naduke

CLI tool that guesses better file names from file content using a local Ollama model, then renames the files while keeping their extensions.

Requirements

Go 1.22+
Ollama installed and running locally or remotely.
The model granite4:3b-h available in Ollama (or another model you specify).

Model download (granite4:3b-h)

ollama pull granite4:3b-h

Installation

go install github.com/takai/naduke/cmd/naduke@latest

Or build locally:

go build -o naduke ./cmd/naduke

Usage

naduke [options] FILE...

Options (from naduke -h):

-host Ollama host (default: localhost)
-port Ollama port (default: 11434)
-server Full Ollama server URL (overrides host/port)
-model Model name (default: granite4:3b-h)
-temperature Sampling temperature (default: 0.0)
-top_k Top-k sampling (default: 1)
-top_p Top-p sampling (default: 1.0)
-repeat_penalty Repeat penalty (default: 1.0)
-dry-run Show suggested names without renaming (note: actual rename run may produce a different suggestion because LLM outputs can vary)
-prefix Prefix to prepend to the generated name
-dir Destination directory for renamed files (default: same as source)
-h, -help Show help

Examples:

# Basic rename with defaults
naduke notes/todo.txt

# Multiple files at once
naduke docs/*.md

# Custom server URL
naduke -server http://ollama.example.com:11434 draft.txt

# Add a prefix to suggestions
naduke -prefix meeting_ notes.txt

# Rename into another directory
naduke -dir out/ docs/*.md

Behavior

Reads the first 1,000 characters (up to ~4KB); aborts on NUL bytes or invalid UTF-8.
Sends system/user prompts to /api/chat (no streaming).
Sanitizes model output; if empty after sanitization, uses file.
Keeps the original extension (e.g., draft.md -> summary.md).
Allows choosing a different destination directory via -dir; source file must be reachable and destination dir must exist.
Fails if the destination already exists.
Dry-run prints suggestions only; due to LLM variability, a later non-dry run might produce a different name.
Validates model output against naming rules (single token, lowercase a-z0-9_, max 30 chars, no extension).
Applies an optional prefix as provided, then appends the model output.

Parameter notes (you do not usually need to change these):

temperature: Controls randomness/creativity. Higher = more varied suggestions; lower = safer/more deterministic.
top_k: Limits candidates to the top-K tokens before sampling. Lower = conservative; higher = more diverse.
top_p: Nucleus sampling; keeps the smallest set of tokens whose cumulative probability ≥ top_p. Higher = more diverse; lower = more focused.
repeat_penalty: Penalizes repeating tokens. Higher than 1 discourages repetition; keep near 1 for normal behavior.

Testing

GOCACHE=$(pwd)/.cache/go-build go test ./...

License

MIT

takai/naduke