fix(stream): add anti-buffer headers for real SSE streaming

- Add X-Accel-Buffering: no, Cache-Control: no-cache, Connection: keep-alive
- Prevents Traefik/proxy from buffering SSE events
- Each token is flushed immediately as generated

.github/workflows/notify-parent.yml

@@ -0,0 +1,25 @@
+name: Notify Parent Repo
+
+on:
+  push:
+    branches:
+      - master
+  workflow_dispatch:
+
+jobs:
+  dispatch:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Trigger root monorepo build
+        uses: peter-evans/repository-dispatch@v3
+        with:
+          token: ${{ secrets.DISPATCH_TOKEN }}
+          repository: asepharyana/asepharyana-hub
+          event-type: submodule-updated
+          client-payload: |
+            {
+              "service": "llm-api",
+              "ref": "${{ github.ref }}",
+              "sha": "${{ github.sha }}",
+              "actor": "${{ github.actor }}"
+            }

Cargo.lock

@@ -568,6 +568,7 @@ dependencies = [
  "chrono",
  "futures",
  "llama-cpp-2",
+ "minijinja",
  "serde",
  "serde_json",
  "thiserror 1.0.69",
@@ -615,12 +616,28 @@ version = "2.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
 
+[[package]]
+name = "memo-map"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "38d1115007560874e373613744c6fba374c17688327a71c1476d1a5954cc857b"
+
 [[package]]
 name = "mime"
 version = "0.3.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6877bb514081ee2a7ff5ef9de3281f14a4dd4bceac4c09388074a6b5df8a139a"
 
+[[package]]
+name = "minijinja"
+version = "2.21.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb3d648e68cea56d9858d535ee28f9538404e2dd8cb08ed0bd05dca379477f39"
+dependencies = [
+ "memo-map",
+ "serde",
+]
+
 [[package]]
 name = "minimal-lexical"
 version = "0.2.1"

Cargo.toml

@@ -6,6 +6,7 @@ edition = "2021"
 [dependencies]
 # LLM inference
 llama-cpp-2 = "0.1"
+minijinja = "2"
 
 # HTTP server
 axum = { version = "0.8", features = ["json"] }

src/application/chat/templates/chat_template.jinja

@@ -0,0 +1,53 @@
+{{- bos_token }}
+{%- if tools %}
+    {%- set tool_definitions %}
+        {{- "# Tools\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+        {%- for tool in tools %}
+            {{- "\n" }}
+            {{- tool | tojson }}
+        {%- endfor %}
+        {{- '\n</tools>\n\nTool usage guidelines:\n- You may call zero or more functions. If no function calls are needed, just answer normally.\n- When calling a function, use: <function name="name"><param name="key">value</param></function>' }}
+    {%- endset %}
+
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' + tool_definitions }}
+    {%- else %}
+        {{- tool_definitions }}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = message.content %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- if message.tool_calls %}
+            {%- set content = message.content %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.type == "function" %}
+                    {%- set content = content + "\n<tool_call>\n<function=" + tool_call.function.name + ">\n" + (tool_call.function.arguments | tojson) + "\n</function>\n</tool_call>" %}
+                {%- endif %}
+            {%- endfor %}
+            {{- '<|im_start|>assistant\n' + content + '<|im_end|>\n' }}
+        {%- else %}
+            {{- '<|im_start|>assistant\n' + content + '<|im_end|>\n' }}
+        {%- endif %}
+    {%- elif message.role == "tool" %}
+        {{- '<|im_start|>user\n<tool_response>\n' + content + '\n</tool_response><|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined %}
+        {%- if enable_thinking is false %}
+            {{- '<think>\n\n</think>\n\n' }}
+        {%- elif enable_thinking is true %}
+            {{- '<think>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endif %}

src/application/chat/use_cases.rs

@@ -1,97 +1,95 @@
 //! Chat completion use cases.
 //!
-//! Orchestrates prompt building, sampler construction, and output parsing.
+//! Orchestrates prompt building using the model's baked-in Jinja template
-//! These are pure functions with no framework dependencies.
+//! via the `minijinja` crate, sampler construction, and output parsing.
 
