fix(stream): robust reasoning/content split + flush un-tagged output

The incremental </think> search missed tags split across tokens, inverting
reasoning/content classification. Detect the first </think> on the full
buffer and track the content boundary as a byte offset. If the model never
closes </think>, flush the buffered text as content so clients always
receive the response. Chat UI now renders reasoning_content too.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
asepharyana
2026-08-03 11:03:26 +07:00
co-authored by Claude Opus 5
parent b636496497
commit 7cec411cba
4 changed files with 134 additions and 50 deletions
+2 -2
View File
@@ -1,6 +1,6 @@
pub mod use_cases;
pub use use_cases::{
build_prompt, build_sampler, clean_text, parse_tool_calls, split_stream_chunk, validate_model,
SamplerParams,
build_prompt, build_sampler, clean_text, parse_tool_calls, split_stream_chunk, strip_markup,
validate_model, SamplerParams,
};
+58 -35
View File
@@ -335,7 +335,7 @@ pub fn parse_tool_calls(text: &str) -> (String, Vec<ToolCall>) {
/// Handles both fixed tags (`<|im_end|>`, `<think>`, `<tool_call>`, …) and the
/// attribute-bearing openers used by this model's tool format (`<function=…>`,
/// `<parameter=…>`), even when a tag straddles a token boundary.
fn strip_markup(text: &str) -> String {
pub fn strip_markup(text: &str) -> String {
let mut out = String::with_capacity(text.len());
let mut rest = text;
@@ -383,22 +383,42 @@ fn strip_markup(text: &str) -> String {
out
}
/// Split an incremental streamed text fragment into `(reasoning, content)`
/// deltas for SSE.
/// Compute `(reasoning, content)` deltas for an incremental streamed fragment.
///
/// * `think_done` means the `</think>` boundary was already crossed **before**
/// this fragment (i.e. it is not the chunk containing the first `</think>`).
/// * Whitespace **inside** a fragment is preserved — only the whitespace
/// sitting immediately around the `</think>` boundary is trimmed, so
/// reasoning does not end with a dangling newline and content does not begin
/// with one. (Trimming every fragment corrupted inter-word spaces.)
/// * Before the first `</think>`, everything is emitted as `reasoning_content`;
/// after it, as `content`. A stray second `</think>` is stripped, not split.
pub fn split_stream_chunk(new_text: &str, think_done: bool) -> (Option<String>, String) {
if !think_done {
if let Some(pos) = new_text.find("</think>") {
let mut before = strip_markup(&new_text[..pos]);
let mut after = strip_markup(&new_text[pos + 8..]);
/// * `new_text` — the not-yet-emitted fragment (`text_buf[sent_len..]`).
/// * `sent_len` — byte offset in the full buffer where `new_text` begins.
/// * `content_start` — byte offset in the full buffer where the content phase
/// begins (immediately after the first `</think>`); `None` while still
/// thinking.
///
/// The boundary is a position in the *full* buffer, not a string search in the
/// fragment — this stays correct even when `</think>` is split across tokens.
/// Whitespace inside a fragment is preserved; only the edges around the
/// boundary are trimmed. A stray second `</think>` is stripped, not re-split.
pub fn split_stream_chunk(
new_text: &str,
sent_len: usize,
content_start: Option<usize>,
) -> (Option<String>, String) {
match content_start {
// Still thinking — everything is reasoning.
None => {
let cleaned = strip_markup(new_text);
let reasoning = if cleaned.is_empty() {
None
} else {
Some(cleaned)
};
(reasoning, String::new())
}
// Boundary already emitted — everything is content.
Some(cs) if cs <= sent_len => (None, strip_markup(new_text)),
// Boundary falls inside this fragment (or beyond it, defensively).
