valknar/llmx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: context compaction (#3446)

Browse Source 
## Compact feature:
1. Stops the model when the context window become too large
2. Add a user turn, asking for the model to summarize
3. Build a bridge that contains all the previous user message + the
summary. Rendered from a template
4. Start sampling again from a clean conversation with only that bridge
This commit is contained in:
jif-oai
2025-09-12 13:07:10 -07:00
committed by GitHub
parent d4848e558b
commit ea225df22e
14 changed files with 1243 additions and 326 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
codex-rs/core/src/openai_model_info.rs
View File

@@ -12,6 +12,19 @@ pub(crate) struct ModelInfo {
/// Maximum number of output tokens that can be generated for the model.
pub(crate) max_output_tokens: u64,
/// Token threshold where we should automatically compact conversation history.
pub(crate) auto_compact_token_limit: Option<i64>,
}
impl ModelInfo {
const fn new(context_window: u64, max_output_tokens: u64) -> Self {
Self {
context_window,
max_output_tokens,
auto_compact_token_limit: None,
}
}
}
pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
@@ -20,73 +33,37 @@ pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
// OSS models have a 128k shared token pool.
// Arbitrarily splitting it: 3/4 input context, 1/4 output.
// https://openai.com/index/gpt-oss-model-card/
"gpt-oss-20b" => Some(ModelInfo {
context_window: 96_000,
max_output_tokens: 32_000,
}),
"gpt-oss-120b" => Some(ModelInfo {
context_window: 96_000,
max_output_tokens: 32_000,
}),
"gpt-oss-20b" => Some(ModelInfo::new(96_000, 32_000)),
"gpt-oss-120b" => Some(ModelInfo::new(96_000, 32_000)),
// https://platform.openai.com/docs/models/o3
"o3" => Some(ModelInfo {
context_window: 200_000,
max_output_tokens: 100_000,
}),
"o3" => Some(ModelInfo::new(200_000, 100_000)),
// https://platform.openai.com/docs/models/o4-mini
"o4-mini" => Some(ModelInfo {
context_window: 200_000,
max_output_tokens: 100_000,
}),
"o4-mini" => Some(ModelInfo::new(200_000, 100_000)),
// https://platform.openai.com/docs/models/codex-mini-latest
"codex-mini-latest" => Some(ModelInfo {
context_window: 200_000,
max_output_tokens: 100_000,
}),
"codex-mini-latest" => Some(ModelInfo::new(200_000, 100_000)),
// As of Jun 25, 2025, gpt-4.1 defaults to gpt-4.1-2025-04-14.
// https://platform.openai.com/docs/models/gpt-4.1
"gpt-4.1" | "gpt-4.1-2025-04-14" => Some(ModelInfo {
context_window: 1_047_576,
max_output_tokens: 32_768,
}),
"gpt-4.1" | "gpt-4.1-2025-04-14" => Some(ModelInfo::new(1_047_576, 32_768)),
// As of Jun 25, 2025, gpt-4o defaults to gpt-4o-2024-08-06.
// https://platform.openai.com/docs/models/gpt-4o
"gpt-4o" | "gpt-4o-2024-08-06" => Some(ModelInfo {
context_window: 128_000,
max_output_tokens: 16_384,
}),
"gpt-4o" | "gpt-4o-2024-08-06" => Some(ModelInfo::new(128_000, 16_384)),
// https://platform.openai.com/docs/models/gpt-4o?snapshot=gpt-4o-2024-05-13
"gpt-4o-2024-05-13" => Some(ModelInfo {
context_window: 128_000,
max_output_tokens: 4_096,
}),
"gpt-4o-2024-05-13" => Some(ModelInfo::new(128_000, 4_096)),
// https://platform.openai.com/docs/models/gpt-4o?snapshot=gpt-4o-2024-11-20
"gpt-4o-2024-11-20" => Some(ModelInfo {
context_window: 128_000,
max_output_tokens: 16_384,
}),
"gpt-4o-2024-11-20" => Some(ModelInfo::new(128_000, 16_384)),
// https://platform.openai.com/docs/models/gpt-3.5-turbo
"gpt-3.5-turbo" => Some(ModelInfo {
context_window: 16_385,
max_output_tokens: 4_096,
}),
"gpt-3.5-turbo" => Some(ModelInfo::new(16_385, 4_096)),
_ if slug.starts_with("gpt-5") => Some(ModelInfo {
context_window: 272_000,
max_output_tokens: 128_000,
}),
_ if slug.starts_with("gpt-5") => Some(ModelInfo::new(272_000, 128_000)),
_ if slug.starts_with("codex-") => Some(ModelInfo {
context_window: 272_000,
max_output_tokens: 128_000,
}),
_ if slug.starts_with("codex-") => Some(ModelInfo::new(272_000, 128_000)),
_ => None,
}