+use std::collections::HashMap;
+
+use minijinja::{Environment, Value};
 use llama_cpp_2::sampling::LlamaSampler;
 
-use crate::domain::entity::{ChatMessage, ChatRequest, ToolCall, ToolCallFunction, ToolDef};
+use crate::domain::entity::{ChatMessage, ChatRequest, ToolCall, ToolCallFunction};
 
-/// Build a prompt string from conversation messages and optional tool definitions.
+/// Build a prompt from messages using the GGUF's Jinja chat template.
 ///
-/// Uses ChatML format with `<|im_start|>` / `<|im_end|>` delimiters. Tool
+/// Renders the model's baked-in template via minijinja, passing the message
-/// definitions are injected into the first system message.
+/// history, optional tool definitions, and generation-prompt switches.
-pub fn build_prompt(messages: &[ChatMessage], tools: &Option<Vec<ToolDef>>) -> String {
+pub fn build_prompt(
-    let mut prompt = String::new();
+    model: &llama_cpp_2::model::LlamaModel,
+    messages: &[ChatMessage],
+    _tools: &Option<Vec<crate::domain::entity::ToolDef>>,
+) -> Result<String, String> {
+    // Load the GGUF's chat template (embedded at crate build time)
+    let template_str = include_str!("templates/chat_template.jinja");
 
-    for (i, msg) in messages.iter().enumerate() {
+    let mut env = Environment::new();
-        match msg.role.as_str() {
+    env.add_template("chat", template_str)
-            "system" => {
+        .map_err(|e| format!("Template add error: {e}"))?;
-                let mut content = msg.content.clone().unwrap_or_default();
+
-                // Inject tools into the system message (first occurrence)
+    // Register tojson filter (safe: Rust serde_json defaults to ensure_ascii=false)
-                if i == 0 {
+    env.add_filter("tojson", |value: &Value| -> String {
-                    if let Some(tools_list) = tools {
+        serde_json::to_string(value).unwrap_or_default()
-                        if !tools_list.is_empty() {
+    });
-                            let mut tools_text = String::from(
+
-                                "\n\n# Tools\n\nYou have access to the following functions:\n\n<tools>",
+    let tmpl = env.get_template("chat")
-                            );
+        .map_err(|e| format!("Template get error: {e}"))?;
-                            for tool in tools_list {
+
-                                tools_text.push('\n');
+    // Build messages as serde_json::Value for minijinja
-                                tools_text.push_str(
+    let mut msgs_val: Vec<Value> = Vec::new();
-                                    &serde_json::to_string(tool).unwrap_or_default(),
+    for msg in messages {
-                                );
+        let mut m: HashMap<String, Value> = HashMap::new();
-                            }
+        m.insert("role".into(), Value::from(msg.role.clone()));
-                            tools_text.push_str(
+
-                                "\n</tools>\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>",
+        let content = msg.content.clone().unwrap_or_default();
-                            );
+
-                            content.push_str(&tools_text);
+        // For assistant messages, check if there are tool_calls
+        if msg.role == "assistant" {
+            if let Some(tcs) = &msg.tool_calls {
+                // Serialise tool calls per the template's expected format
+                let tcs_val: Vec<Value> = tcs.iter().map(|tc| {
+                    let args: serde_json::Value =
+                        serde_json::from_str(&tc.function.arguments).unwrap_or_default();
+                    Value::from_serialize(&serde_json::json!({
+                        "id": tc.id,
+                        "type": "function",
+                        "function": {
+                            "name": tc.function.name,
+                            "arguments": args,
                         }
-                    }
+                    }))
-                }
+                }).collect();
-                prompt.push_str(&format!("<|im_start|>system\n{}<|im_end|>\n", content));
+                m.insert("tool_calls".into(), Value::from(tcs_val));
-            }
-            "user" => {
-                let content = msg.content.as_deref().unwrap_or("");
-                if msg.tool_call_id.is_some() || msg.name.is_some() {
-                    prompt.push_str(&format!(
-                        "<|im_start|>user\n<tool_response>\n{}\n</tool_response><|im_end|>\n",
-                        content
-                    ));
-                } else {
-                    prompt.push_str(&format!("<|im_start|>user\n{}<|im_end|>\n", content));
-                }
-            }
-            "assistant" => {
-                let content = msg.content.as_deref().unwrap_or("");
-                if let Some(tcs) = &msg.tool_calls {
-                    let mut asst = format!("<|im_start|>assistant\n{}", content);
-                    for tc in tcs {
-                        let args: serde_json::Value =
-                            serde_json::from_str(&tc.function.arguments).unwrap_or_default();
-                        asst.push_str(&format!(
-                            "<tool_call>\n<function={}>\n",
-                            tc.function.name
-                        ));
-                        if let Some(obj) = args.as_object() {
-                            for (k, v) in obj {
-                                let val = match v {
-                                    serde_json::Value::String(s) => s.clone(),
-                                    other => serde_json::to_string(other).unwrap_or_default(),
-                                };
-                                asst.push_str(&format!("<parameter={}>\n{}\n</parameter>\n", k, val));
-                            }
-                        }
-                        asst.push_str("</function>\n</tool_call>");
-                    }
-                    asst.push_str("<|im_end|>\n");
-                    prompt.push_str(&asst);
-                } else {
-                    prompt.push_str(&format!(
-                        "<|im_start|>assistant\n{}<|im_end|>\n",
-                        content
-                    ));
-                }
-            }
-            _ => {
-                let content = msg.content.as_deref().unwrap_or("");
-                prompt.push_str(&format!("<|im_start|>user\n{}<|im_end|>\n", content));
             }
         }
+
+        // Handle tool role messages
+        if msg.role == "tool" {
+            // Wrap in tool_response as the template expects
+            let wrapped = format!("<tool_response>\n{}\n</tool_response>", content);
+            m.insert("content".into(), Value::from(wrapped));
+        } else {
+            m.insert("content".into(), Value::from(content));
+        }
+
+        msgs_val.push(Value::from(m));
     }
 
