvas-rendered PDF with a transparent, natively-selectable text overlay via PDF.js. Every text span lines up 1:1 with its canvas bitmap.
• Smart Definitions — Key terms and explanations extracted automatically on ingest via LLM, displayed as hover cards over the PDF text layer.
• Anomaly Detection — Flags contradictions, missing clauses, and unusual language in your documents.
• Liquid Navigation — Answers include the source page number; clicking jumps the PDF viewer directly to that page.
• Auto-Spawned AI — The app boots its own llama-server sidecar on a free port and shuts it down on exit.
• Embedded Database — SurrealDB with SurrealKv engine. Zero configuration, no separate server process.
• Section Extraction — Pure-regex section heading detection with cross-reference tracking between sections.

Quick Start

# Prerequisites: Rust, Node.js 18+, llama-server on PATH + a GGUF model

git clone https://github.com/your-org/lexis-local.git
cd lexis-local
npm install

# Point at your GGUF model (defaults to ~/.cache/lexis/model.gguf)
export LEXIS_MODEL_PATH=/path/to/your/model.gguf

npm run tauri dev

See CONTRIBUTING.md for detailed setup instructions and development commands.

Requirements

• Rust (latest stable) + Tauri 2 system deps
• Node.js 18+
• llama-server binary (from llama.cpp) on $PATH
• A GGUF model supporting /v1/embeddings and /v1/chat/completions (e.g. nomic-embed-text-v1.5, 768-dim)

Architecture

PDF bytes ─▶ pdf_extract ─▶ documents table
                              │
                              ├─▶ chunk_text (1024/128) ─▶ embed() ─▶ chunks (+ vector, page)
                              └─▶ extract_definitions ─▶ definitions table

ask(question) ─▶ embed ─▶ M-TREE KNN search ─▶ context ─▶ chat ─▶ answer (+ page)

Stack: Tauri 2 (Rust) · React 19 / TypeScript / Tailwind 4 · SurrealDB (SurrealKv) · PDF.js 6 · llama.cpp (sidecar)

Documentation Hub

Document	Description
docs/architecture.md	System architecture with Mermaid diagrams and data flow
docs/api-reference.md	Tauri commands, HTTP endpoints, and SurrealDB schema
CONTRIBUTING.md	Developer onboarding, setup, and code conventions
CODE_OF_CONDUCT.md	Community guidelines
plan.md	Phased build plan and current status
CLAUDE.md	Agent guide and architecture constraints

Development

npm run dev            # Frontend only (Vite hot-reload)
npm run build          # tsc typecheck + Vite production build
npx tsc --noEmit       # Frontend typecheck only

cd src-tauri && cargo check -j 4 && cargo test -j 4 && cd ..

Low-RAM note: use -j 4 with cargo on machines with <16 GB to avoid OOM.

License

MIT