bcode_server 0.0.1-alpha.0

Local bcode daemon runtime
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
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144

//! Context accounting helpers for model requests and structural compaction.
//!
//! Proactive compaction uses a conservative estimate of the complete candidate request assembled
//! for the current round. Structural retention uses canonical serialized model-message accounting
//! so cuts include protocol overhead rather than raw content character counts.

use super::{
    ContentBlock, ModelMessage, ModelTurnRequest, session_events_to_model_messages_with_limit,
};

/// Current version of the local semantic request-context estimator.
pub const LOCAL_CONTEXT_ESTIMATOR_VERSION: u16 = 1;

/// Estimate the complete semantic model-visible request context.
///
/// Host metadata, prompt-cache controls, continuation identifiers, and opaque provider state are
/// excluded because they are not semantic active context.
pub fn local_request_estimate(
    request: &ModelTurnRequest,
) -> bcode_session_models::LocalContextEstimate {
    let tokens = serde_json::to_string(&(
        request.system_prompt.as_ref(),
        &request.messages,
        &request.tools,
        &request.parameters,
        request.structured_output.as_ref(),
        &request.provider_context.request,
    ))
    .map_or(u64::MAX, |serialized| {
        estimated_tokens_from_chars(serialized.chars().count())
    })
    .saturating_add(image_tokens(&request.messages));
    bcode_session_models::LocalContextEstimate {
        tokens,
        algorithm_version: LOCAL_CONTEXT_ESTIMATOR_VERSION,
    }
}

pub fn projected_model_context_chars(
    history: &[bcode_session_models::SessionEvent],
    tool_output_context_chars: usize,
) -> usize {
    session_events_to_model_messages_with_limit(history, tool_output_context_chars)
        .iter()
        .map(model_message_context_chars)
        .sum()
}

pub fn estimated_tokens_from_chars(chars: usize) -> u64 {
    bcode_session_models::estimated_normalized_text_tokens_from_chars(chars)
}

pub fn estimated_model_messages_tokens(messages: &[ModelMessage]) -> u64 {
    estimated_semantic_messages_tokens(messages)
}

fn estimated_semantic_messages_tokens(messages: &[ModelMessage]) -> u64 {
    let structural = serde_json::to_vec(messages).map_or(u64::MAX, |serialized| {
        u64::try_from(serialized.len())
            .unwrap_or(u64::MAX)
            .saturating_add(3)
            / 4
    });
    structural.saturating_add(image_tokens(messages))
}

fn image_tokens(messages: &[ModelMessage]) -> u64 {
    messages
        .iter()
        .flat_map(|message| &message.content)
        .map(|block| match block {
            ContentBlock::Image { image } => image_metadata_tokens(&image.metadata),
            ContentBlock::ToolResult { result } => result
                .content
                .iter()
                .map(|content| match content {
                    bcode_model::ToolResultContent::Image { image } => {
                        image_metadata_tokens(&image.metadata)
                    }
                    bcode_model::ToolResultContent::ImageRef { image } => {
                        image_metadata_tokens(&image.metadata)
                    }
                    bcode_model::ToolResultContent::Text { .. } => 0,
                })
                .sum(),
            ContentBlock::Text { .. }
            | ContentBlock::ToolCall { .. }
            | ContentBlock::CachePoint { .. }
            | ContentBlock::ProviderExtension { .. } => 0,
        })
        .sum()
}

fn image_metadata_tokens(metadata: &bcode_model::ImageMetadata) -> u64 {
    metadata
        .width
        .zip(metadata.height)
        .map_or(0, |(width, height)| {
            u64::from(width)
                .saturating_mul(u64::from(height))
                .saturating_add(749)
                / 750
        })
}

pub fn model_message_context_chars(message: &ModelMessage) -> usize {
    message
        .content
        .iter()
        .map(|block| match block {
            ContentBlock::Text { text } => text.chars().count(),
            ContentBlock::Image { image } => {
                usize::try_from(image_metadata_tokens(&image.metadata).saturating_mul(4))
                    .unwrap_or(usize::MAX)
            }
            ContentBlock::CachePoint { .. } => 0,
            ContentBlock::ToolCall { call } => {
                call.name.chars().count() + call.arguments.to_string().chars().count()
            }
            ContentBlock::ToolResult { result } => result.output.chars().count().saturating_add(
                usize::try_from(
                    result
                        .content
                        .iter()
                        .map(|content| match content {
                            bcode_model::ToolResultContent::Image { image } => {
                                image_metadata_tokens(&image.metadata)
                            }
                            bcode_model::ToolResultContent::ImageRef { image } => {
                                image_metadata_tokens(&image.metadata)
                            }
                            bcode_model::ToolResultContent::Text { text } => {
                                estimated_tokens_from_chars(text.chars().count())
                            }
                        })
                        .sum::<u64>()
                        .saturating_mul(4),
                )
                .unwrap_or(usize::MAX),
            ),
            ContentBlock::ProviderExtension { value } => value.to_string().chars().count(),
        })
        .sum()
}