-    // Generation prompt: non-thinking mode
+    // BOS token for sentencepiece / unigram models
-    prompt.push_str("<|im_start|>assistant\n<think>\n\n</think>\n\n");
+    let bos_token: &str = "<s>";
-    prompt
+
+    // Build context
+    let mut ctx: HashMap<String, Value> = HashMap::new();
+    ctx.insert("bos_token".into(), Value::from(bos_token));
+    ctx.insert("messages".into(), Value::from(msgs_val));
+    ctx.insert("add_generation_prompt".into(), Value::from(true));
+    ctx.insert("enable_thinking".into(), Value::from(true));
+
+    // Render
+    let result = tmpl
+        .render(&ctx)
+        .map_err(|e| format!("Template render error: {e}"))?;
+
+    Ok(result)
 }
 
 /// Parameters for building a [`LlamaSampler`] chain.
@@ -135,7 +133,6 @@ pub fn build_sampler(params: &SamplerParams) -> LlamaSampler {
     let seed = params.seed;
     let mut samplers: Vec<LlamaSampler> = Vec::new();
 
-    // Repetition/frequency/presence penalties
     let repeat = repeat_penalty.unwrap_or(1.0);
     let freq = frequency_penalty.unwrap_or(0.0);
     let present = presence_penalty.unwrap_or(0.0);
@@ -143,22 +140,18 @@ pub fn build_sampler(params: &SamplerParams) -> LlamaSampler {
         samplers.push(LS::penalties(64, repeat, freq, present));
     }
 
-    // top_k
     if let Some(k) = top_k {
         samplers.push(LS::top_k(k as i32));
     }
 
-    // top_p
     if let Some(p) = top_p {
         samplers.push(LS::top_p(p, 1));
     }
 
-    // min_p
     if let Some(p) = min_p {
         samplers.push(LS::min_p(p, 1));
     }
 
-    // Temperature + final selector
     let temp = temperature.unwrap_or(0.0);
     if temp <= 0.0 {
         samplers.push(LS::greedy());
@@ -173,19 +166,36 @@ pub fn build_sampler(params: &SamplerParams) -> LlamaSampler {
     LlamaSampler::chain_simple(samplers)
 }
 
-
 // ═══════════════════════════════════════════════════════════════
 //  TEXT PROCESSING
 // ═══════════════════════════════════════════════════════════════
 
-/// Remove special tokens from generated text.
+/// Remove special tokens from generated text and separate reasoning.
-pub fn clean_text(text: &str) -> String {
+///
-    text.replace("<|im_end|>", "")
+/// For thinking models, returns (reasoning, cleaned_answer).
-        .replace("<|im_start|>", "")
+pub fn clean_text(text: &str) -> (String, String) {
-        .replace("<think>", "")
+    let text = text.replace("<|im_end|>", "")
-        .replace("</think>", "")
+        .replace("<|im_start|>", "");
-        .trim()
+
-        .to_string()
+    // Separate reasoning (between <think>/</think>) from answer
+    let text = text.trim();
+    let (reasoning, answer) = if let Some(close_idx) = text.find("</think>") {
+        let reasoning = text[..close_idx].trim()
+            .trim_start_matches("<think>")
+            .trim()
+            .to_string();
+        let answer = text[close_idx + 8..].trim().to_string();
+        (reasoning, answer)
+    } else if text.contains("<think>") {
+        // Still thinking — everything is reasoning
+        let reasoning = text.trim_start_matches("<think>").trim().to_string();
+        (reasoning, String::new())
+    } else {
+        (String::new(), text.to_string())
+    };
+
+    let answer = answer.replace("<think>", "").replace("</think>", "");
+    (reasoning, answer.trim().to_string())
 }
 
