i/Aether
1
0
Fork 0
mirror of https://github.com/fawney19/Aether.git synced 2026-10-13 12:39:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: 修复缓存 token usage 转换语义

Browse Source 
统一 OpenAI、Gemini、Claude 之间缓存 token 的 usage 语义，避免 Claude 侧重复统计缓存输入 token。

同时补充 stream 合并逻辑、字段注释和覆盖转换链路的测试。
This commit is contained in:
stabey
2026-05-27 23:49:07 +08:00
parent 14ad6e9b75
commit 739e39e1ca
9 changed files with 274 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/aether-ai-formats/src/formats/shared/stream_core/common.rs
+60 -10
View File
@@ -99,6 +99,7 @@ pub fn canonical_usage_from_openai_usage(value: Option<&Value>) -> Option<Canoni
}
Some(CanonicalUsage {
input_tokens,
input_tokens_include_cache: cache_read_tokens > 0 || cache_creation_tokens > 0,
output_tokens,
total_tokens,
cache_creation_tokens,
@@ -247,6 +248,7 @@ pub fn canonical_usage_from_claude_usage(value: Option<&Value>) -> Option<Canoni
.unwrap_or(0);
Some(CanonicalUsage {
input_tokens,
input_tokens_include_cache: false,
output_tokens,
total_tokens: input_tokens
.saturating_add(output_tokens)
@@ -314,22 +316,27 @@ pub fn canonical_usage_from_gemini_usage(value: Option<&Value>) -> Option<Canoni
let usage = value?.as_object()?;
let input_tokens = usage
.get("promptTokenCount")
.or_else(|| usage.get("prompt_token_count"))
.and_then(Value::as_u64)
.unwrap_or(0);
let output_tokens = usage
.get("candidatesTokenCount")
.or_else(|| usage.get("candidates_token_count"))
.and_then(Value::as_u64)
.unwrap_or(0);
let reasoning_tokens = usage
.get("thoughtsTokenCount")
.or_else(|| usage.get("thoughts_token_count"))
.and_then(Value::as_u64)
.unwrap_or(0);
let cache_read_tokens = usage
.get("cachedContentTokenCount")
.or_else(|| usage.get("cached_content_token_count"))
.and_then(Value::as_u64)
.unwrap_or(0);
let total_tokens = usage
.get("totalTokenCount")
.or_else(|| usage.get("total_token_count"))
.and_then(Value::as_u64)
.unwrap_or(
input_tokens
@@ -338,6 +345,7 @@ pub fn canonical_usage_from_gemini_usage(value: Option<&Value>) -> Option<Canoni
);
Some(CanonicalUsage {
input_tokens,
input_tokens_include_cache: cache_read_tokens > 0,
output_tokens: output_tokens.saturating_add(reasoning_tokens),
total_tokens,
cache_read_tokens,
@@ -490,12 +498,13 @@ pub fn build_openai_chat_usage_chunk_from_usage(
model: &str,
usage: &CanonicalUsage,
) -> Value {
let input_tokens = inclusive_input_tokens_from_usage(usage);
build_openai_chat_usage_chunk_with_cache(
id,
model,
usage.input_tokens,
input_tokens,
usage.output_tokens,
usage.total_tokens,
inclusive_total_tokens_from_usage(usage, input_tokens),
usage.reasoning_tokens,
cache_creation_tokens_for_usage(usage),
usage.cache_read_tokens,
@@ -504,12 +513,16 @@ pub fn build_openai_chat_usage_chunk_from_usage(
pub fn openai_responses_usage_from_usage(usage: &CanonicalUsage) -> Value {
let mut output = Map::new();
output.insert("input_tokens".to_string(), Value::from(usage.input_tokens));
let input_tokens = inclusive_input_tokens_from_usage(usage);
output.insert("input_tokens".to_string(), Value::from(input_tokens));
output.insert(
"output_tokens".to_string(),
Value::from(usage.output_tokens),
);
output.insert("total_tokens".to_string(), Value::from(usage.total_tokens));
output.insert(
"total_tokens".to_string(),
Value::from(inclusive_total_tokens_from_usage(usage, input_tokens)),
);
if usage.reasoning_tokens > 0 {
output.insert(
"output_tokens_details".to_string(),
@@ -527,7 +540,10 @@ pub fn openai_responses_usage_from_usage(usage: &CanonicalUsage) -> Value {
pub fn claude_usage_from_usage(usage: &CanonicalUsage) -> Value {
let mut output = Map::new();
output.insert("input_tokens".to_string(), Value::from(usage.input_tokens));
output.insert(
"input_tokens".to_string(),
Value::from(claude_input_tokens_from_usage(usage)),
);
output.insert(
"output_tokens".to_string(),
Value::from(usage.output_tokens),
@@ -560,18 +576,16 @@ pub fn claude_usage_from_usage(usage: &CanonicalUsage) -> Value {
pub fn gemini_usage_metadata_from_usage(usage: &CanonicalUsage) -> Value {
