Skip to content

GitLab

"...git@developer.sourcefind.cn:2222/OpenDAS/vllm_cscc.git" did not exist on "0168f69e50898fd5f09ac64a0d735039e57e7806"
Unverified Commit d82b0050 authored by MatejKosec's avatar MatejKosec Committed by GitHub
Browse files

feat: interleaved thinking support in reasoning parser (#6422) 


Signed-off-by: default avatarMatej Kosec <mkosec@nvidia.com>
parent 7409bd3a
Hide whitespace changes
Inline Side-by-side
Showing with 662 additions and 105 deletions
lib/parsers/src/reasoning/base_parser.rs
View file @ d82b0050
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. // SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0 // SPDX-License-Identifier: Apache-2.0
//! # Reasoning and Tool Call Interplay
//!
//! Models like GLM-4.5/4.7 and Qwen3 interleave reasoning blocks with tool calls:
//!
//! ```text
//! <think>reasoning about what tool to call</think>
//! <tool_call>get_weather<arg_key>city</arg_key><arg_value>Beijing</arg_value></tool_call>
//! <think>reasoning about the result</think>
//! <tool_call>summarize<arg_key>text</arg_key><arg_value>...</arg_value></tool_call>
//! ```
//!
//! The reasoning parser and the tool call parser are **independent, sequential** stages:
//!
//! 1. **Reasoning parser** (`BasicReasoningParser`) splits the stream into:
//! - `reasoning_content`: everything inside `<think>...</think>` blocks
//! - `normal_text`: everything outside (including tool call tags)
//! 2. **Tool call parser** (`glm47` / others) then processes `normal_text` to extract
//! `<tool_call>...</tool_call>` blocks.
//!
//! This means tool calls **must** appear outside `<think>` blocks to be detected.
//! If a model erroneously emits a tool call inside a `<think>` block (observed in
//! GLM-4.7 under very long contexts), the tool call parser will not see it.
//!
//! ## `force_reasoning` and tokenizer behavior
//!
//! Some models (e.g. GLM-5-FP8 served via ZAI) consume `<think>` as a special
//! tokenizer token and never emit it as literal text. In that case use
//! `force_reasoning=true` (`deepseek_r1` parser), which treats all output as
//! reasoning until `</think>` is seen. Models that do emit `<think>` as text
//! (standard serving, Qwen3, GLM-4.5) should use `force_reasoning=false`
//! (`glm45`, `nemotron_deci`, `qwen3` parsers).
use crate::{ParserResult, ReasoningParser}; use crate::{ParserResult, ReasoningParser};
/// Returns the length of the longest suffix of `s` that is also a prefix of `delim`.
///
/// Ported from ollama's `thinking/parser.go::overlap()`. Used to detect partial
/// tags split across streaming chunk boundaries (e.g., `"Hello world <th"` where
/// `<th` is a prefix of `<think>`).
fn overlap(s: &str, delim: &str) -> usize {
let max = delim.len().min(s.len());
for i in (1..=max).rev() {
if !delim.is_char_boundary(i) {
continue; // Skip mid-codepoint positions (e.g., multi-byte `◁` in Kimi tags)
}
if s.ends_with(&delim[..i]) {
return i;
}
}
0
}
#[derive(Default, Debug, Clone)] #[derive(Default, Debug, Clone)]
pub struct BasicReasoningParser { pub struct BasicReasoningParser {
think_start_token: String, think_start_token: String,
...@@ -33,7 +83,8 @@ impl BasicReasoningParser { ...@@ -33,7 +83,8 @@ impl BasicReasoningParser {
impl ReasoningParser for BasicReasoningParser { impl ReasoningParser for BasicReasoningParser {
fn detect_and_parse_reasoning(&mut self, text: &str, _token_ids: &[u32]) -> ParserResult { fn detect_and_parse_reasoning(&mut self, text: &str, _token_ids: &[u32]) -> ParserResult {
let in_reasoning = self._in_reasoning || text.contains(&self.think_start_token); let has_think_tag = text.contains(&self.think_start_token);
let in_reasoning = self._in_reasoning || has_think_tag;
if !in_reasoning { if !in_reasoning {
return ParserResult { return ParserResult {
normal_text: text.to_string(), normal_text: text.to_string(),
...@@ -41,24 +92,53 @@ impl ReasoningParser for BasicReasoningParser { ...@@ -41,24 +92,53 @@ impl ReasoningParser for BasicReasoningParser {
}; };
} }