 /// Parse tool calls from generated text in the format:
@@ -264,7 +274,6 @@ pub fn parse_tool_calls(text: &str) -> (String, Vec<ToolCall>) {
                 continue;
             }
         }
-        // Save last param
         if let Some(p) = current_param.take() {
             args_map.insert(p, serde_json::Value::String(current_value.trim().to_string()));
         }
@@ -283,12 +292,9 @@ pub fn parse_tool_calls(text: &str) -> (String, Vec<ToolCall>) {
         idx = end;
     }
 
-    // Remove tool_call blocks from the text
     clean = clean.replace("<tool_call>", "").replace("</tool_call>", "");
-    // XML-like tags are already fully parsed; remaining text is the content
     clean = clean.trim().to_string();
-    // Strip remaining XML tags that aren't part of clean
+    let (_reasoning, cleaned) = clean_text(&clean);
-    let cleaned = clean_text(&clean);
 
     (cleaned, tool_calls)
 }

src/config/mod.rs

@@ -4,8 +4,8 @@
 
 use std::sync::LazyLock;
 
-const DEFAULT_MODEL_PATH: &str = "/models/MiniCPM-V-4.6-Q4_K_M.gguf";
+const DEFAULT_MODEL_PATH: &str = "/models/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-Q8_0.gguf";
-pub const MODEL_ID: &str = "minicpm-v-4.6";
+pub const MODEL_ID: &str = "minicpm5-1b-fable5-v2-thinking";
 
 /// Application configuration loaded at startup from environment variables.
 #[derive(Debug, Clone)]
@@ -44,7 +44,7 @@ impl AppConfig {
                 .and_then(|v| v.parse().ok())
                 .unwrap_or(8080),
             log_level: std::env::var("RUST_LOG").unwrap_or_else(|_| "info".to_string()),
-            n_ctx: 2048,
+            n_ctx: 8192,
             n_batch: 512,
             n_threads: 4,
         }

src/domain/entity/mod.rs

@@ -96,8 +96,11 @@ pub struct Choice {
 #[derive(Serialize)]
 pub struct ResponseMessage {
     pub role: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub content: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning_content: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_calls: Option<Vec<ToolCall>>,
 }
 
@@ -165,6 +168,8 @@ pub struct SseDelta {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub content: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning_content: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_calls: Option<Vec<ToolCall>>,
 }
 

src/infrastructure/llama/engine.rs

@@ -219,6 +219,16 @@ impl LlamaEngine {
 
         let mut current = inner.sample(sampler);
 
+        // Skip leading EOS tokens (like <|im_end|> as first token)
+        while output.is_empty() && self.model.is_eog_token(current) {
+            let pos = input_tokens.len() as i32 + output.len() as i32;
+            if let Err(e) = inner.decode(current, pos) {
+                tracing::info!("  Decode error: {e}");
+                break;
+            }
+            current = inner.sample(sampler);
+        }
+
         for _ in 0..max_tokens {
             if self.model.is_eog_token(current) {
                 break;

src/presentation/handler/chat.rs

@@ -5,6 +5,7 @@ use std::sync::Arc;
 
 use axum::extract::State;
 use axum::response::sse::{Event, KeepAlive, Sse};
+use axum::http::HeaderMap;
 use axum::{Json, response::{IntoResponse, Response}};
 use chrono::Utc;
 use tokio::sync::mpsc;
@@ -26,7 +27,8 @@ pub async fn chat_completions(
 ) -> Result<Response, AppError> {
     let max_tokens = req.max_tokens.unwrap_or(256).min(1024);
     let stop = req.stop.clone().unwrap_or_default();
-    let prompt = chat::build_prompt(&req.messages, &req.tools);
+    let prompt = chat::build_prompt(&state.engine.model, &req.messages, &req.tools)
+        .map_err(|e| AppError::LlmError(e))?;
 