let visible_output_tokens = usage.output_tokens.saturating_sub(usage.reasoning_tokens);
let input_tokens = inclusive_input_tokens_from_usage(usage);
let mut output = Map::new();
output.insert(
"promptTokenCount".to_string(),
Value::from(usage.input_tokens),
);
output.insert("promptTokenCount".to_string(), Value::from(input_tokens));
output.insert(
"candidatesTokenCount".to_string(),
Value::from(visible_output_tokens),
);
output.insert(
"totalTokenCount".to_string(),
Value::from(usage.total_tokens),
Value::from(inclusive_total_tokens_from_usage(usage, input_tokens)),
);
if usage.reasoning_tokens > 0 {
output.insert(
@@ -649,3 +663,39 @@ fn cache_creation_tokens_for_usage(usage: &CanonicalUsage) -> u64 {
.saturating_add(usage.cache_creation_ephemeral_1h_tokens)
}
}
fn cache_input_tokens_for_usage(usage: &CanonicalUsage) -> u64 {
usage
.cache_read_tokens
.saturating_add(cache_creation_tokens_for_usage(usage))
}
fn claude_input_tokens_from_usage(usage: &CanonicalUsage) -> u64 {
if usage.input_tokens_include_cache {
usage
.input_tokens
.saturating_sub(cache_input_tokens_for_usage(usage))
} else {
usage.input_tokens
}
}
fn inclusive_input_tokens_from_usage(usage: &CanonicalUsage) -> u64 {
if usage.input_tokens_include_cache {
usage.input_tokens
} else {
usage
.input_tokens
.saturating_add(cache_input_tokens_for_usage(usage))
}
}
fn inclusive_total_tokens_from_usage(usage: &CanonicalUsage, input_tokens: u64) -> u64 {
if usage.total_tokens > 0
&& (usage.input_tokens_include_cache || cache_input_tokens_for_usage(usage) == 0)
{
usage.total_tokens
} else {
input_tokens.saturating_add(usage.output_tokens)
}
}
crates/aether-ai-formats/src/formats/shared/sync_products.rs
+13 -22
View File
@@ -26,8 +26,9 @@ use crate::formats::shared::response::{
sanitize_claude_read_tool_inputs,
};
use crate::formats::shared::stream_core::common::{
content_part_from_openai_image_generation_item, map_openai_finish_reason_to_gemini,
parse_json_arguments_value, CanonicalContentPart, CanonicalStreamEvent, CanonicalUsage,
content_part_from_openai_image_generation_item, gemini_usage_metadata_from_usage,
map_openai_finish_reason_to_gemini, parse_json_arguments_value, CanonicalContentPart,
CanonicalStreamEvent, CanonicalUsage,
};
#[derive(Clone, Debug, PartialEq)]
@@ -3044,26 +3045,7 @@ fn gemini_sync_part_from_canonical_content_part(part: CanonicalContentPart) -> V
}
fn gemini_usage_metadata_from_canonical(usage: CanonicalUsage) -> Value {
let mut usage_metadata = Map::new();
usage_metadata.insert(
"promptTokenCount".to_string(),
Value::from(usage.input_tokens),
);
usage_metadata.insert(
"candidatesTokenCount".to_string(),
Value::from(usage.output_tokens.saturating_sub(usage.reasoning_tokens)),
);
usage_metadata.insert(
"totalTokenCount".to_string(),
Value::from(usage.total_tokens),
);
if usage.reasoning_tokens > 0 {
usage_metadata.insert(
"thoughtsTokenCount".to_string(),
Value::from(usage.reasoning_tokens),
);
}
Value::Object(usage_metadata)
gemini_usage_metadata_from_usage(&usage)
}
fn parse_data_url(value: &str) -> Option<(String, String)> {
@@ -4213,6 +4195,9 @@ mod tests {
"prompt_tokens": 2,
"completion_tokens": 4,
"total_tokens": 6,
"prompt_tokens_details": {
"cached_tokens": 1
},
"completion_tokens_details": {
"reasoning_tokens": 1
}
@@ -4230,6 +4215,9 @@ mod tests {
)
.expect("canonical openai chat -> claude");
assert_eq!(converted_claude, legacy_claude);
assert_eq!(converted_claude["usage"]["input_tokens"], 1);
assert_eq!(converted_claude["usage"]["cache_read_input_tokens"], 1);
assert_eq!(converted_claude["usage"]["output_tokens"], 4);
let legacy_gemini =
convert_openai_chat_response_to_gemini_chat(&provider_body_json, &report_context)
@@ -4341,6 +4329,9 @@ mod tests {
converted_claude["content"][3]["content"],
json!({"ok": true})
);
assert_eq!(converted_claude["usage"]["input_tokens"], 1);
assert_eq!(converted_claude["usage"]["cache_read_input_tokens"], 2);
assert_eq!(converted_claude["usage"]["output_tokens"], 5);
let converted_gemini = convert_standard_chat_response(
&provider_body_json,