monkeSearch

---

Read the technical report at: monkesearch.github.io

---

A prototype for searching your files with natural language — fully offline, aimed to run on potato PCs. No GPU required, no cloud API calls, nothing leaves your machine. Currently macOS-only (uses Spotlight) with cross-platform support in the works.

How It Works

Any natural language file search query can be broken into 3 constituents:

1. File type — what kind of file (pdf, image, python script, etc.)
2. Temporal context — when (3 days ago, last week, 7 months ago)
3. Misc keywords — any remaining context (project name, topic, content)

monkeSearch uses a small local LLM (like LFM 1.2B) to parse queries and convert them into native macOS Spotlight search predicates.

Quick Start

# Terminal 1: Start llama-server (keep running) (or point to any openai compatible endpoint)
llama-server --hf-repo LiquidAI/LFM2.5-1.2B-Instruct-GGUF --hf-file LFM2.5-1.2B-Instruct-Q8_0.gguf --port 8080

# Terminal 2: Search your files
cd app/
python parser.py "photos from yesterday"

As a Module

from parser import FileSearchParser

searcher = FileSearchParser()
results, parsed_data, misc = searcher.search("python files from last week")
for path in results:
    print(path)

Requirements

• macOS (for Spotlight integration)
• Python 3.8+
• llama-server from llama.cpp (install via brew install llama.cpp)
• A GGUF model (default: LFM 2.5 1.2B, ~700MB)

Configuration

Set MONKE_SERVER_URL to point to your llama-server (default: http://localhost:8080/v1):

export MONKE_SERVER_URL="http://192.168.1.42:8080/v1"
python parser.py "photos from last week"

Branches

Branch	Approach	Platform
main / dev	LLM → Spotlight NSPredicate (this branch)	macOS
vectordb	Vector DB (LEANN/ChromaDB) + semantic embeddings	macOS / Linux / Windows

Example Queries

Query	What It Finds
"photos from yesterday"	Image files modified in the last day
"python scripts from 3 days ago"	.py files from 3 days ago
"pdf invoices from last month"	PDFs with "invoices" modified in the last month
"code files"	Source code files of any language
"videos from 2 years ago"	Video files modified ~2 years ago

Limitations

• Spotlight-indexed files only (more like a feature, you can scope your search)
• Metadata-only — file content search planned
• Small LLM tradeoff — tiny models can misunderstand complex queries
• Basic temporal — simple time expressions only
• macOS-only — see the vectordb branch for cross-platform

License

Apache-2.0