// The text is considered to be in a reasoning block. // If force_reasoning and no start tag, treat entire text as reasoning
let processed_text = text.replace(&self.think_start_token, "").trim().to_string(); if self._in_reasoning && !has_think_tag && !text.contains(&self.think_end_token) {
if !processed_text.contains(&self.think_end_token) {
// Assume reasoning was truncated before `think_end_token`
return ParserResult { return ParserResult {
normal_text: String::new(), normal_text: String::new(),
reasoning_text: processed_text, reasoning_text: text.to_string(),
}; };
} }
// Extract reasoning content // Extract all <think>...</think> pairs using cursor-based iteration
let splits: Vec<&str> = processed_text.splitn(2, &self.think_end_token).collect(); let mut reasoning_parts = Vec::new();
let reasoning_text = splits.first().unwrap_or(&"").to_string(); let mut normal_parts = Vec::new();
let normal_text = splits let mut cursor = 0;
.get(1) let mut currently_reasoning = self._in_reasoning;
.map(|s| s.trim().to_string())
.unwrap_or_default(); while cursor < text.len() {
if currently_reasoning {
// We're inside a reasoning block — look for end token
if let Some(end_offset) = text[cursor..].find(&self.think_end_token) {
reasoning_parts.push(&text[cursor..cursor + end_offset]);
cursor += end_offset + self.think_end_token.len();
currently_reasoning = false;
} else {
// No end token — rest is reasoning (truncated)
reasoning_parts.push(&text[cursor..]);
cursor = text.len();
}
} else {
// We're in normal text — look for start token
if let Some(start_offset) = text[cursor..].find(&self.think_start_token) {
normal_parts.push(&text[cursor..cursor + start_offset]);
cursor += start_offset + self.think_start_token.len();
currently_reasoning = true;
} else {
// No more think blocks — rest is normal text
normal_parts.push(&text[cursor..]);
cursor = text.len();
}
}
}
let reasoning_text = reasoning_parts.join("").trim().to_string();
let normal_text = normal_parts.join("").trim().to_string();
// Note: self._in_reasoning is intentionally NOT updated here. This method is
// documented to "reset or ignore internal streaming state" (see trait doc). Callers
// should not mix detect_and_parse_reasoning with parse_reasoning_streaming_incremental
// on the same parser instance.
ParserResult { ParserResult {
normal_text, normal_text,
...@@ -71,86 +151,93 @@ impl ReasoningParser for BasicReasoningParser { ...@@ -71,86 +151,93 @@ impl ReasoningParser for BasicReasoningParser {
text: &str, text: &str,
_token_ids: &[u32], _token_ids: &[u32],
) -> ParserResult { ) -> ParserResult {
// Incrementally parse the streaming text
self._buffer.push_str(text); self._buffer.push_str(text);
let mut current_text = self._buffer.to_string();
// If the current text is a prefix of the think token, keep buffering.
// Only buffer for start token if we haven't found it yet.
// Only buffer for end token if we're currently inside a reasoning block.
// After reasoning ends, all content passes through as normal text.
if !self.stripped_think_start
&& self.think_start_token.starts_with(&current_text)
&& self.think_start_token.as_str() != current_text.as_str()
{
return ParserResult {
normal_text: String::new(),
reasoning_text: String::new(),
};
}
if self._in_reasoning
&& self.think_end_token.starts_with(&current_text)
&& self.think_end_token.as_str() != current_text.as_str()
{
return ParserResult {
normal_text: String::new(),
reasoning_text: String::new(),
};
}
// Strip `<think>` token if present let mut accumulated_normal = String::new();
if !self.stripped_think_start && current_text.contains(&self.think_start_token) { let mut accumulated_reasoning = String::new();
current_text = current_text.replace(&self.think_start_token, "");
self._buffer = current_text.to_string(); // Loop to exhaust all state transitions within a single chunk. Without this,
self.stripped_think_start = true; // a chunk containing two complete <think>...</think> blocks would process only
self._in_reasoning = true; // the first transition and buffer the rest, risking content loss at end-of-stream.