Some(cs) => {
let rel = (cs - sent_len).min(new_text.len());
let (before, after) = new_text.split_at(rel);
let mut before = strip_markup(before);
let mut after = strip_markup(after);
while before.ends_with(['\n', ' ', '\t']) {
before.pop();
@@ -413,19 +433,7 @@ pub fn split_stream_chunk(new_text: &str, think_done: bool) -> (Option<String>,
Some(before)
};
(reasoning, after)
} else {
// Still thinking — everything is reasoning.
let cleaned = strip_markup(new_text);
let reasoning = if cleaned.is_empty() {
None
} else {
Some(cleaned)
};
(reasoning, String::new())
}
} else {
// Think phase already ended — everything is content; strip stray markup.
(None, strip_markup(new_text))
}
}
@@ -471,7 +479,7 @@ mod tests {
#[test]
fn split_chunk_before_think_is_reasoning() {
let (reasoning, content) = split_stream_chunk("Hello ", false);
let (reasoning, content) = split_stream_chunk("Hello ", 0, None);
assert_eq!(reasoning.as_deref(), Some("Hello "));
assert_eq!(content, "");
}
@@ -479,21 +487,23 @@ mod tests {
#[test]
fn split_chunk_preserves_internal_spaces() {
// Regression: trimming every fragment used to eat inter-word spaces.
let (r1, _) = split_stream_chunk("Hello", false);
let (r2, _) = split_stream_chunk(" world", false);
let (r1, _) = split_stream_chunk("Hello", 0, None);
let (r2, _) = split_stream_chunk(" world", 5, None);
assert_eq!(format!("{}{}", r1.unwrap(), r2.unwrap()), "Hello world");
}
#[test]
fn split_chunk_boundary_trims_only_edges() {
let (reasoning, content) = split_stream_chunk("question\n</think>\n\nAnswer ", false);
// text_buf = "question\n</think>\n\nAnswer "; boundary right after the
// tag at byte 17.
let (reasoning, content) = split_stream_chunk("question\n</think>\n\nAnswer ", 0, Some(17));
assert_eq!(reasoning.as_deref(), Some("question"));
assert_eq!(content, "Answer ");
}
#[test]
fn split_chunk_after_think_is_content() {
let (reasoning, content) = split_stream_chunk(" answer", true);
let (reasoning, content) = split_stream_chunk(" answer", 0, Some(0));
assert_eq!(reasoning, None);
assert_eq!(content, " answer");
}
@@ -502,16 +512,29 @@ mod tests {
fn split_chunk_stray_think_tag_is_stripped_not_split() {
// A second </think> (already past the boundary) must not restart
// reasoning classification.
let (reasoning, content) = split_stream_chunk("...</think>more", true);
let (reasoning, content) = split_stream_chunk("...</think>more", 0, Some(0));
assert_eq!(reasoning, None);
assert_eq!(content, "...more");
}
#[test]
fn split_chunk_boundary_straddling_tokens() {
// `</think>` split as "</think" + ">" across two fragments: the boundary
// is detected on the full buffer, so the answer still becomes content.
let (r1, _) = split_stream_chunk("reasoning...</think", 0, None);
assert!(r1.is_some());
let (r2, c2) = split_stream_chunk("\n\n2 + 2 = 4.", 18, Some(18));
assert_eq!(r2, None);
assert_eq!(c2, "\n\n2 + 2 = 4.");
}
#[test]
fn split_chunk_strips_special_and_markup() {
// boundary at byte 30 (right after "</think>").
let (reasoning, content) = split_stream_chunk(
"<|im_end|><think>Hello</think>\n<tool_call><function=get_weather>",
false,
0,
Some(30),
);
assert_eq!(reasoning.as_deref(), Some("Hello"));
assert_eq!(content, "");
+40 -6
View File
@@ -72,6 +72,13 @@
display: flex; align-items: center; gap: 6px;
}
.msg .tool-call::before { content: '\1F527'; }
.msg .reasoning {
font-size: 12px; font-style: italic;
color: var(--text2);
white-space: pre-wrap; word-break: break-word;
border-bottom: 1px solid var(--border);
padding-bottom: 8px; margin-bottom: 8px;
}
.msg.error {
background: #2a1818; border-color: #4a2828; color: #f08080;
}
@@ -186,8 +193,33 @@ async function send() {
const decoder = new TextDecoder();
let buffer = '';
let full = '';
let reasoning = '';
let toolCalls = null;
// Re-render the assistant bubble: reasoning (muted) above the answer.
