chat-mlx 0.0.0

Local-inference chat-rs provider (and CLI) for MiniCPM5 / Llama / Qwen models on Apple Silicon via MLX.
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
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72

use chat_core::types::messages::content::{CompleteReasonEnum, Content, RoleEnum};
use chat_core::types::messages::parts::{PartEnum, Parts};
use chat_core::types::messages::reasoning::Reasoning;
use chat_core::types::messages::text::Text;
use chat_core::types::metadata::Metadata;
use chat_core::types::metadata::usage::Usage;
use chat_core::types::response::ChatResponse;
use serde_json::Value;
use tools_rs::FunctionCall;

/// Assemble a [`ChatResponse`] from the split-out pieces of one generation.
/// Used by both the completion and streaming paths so they stay equivalent.
///
/// Part order is reasoning → (structured | text + tool calls). A `Structured`
/// value, when present, is emitted as the last part (and supersedes text/tools,
/// which a structured request won't have). Tool calls set the finish reason to
/// `ToolCall`, which drives the chat loop to execute them and call again.
#[allow(clippy::too_many_arguments)]
pub fn build(
    model_id: &str,
    reasoning_text: String,
    text: String,
    calls: Vec<FunctionCall>,
    structured: Option<Value>,
    input_tokens: usize,
    output_tokens: usize,
    max_tokens: usize,
) -> ChatResponse {
    let complete_reason = if !calls.is_empty() {
        CompleteReasonEnum::ToolCall
    } else if output_tokens >= max_tokens {
        CompleteReasonEnum::MaxTokens
    } else {
        CompleteReasonEnum::Stop
    };

    let mut parts = Vec::new();
    if !reasoning_text.trim().is_empty() {
        parts.push(PartEnum::Reasoning(Reasoning::new(reasoning_text)));
    }
    if let Some(value) = structured {
        parts.push(PartEnum::Structured(value));
    } else {
        if !text.trim().is_empty() {
            parts.push(PartEnum::Text(Text::new(text)));
        }
        for call in calls {
            parts.push(PartEnum::from_function_call(call));
        }
    }

    let content = Content {
        role: RoleEnum::Model,
        parts: Parts(parts),
        complete_reason,
    };

    let metadata = Metadata {
        model_slug: Some(model_id.to_string()),
        usage: Usage {
            input_tokens,
            output_tokens,
            total_tokens: input_tokens + output_tokens,
        },
        ..Default::default()
    };

    ChatResponse {
        metadata: Some(metadata),
        content,
    }
}