GeneralBots/generalbots
1
0
Fork 0
Code Activity Actions 5

fix: handle GLM 4.7 reasoning_content and chat_template_kwargs
All checks were successful
BotServer CI/CD / build (push) Successful in 3m29s
Details

Browse source
This commit is contained in:
Rodrigo Rodriguez (Pragmatismo) 2026-04-14 17:14:04 -03:00
parent 44026ba073
commit 62cdf1c638
1 changed files with 40 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

47
src/llm/mod.rs
View file

@ -402,10 +402,15 @@ impl LLMProvider for OpenAIClient {
"top_p": 1.0 "top_p": 1.0
}); });
// Kimi K2.5 factory: enable thinking mode via chat_template_kwargs // GLM 4.7 / Kimi K2.5 factory: enable thinking mode via chat_template_kwargs
if model.contains("kimi") { if model.contains("kimi") || model.contains("glm") {
request_body["chat_template_kwargs"] = serde_json::json!({"thinking": true}); let kwargs = if model.contains("glm") {
info!("Kimi factory: enabled thinking mode (chat_template_kwargs)"); serde_json::json!({"enable_thinking": true, "clear_thinking": false})
} else {
serde_json::json!({"thinking": true})
};
request_body["chat_template_kwargs"] = kwargs;
info!("Model factory: enabled thinking mode for {} (chat_template_kwargs)", model);
} }
// Add tools to the request if provided // Add tools to the request if provided
@ -435,16 +440,44 @@ impl LLMProvider for OpenAIClient {
let mut stream_state = String::new(); // State for the handler to track thinking tags across chunks let mut stream_state = String::new(); // State for the handler to track thinking tags across chunks
let mut stream = response.bytes_stream(); let mut stream = response.bytes_stream();
let mut in_reasoning = false;
while let Some(chunk_result) = stream.next().await { while let Some(chunk_result) = stream.next().await {
let chunk = chunk_result?; let chunk = chunk_result?;
let chunk_str = String::from_utf8_lossy(&chunk); let chunk_str = String::from_utf8_lossy(&chunk);
for line in chunk_str.lines() { for line in chunk_str.lines() {
if line.starts_with("data: ") && !line.contains("[DONE]") { if line.starts_with("data: ") && !line.contains("[DONE]") {
if let Ok(data) = serde_json::from_str::<Value>(&line[6..]) { if let Ok(data) = serde_json::from_str::<Value>(&line[6..]) {
// Handle reasoning_content (GLM 4.7 / deepseek / etc)
if let Some(reasoning) = data["choices"][0]["delta"]["reasoning_content"].as_str() {
if !reasoning.is_empty() {
if !in_reasoning {
in_reasoning = true;
}
// Send thinking tokens to UI via JSON stringified message
let thinking_msg = serde_json::json!({
"type": "thinking",
"content": reasoning
}).to_string();
let _ = tx.send(thinking_msg).await;
}
}
// Handle regular content
if let Some(content) = data["choices"][0]["delta"]["content"].as_str() { if let Some(content) = data["choices"][0]["delta"]["content"].as_str() {
let processed = handler.process_content_streaming(content, &mut stream_state); if !content.is_empty() {
if !processed.is_empty() { if in_reasoning {
let _ = tx.send(processed).await; in_reasoning = false;
// Send clear signal
let clear_msg = serde_json::json!({
"type": "thinking_clear"
}).to_string();
let _ = tx.send(clear_msg).await;
}
let processed = handler.process_content_streaming(content, &mut stream_state);
if !processed.is_empty() {
let _ = tx.send(processed).await;
}
} }
} }