} loop {
// Handle end of reasoning block let current_text = self._buffer.clone();
let mut think_end_idx = current_text.len();
if self._in_reasoning { // Strip leading <think> tag if not yet stripped. Handles two cases:
think_end_idx = current_text // 1. force_reasoning=true where the model also emits <think> as text
.find(&self.think_end_token) // 2. First call where <think> arrives at buffer position 0
.unwrap_or(current_text.len()); // Mid-text <think> (position > 0) falls through to the find() branch below.
} if !self.stripped_think_start
if self._in_reasoning && think_end_idx < current_text.len() { && current_text.starts_with(self.think_start_token.as_str())
let reasoning_text = &current_text[..think_end_idx]; {
self._buffer.clear(); self._buffer = current_text[self.think_start_token.len()..].to_string();
self._in_reasoning = false; self.stripped_think_start = true;
let start_idx = think_end_idx + self.think_end_token.len(); self._in_reasoning = true;
let normal_text = if start_idx < current_text.len() { continue;
&current_text[start_idx..]
} else {
""
};
return ParserResult {
normal_text: normal_text.to_string(),
reasoning_text: reasoning_text.to_string(),
};
}
// Continue with reasoning content
if self._in_reasoning && self.stream_reasoning {
// Stream the content immediately
let reasoning_text = current_text;
self._buffer.clear();
ParserResult {
normal_text: String::new(),
reasoning_text,
}
} else if !self._in_reasoning {
// If we're not in a reasoning block return as normal text
let normal_text = current_text;
self._buffer.clear();
ParserResult {
normal_text,
reasoning_text: String::new(),
} }
} else {
// If we are in a reasoning block but no end token is found, return the current buffer if self._in_reasoning {
ParserResult { if let Some(end_idx) = current_text.find(self.think_end_token.as_str()) {
normal_text: String::new(), // End of reasoning block: accumulate content and transition out.
reasoning_text: String::new(), accumulated_reasoning.push_str(&current_text[..end_idx]);
let after_end = end_idx + self.think_end_token.len();
self._buffer = current_text[after_end..].to_string();
self._in_reasoning = false;
self.stripped_think_start = false; // Allow detecting next <think> block
continue; // Process remainder — may contain further blocks
} else {
// No complete end token — check for partial at end of buffer
// (e.g., "reasoning content</th" where "</th" is a prefix of "</think>").
if self.stream_reasoning {
let ol = overlap(&current_text, &self.think_end_token);
if ol >= 2 {
let safe_end = current_text.len() - ol;
if safe_end > 0 {
accumulated_reasoning.push_str(&current_text[..safe_end]);
}
self._buffer = current_text[safe_end..].to_string();
} else {
accumulated_reasoning.push_str(&current_text);
self._buffer.clear();
}
}
// When stream_reasoning=false, buffer retains all content until
// </think> arrives — no overlap check needed.
break;
}
} else {
// Not in reasoning — look for the next <think> block.
if let Some(think_pos) = current_text.find(self.think_start_token.as_str()) {
accumulated_normal.push_str(&current_text[..think_pos]);
let after_start = think_pos + self.think_start_token.len();
self._buffer = current_text[after_start..].to_string();
self._in_reasoning = true;
self.stripped_think_start = true;
continue; // Process reasoning content
} else {
// No complete start token — check for partial at end of buffer
// (e.g., "Hello world <th" where "<th" is a prefix of "<think>").
// Require overlap >= 2 so a lone `<` passes through for tool call
// XML tags like `<invoke>` or `<minimax:tool_call>`.
