litesearch

A fully-local semantic search CLI. Ingest documents, embed them locally, and run vector search — no external services required.

How it works

litesearch ingest runs a document through a four-stage pipeline:

1. Parse: extracts plain text from the file using LiteParse
2. Chunk: splits the text into overlapping segments using Chonkie (default: 512 characters)
3. Embed: generates 768-dimensional embeddings locally using nomic-ai/nomic-embed-text-v1.5 via @huggingface/transformers
4. Store: upserts vectors and payloads into a local Qdrant Edge shard persisted at .litesearch.qdrant/

litesearch retrieve embeds the query with the same model and performs a cosine similarity search against the local shard.

The vector store is backed by qdrant-edge-utils, a Rust native addon (napi-rs) that wraps qdrant-edge for on-disk mmap-backed storage.

Requirements

• Bun
• Rust toolchain (only needed if rebuilding qdrant-edge-utils from source)

Installation

bun install

Usage

Ingest a document

bun run ingest <file> [options]

Option	Description
-c, --config <path>	Path to a LiteParse config file
-s, --chunk-size <number>	Chunk size in characters (default: 512)

Search

bun run retrieve <query> [options]

Option	Description
-f, --files <paths...>	Restrict search to specific ingested document paths
-l, --limit <number>	Maximum number of results (default: 10)
-t, --score-threshold <number>	Minimum cosine similarity score

Example

bun run ingest ./data/report.pdf
bun run retrieve "quarterly revenue breakdown" --limit 5

Output:

[0.8723] ./data/report.pdf
Revenue for Q3 reached $4.2M, up 18% from the prior quarter...

[0.8341] ./data/report.pdf
The board approved a revised forecast of $17M for the full fiscal year...

TUI

You can also use the terminal interface to access the ingestion and retrieval functionalities interactively:

bun run tui

You can then follow the prompts from terminal, as in this image:

Project structure

src/
  index.ts       # CLI entry point (commander)
  pipelines.ts   # ingest and retrieve orchestration
  parsing.ts     # document parsing via LiteParse
  chunk.ts       # text chunking via Chonkie
  embed.ts       # local embedding via transformers.js
  store.ts       # vector upsert and search
  tui.ts         # utilities to run the terminal interface

packages/
  qdrant-edge-utils/
    src/lib.rs   # Rust native addon wrapping qdrant-edge

Vector store

The local shard is stored in .litesearch.qdrant/ in the working directory. This directory is created automatically on first ingest. To reset the store, delete this directory.

Development

bun run lint
bun run format