ollama-vision 0.2.0

Robust Ollama vision model toolkit for image tagging and captioning
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

use ollama_vision::{OllamaVisionConfig, TagOptions};
use std::path::Path;

/// Demonstrates tagging with a reasoning/thinking model (e.g., deepseek-r1).
///
/// Thinking models wrap their reasoning in `<think>...</think>` blocks.
/// The parser automatically strips these before extracting tags.
#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let image_path = std::env::args().nth(1).unwrap_or_else(|| {
        eprintln!("Usage: thinking_mode <image_path> [model]");
        eprintln!("  model defaults to 'minicpm-v' (supports thinking)");
        std::process::exit(1);
    });

    let model = std::env::args().nth(2).unwrap_or("minicpm-v".to_string());

    let config = OllamaVisionConfig::with_model(&model);
    let client = reqwest::Client::new();

    // Disable JSON format constraint — thinking models often struggle with it
    let options = TagOptions {
        request_json_format: false,
        ..Default::default()
    };

    println!("Tagging {} with thinking model '{}'...", image_path, model);
    println!("(JSON format constraint disabled for thinking model compatibility)");

    let tags = ollama_vision::tag_image(&client, &config, Path::new(&image_path), &options).await?;

    println!("\nTags ({}):", tags.len());
    for tag in &tags {
        println!("  - {}", tag);
    }

    Ok(())
}