function render() {
el.innerHTML = '';
if (reasoning) {
const r = document.createElement('div');
r.className = 'reasoning';
r.textContent = reasoning;
el.appendChild(r);
}
if (full) {
const c = document.createElement('div');
c.textContent = full;
el.appendChild(c);
}
if (toolCalls?.length) {
for (const tc of toolCalls) {
const t = document.createElement('div');
t.className = 'tool-call';
t.textContent = 'Calling tool: ' + (tc.function?.name || 'tool');
el.appendChild(t);
}
}
}
while (true) {
const { done, value } = await reader.read();
if (done) break;
@@ -206,18 +238,20 @@ async function send() {
const delta = chunk.choices?.[0]?.delta;
const finish = chunk.choices?.[0]?.finish_reason;
if (delta?.reasoning_content) {
reasoning += delta.reasoning_content;
render();
}
if (delta?.content) {
full += delta.content;
el.textContent = full;
render();
}
if (delta?.tool_calls) {
toolCalls = delta.tool_calls;
render();
}
if (finish === 'tool_calls' && toolCalls) {
const t = document.createElement('div');
t.className = 'tool-call';
t.textContent = 'Calling tool: ' + toolCalls.map(tc => tc.function?.name).join(', ');
el.appendChild(t);
if (finish === 'tool_calls') {
render();
}
} catch (e) { /* skip malformed chunk */ }
}
+34 -7
View File
@@ -176,7 +176,9 @@ async fn handle_streaming(
let mut sampler = SendSampler(chat::build_sampler(&params));
let mut text_buf = String::new();
let mut sent_len: usize = 0;
let mut think_done = false;
// Byte offset in text_buf where the content phase begins (right after
// the first `</think>`); None while still thinking.
let mut content_start: Option<usize> = None;
let outcome = engine.generate(
&input_tokens,
@@ -187,11 +189,13 @@ async fn handle_streaming(
&mut |_token, piece| {
text_buf.push_str(piece);
// `was_thinking` is passed to split_stream_chunk so the chunk
// containing the first </think> is treated as the boundary.
let was_thinking = !think_done;
if !think_done && text_buf.contains("</think>") {
think_done = true;
// Robust boundary detection on the *full* buffer — a `</think>`
// tag may be split across tokens, which would defeat a search
// over the incremental fragment only.
if content_start.is_none() {
if let Some(pos) = text_buf.find("</think>") {
content_start = Some(pos + 8);
}
}
let new_text = &text_buf[sent_len..];
@@ -199,7 +203,8 @@ async fn handle_streaming(
return true;
}
let (reasoning, content) = chat::split_stream_chunk(new_text, was_thinking);
let (reasoning, content) =
chat::split_stream_chunk(new_text, sent_len, content_start);
if let Some(reasoning) = reasoning {
let chunk = SseChunk::delta(
@@ -251,6 +256,28 @@ async fn handle_streaming(
total_tokens: prompt_tokens + completion_tokens,
};
// If the model never emitted `</think>`, everything was streamed
// as reasoning_content. Flush it as content so the client always
// receives the response text.
if content_start.is_none() && !text_buf.is_empty() {
let cleaned = chat::strip_markup(&text_buf);
if !cleaned.is_empty() {
let chunk = SseChunk::delta(
chat_id.clone(),
created,
model_name.clone(),
SseDelta {
role: None,
content: Some(cleaned),
tool_calls: None,
reasoning_content: None,
},
);
let event = serde_json::to_string(&chunk).unwrap();
let _ = tx.blocking_send(Ok(Event::default().data(event)));
}
}
// Single-shot tool-calls delta (this model emits whole blocks).
let mut sent_tool_calls = false;
if outcome.finish == FinishReason::ToolCalls {