let ol = overlap(&current_text, &self.think_start_token);
if ol >= 2 {
let safe_end = current_text.len() - ol;
if safe_end > 0 {
accumulated_normal.push_str(&current_text[..safe_end]);
}
self._buffer = current_text[safe_end..].to_string();
} else {
accumulated_normal.push_str(&current_text);
self._buffer.clear();
}
break;
}
} }
} }
ParserResult {
normal_text: accumulated_normal,
reasoning_text: accumulated_reasoning,
}
} }
} }
...@@ -222,9 +309,8 @@ mod tests { ...@@ -222,9 +309,8 @@ mod tests {
"<think>first reasoning</think> middle <think>second reasoning</think> end", "<think>first reasoning</think> middle <think>second reasoning</think> end",
&[], &[],
); );
// The current implementation only handles the first occurrence properly assert_eq!(result.normal_text, "middle end");
assert_eq!(result.normal_text, "middle second reasoning</think> end"); assert_eq!(result.reasoning_text, "first reasoningsecond reasoning");
assert_eq!(result.reasoning_text, "first reasoning");
} }
#[test] #[test]
...@@ -236,11 +322,11 @@ mod tests { ...@@ -236,11 +322,11 @@ mod tests {
assert_eq!(result1.normal_text, " middle"); assert_eq!(result1.normal_text, " middle");
assert_eq!(result1.reasoning_text, "first reasoning"); assert_eq!(result1.reasoning_text, "first reasoning");
// Basic parser assumes only one reasoning block at a time // Second reasoning block: space before <think> is normal prefix, reasoning extracted
let result2 = parser let result2 = parser
.parse_reasoning_streaming_incremental(" <think>second reasoning</think> end", &[]); .parse_reasoning_streaming_incremental(" <think>second reasoning</think> end", &[]);
assert_eq!(result2.normal_text, " <think>second reasoning</think> end"); assert_eq!(result2.reasoning_text, "second reasoning");
assert_eq!(result2.reasoning_text, ""); assert_eq!(result2.normal_text, " end"); // " " prefix + " end" suffix
} }
#[test] #[test]
...@@ -334,10 +420,11 @@ mod tests { ...@@ -334,10 +420,11 @@ mod tests {
"<think>outer <think>inner</think> reasoning</think> normal", "<think>outer <think>inner</think> reasoning</think> normal",
&[], &[],
); );
// Current implementation should handle this by finding the first closing tag // Cursor-based parsing: first <think> starts reasoning, first </think> ends it.
// "outer <think>inner" is reasoning (inner <think> is just text within reasoning).
// " reasoning</think> normal" is normal text (stray </think> passes through).
assert_eq!(result.reasoning_text, "outer <think>inner");
assert_eq!(result.normal_text, "reasoning</think> normal"); assert_eq!(result.normal_text, "reasoning</think> normal");
// All <think> tags are stripped, so <think>inner is not included
assert_eq!(result.reasoning_text, "outer inner");
} }
#[test] #[test]
...@@ -364,9 +451,10 @@ mod tests { ...@@ -364,9 +451,10 @@ mod tests {
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true); BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let result = parser let result = parser
.detect_and_parse_reasoning("<think>first <think>second reasoning</think> normal", &[]); .detect_and_parse_reasoning("<think>first <think>second reasoning</think> normal", &[]);
// Should handle by replacing all opening tags and using first closing tag // Cursor-based: first <think> opens reasoning, finds first </think>.
// Inner <think> is just text within the reasoning block.
assert_eq!(result.reasoning_text, "first <think>second reasoning");
assert_eq!(result.normal_text, "normal"); assert_eq!(result.normal_text, "normal");
assert_eq!(result.reasoning_text, "first second reasoning");
} }
#[test] #[test]
...@@ -399,7 +487,7 @@ mod tests { ...@@ -399,7 +487,7 @@ mod tests {
#[test] #[test]
fn test_streaming_reset_state_after_complete_block() { fn test_streaming_reset_state_after_complete_block() {
let mut parser = let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, false); BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
// Process complete reasoning block // Process complete reasoning block
let result1 = let result1 =
...@@ -412,12 +500,28 @@ mod tests { ...@@ -412,12 +500,28 @@ mod tests {
assert_eq!(result2.normal_text, " more normal text"); assert_eq!(result2.normal_text, " more normal text");
assert_eq!(result2.reasoning_text, ""); assert_eq!(result2.reasoning_text, "");
// Basic parser does not expect more than one reasoning block at a time // Subsequent reasoning blocks should now be parsed (interleaved thinking)
// So this should not affect the state // The leading " " before <think> is normal-text prefix; " final" is suffix.