     // Tokenize
     let input_tokens = state
@@ -71,7 +73,8 @@ async fn handle_non_streaming(
         .generate(&input_tokens, &mut sampler, max_tokens, &stop)
         .await?;
 
-    let (output_text, tool_calls) = chat::parse_tool_calls(&chat::clean_text(&raw_text));
+    let (reasoning, cleaned) = chat::clean_text(&raw_text);
+    let (output_text, tool_calls) = chat::parse_tool_calls(&cleaned);
 
     let completion_tokens = output_tokens.len() as u32;
     info!("  {} generated tokens", completion_tokens);
@@ -84,6 +87,8 @@ async fn handle_non_streaming(
         "length"
     };
 
+    let reasoning_opt = if reasoning.is_empty() { None } else { Some(reasoning) };
+
     Ok(Json(ChatResponse {
         id: chat_id,
         object: "chat.completion".into(),
@@ -94,6 +99,7 @@ async fn handle_non_streaming(
             message: ResponseMessage {
                 role: "assistant".into(),
                 content: Some(output_text),
+                reasoning_content: reasoning_opt,
                 tool_calls: if tool_calls.is_empty() {
                     None
                 } else {
@@ -140,6 +146,7 @@ async fn handle_streaming(
                     role: Some("assistant".into()),
                     content: None,
                     tool_calls: None,
+                    reasoning_content: None,
                 },
                 finish_reason: None,
             }],
@@ -162,6 +169,8 @@ async fn handle_streaming(
 
         let mut count = 0u32;
         let mut text_buf = String::new();
+        let mut sent_len: usize = 0;     // how many chars of text_buf have been sent
+        let mut think_done: bool = false; // true once </think> seen
         let mut current = inner.sample(&mut sampler);
 
         loop {
@@ -177,6 +186,7 @@ async fn handle_streaming(
                             role: None,
                             content: None,
                             tool_calls: None,
+                            reasoning_content: None,
                         },
                         finish_reason: Some("length".into()),
                     }],
@@ -186,6 +196,40 @@ async fn handle_streaming(
                 break;
             }
 
+            // Skip leading EOS tokens (like <|im_end|> at start of generation)
+            if state.engine.is_eog(current) && text_buf.is_empty() {
+                // safety bound: don't skip more than 10
+                if count >= max_tokens || count > 10 {
+                    let chunk = serde_json::to_string(&SseChunk {
+                        id: chat_id.clone(),
+                        object: "chat.completion.chunk".into(),
+                        created,
+                        model: model_name.clone(),
+                        choices: vec![SseChoice {
+                            index: 0,
+                            delta: SseDelta {
+                                role: None,
+                                content: None,
+                                tool_calls: None,
+                                reasoning_content: None,
+                            },
+                            finish_reason: Some("stop".into()),
+                        }],
+                    })
+                    .unwrap();
+                    let _ = tx.send(Ok(Event::default().data(chunk))).await;
+                    break;
+                }
+                count += 1;
+                let pos = input_tokens.len() as i32 + count as i32;
+                if let Err(e) = inner.decode(current, pos) {
+                    info!("  Decode error: {e}");
+                    break;
+                }
+                current = inner.sample(&mut sampler);
+                continue;
+            }
+
             if state.engine.is_eog(current) {
                 let reason = if has_tools && text_buf.contains("<tool_call>") {
                     "tool_calls"
@@ -203,6 +247,7 @@ async fn handle_streaming(
                             role: None,
                             content: None,
                             tool_calls: None,
+                            reasoning_content: None,
                         },
                         finish_reason: Some(reason.into()),
                     }],
@@ -213,9 +258,97 @@ async fn handle_streaming(
             }
 
             let piece = state.engine.decode_token(current);
-            let content = chat::clean_text(&piece);
 
-            if !content.is_empty() {
+            // Push into buffer
+            text_buf.push_str(&piece);
+
+            // Detect </think> transition
+            if !think_done && text_buf.contains("</think>") {
+                think_done = true;
+            }
+
+            // Find new text since last send
+            let new_text = &text_buf[sent_len..]; // everything not yet streamed
+            if new_text.is_empty() {
+                // Nothing new to send; skip straight to decode
+                let pos = input_tokens.len() as i32 + count as i32;
+                if let Err(e) = inner.decode(current, pos) {
+                    info!("  Decode error: {e}");
+                    break;
+                }
+                count += 1;
+                current = inner.sample(&mut sampler);
+                continue;
+            }
+
+            // Strip special tokens from the NEW text chunk
+            // Split at </think> if present — before goes to reasoning, after to content
+            let (reasoning_part, content_part) = if let Some(pos) = new_text.find("</think>") {
+                let before = new_text[..pos]
+                    .replace("<|im_end|>", "")
+                    .replace("<|im_start|>", "")
+                    .replace("<think>", "")
+                    .trim()
+                    .to_string();
+                let after = new_text[pos + 8..]
+                    .replace("<|im_end|>", "")
+                    .replace("<|im_start|>", "")
+                    .replace("<think>", "")
+                    .trim()
+                    .to_string();
+                think_done = true;
+                (Some(before), after)
+            } else if think_done {
+                let cleaned = new_text
+                    .replace("<|im_end|>", "")
+                    .replace("<|im_start|>", "")
+                    .replace("<think>", "")
+                    .trim()
+                    .to_string();
+                (None, cleaned)
+            } else {
+                let cleaned = new_text
+                    .replace("<|im_end|>", "")
+                    .replace("<|im_start|>", "")
+                    .replace("<think>", "")
+                    .trim()
+                    .to_string();
+                (Some(cleaned), String::new())
+            };
+
+            // Send reasoning part (before </think>, or entire text if still thinking)
+            if let Some(ref r) = reasoning_part {
+                if !r.is_empty() {
+                    let delta = SseDelta {
+                        role: None,
+                        content: None,
+                        reasoning_content: Some(r.clone()),
+                        tool_calls: None,
+                    };
+                    let chunk = serde_json::to_string(&SseChunk {
+                        id: chat_id.clone(),
+                        object: "chat.completion.chunk".into(),
+                        created,
+                        model: model_name.clone(),
+                        choices: vec![SseChoice {
+                            index: 0,
+                            delta,
+                            finish_reason: None,
+                        }],
+                    })
+                    .unwrap();
+                    let _ = tx.send(Ok(Event::default().data(chunk))).await;
+                }
+            }
+
+            // Send content part (after </think>, or never if model doesn't think)
+            if !content_part.is_empty() {
+                let delta = SseDelta {
+                    role: None,
+                    content: Some(content_part),
+                    reasoning_content: None,
+                    tool_calls: None,
+                };
                 let chunk = serde_json::to_string(&SseChunk {
                     id: chat_id.clone(),
                     object: "chat.completion.chunk".into(),
@@ -223,11 +356,7 @@ async fn handle_streaming(
                     model: model_name.clone(),
                     choices: vec![SseChoice {
                         index: 0,
-                        delta: SseDelta {
+                        delta,
-                            role: None,
-                            content: Some(content.clone()),
-                            tool_calls: None,
-                        },
                         finish_reason: None,
                     }],
                 })
@@ -237,7 +366,7 @@ async fn handle_streaming(
                 }
             }
 
-            text_buf.push_str(&piece);
+            sent_len = text_buf.len();
 
             // Check stop sequences
             let mut stop_now = false;
@@ -259,6 +388,7 @@ async fn handle_streaming(
                             role: None,
                             content: None,
                             tool_calls: None,
+                            reasoning_content: None,
                         },
                         finish_reason: Some("stop".into()),
                     }],
@@ -284,6 +414,7 @@ async fn handle_streaming(
                                 role: None,
                                 content: None,
                                 tool_calls: None,
+                                reasoning_content: None,
                             },
                             finish_reason: Some("tool_calls".into()),
                         }],
@@ -305,6 +436,10 @@ async fn handle_streaming(
     });
 
     let stream = ReceiverStream::new(rx);
+    let mut headers = HeaderMap::new();
+    headers.insert("X-Accel-Buffering", "no".parse().unwrap());
+    headers.insert("Cache-Control", "no-cache".parse().unwrap());
+    headers.insert("Connection", "keep-alive".parse().unwrap());
     let sse = Sse::new(stream).keep_alive(KeepAlive::default());
-    Ok(sse.into_response())
+    Ok((headers, sse).into_response())
 }

src/presentation/handler/health.rs

@@ -8,6 +8,6 @@ use crate::domain::entity::HealthResponse;
 pub async fn health_check() -> Json<HealthResponse> {
     Json(HealthResponse {
         status: "ok".into(),
-        model: format!("{MODEL_ID}-q4_k_m"),
+        model: format!("{MODEL_ID}-q8_0"),
     })
 }