let result3 = parser let result3 = parser
.parse_reasoning_streaming_incremental(" <think>new reasoning</think> final", &[]); .parse_reasoning_streaming_incremental(" <think>new reasoning</think> final", &[]);
assert_eq!(result3.normal_text, " <think>new reasoning</think> final"); assert_eq!(result3.reasoning_text, "new reasoning");
assert_eq!(result3.reasoning_text, ""); assert_eq!(result3.normal_text, " final"); // " " prefix + " final" suffix
// Same test with separate chunks for clarity
let mut parser2 =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser2.parse_reasoning_streaming_incremental("<think>first</think> normal", &[]);
assert_eq!(r1.reasoning_text, "first");
assert_eq!(r1.normal_text, " normal");
let r2 = parser2.parse_reasoning_streaming_incremental(" between", &[]);
assert_eq!(r2.normal_text, " between");
assert_eq!(r2.reasoning_text, "");
let r3 = parser2.parse_reasoning_streaming_incremental("<think>second</think> final", &[]);
assert_eq!(r3.reasoning_text, "second");
assert_eq!(r3.normal_text, " final");
} }
#[test] #[test]
...@@ -474,4 +578,455 @@ mod tests { ...@@ -474,4 +578,455 @@ mod tests {
let r6 = parser.parse_reasoning_streaming_incremental("invoke name=\"get_weather\">", &[]); let r6 = parser.parse_reasoning_streaming_incremental("invoke name=\"get_weather\">", &[]);
assert_eq!(r6.normal_text, "invoke name=\"get_weather\">"); assert_eq!(r6.normal_text, "invoke name=\"get_weather\">");
} }
#[test]
fn test_interleaved_streaming_across_chunks() {
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>thought 1</think>", &[]);
assert_eq!(r1.reasoning_text, "thought 1");
assert_eq!(r1.normal_text, "");
let r2 = parser.parse_reasoning_streaming_incremental(" answer 1 ", &[]);
assert_eq!(r2.normal_text, " answer 1 ");
assert_eq!(r2.reasoning_text, "");
let r3 = parser.parse_reasoning_streaming_incremental("<think>thought 2</think>", &[]);
assert_eq!(r3.reasoning_text, "thought 2");
assert_eq!(r3.normal_text, "");
let r4 = parser.parse_reasoning_streaming_incremental(" answer 2", &[]);
assert_eq!(r4.normal_text, " answer 2");
assert_eq!(r4.reasoning_text, "");
let r5 = parser.parse_reasoning_streaming_incremental("<think>thought 3</think>", &[]);
assert_eq!(r5.reasoning_text, "thought 3");
assert_eq!(r5.normal_text, "");
let r6 = parser.parse_reasoning_streaming_incremental(" final answer", &[]);
assert_eq!(r6.normal_text, " final answer");
assert_eq!(r6.reasoning_text, "");
}
#[test]
fn test_three_reasoning_blocks_non_streaming() {
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let result = parser.detect_and_parse_reasoning(
"<think>A</think> one <think>B</think> two <think>C</think> three",
&[],
);
assert_eq!(result.reasoning_text, "ABC");
assert_eq!(result.normal_text, "one two three");
}
#[test]
fn test_streaming_transition_chunk() {
// </think> and <think> arrive in the same chunk.
// With loop-based processing, the second block's opening content is emitted
// immediately (stream_reasoning=true) rather than buffered until the next call.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>first", &[]);
assert_eq!(r1.reasoning_text, "first");
// Mid-chunk transition: </think> then normal text then <think> with more content.
// The loop transitions out of reasoning, emits " middle " as normal text, enters
// the next reasoning block, and streams "second" immediately.
let r2 = parser.parse_reasoning_streaming_incremental("</think> middle <think>second", &[]);
assert_eq!(r2.reasoning_text, "second");
assert_eq!(r2.normal_text, " middle ");
// Continuation of second reasoning block
let r3 = parser.parse_reasoning_streaming_incremental(" more</think> end", &[]);
assert_eq!(r3.reasoning_text, " more");
assert_eq!(r3.normal_text, " end");
}
#[test]
fn test_interleaved_with_force_reasoning() {
// deepseek_r1 mode: force_reasoning=true, first tokens are reasoning without <think>
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), true, true);
// No <think> tag — treated as reasoning because force_reasoning=true
let r1 = parser.parse_reasoning_streaming_incremental("initial reasoning", &[]);
assert_eq!(r1.reasoning_text, "initial reasoning");
assert_eq!(r1.normal_text, "");
// End of forced reasoning block
let r2 = parser.parse_reasoning_streaming_incremental("</think> answer", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, " answer");
// Second reasoning block with explicit <think>
let r3 =
parser.parse_reasoning_streaming_incremental("<think>second thought</think> done", &[]);
assert_eq!(r3.reasoning_text, "second thought");
assert_eq!(r3.normal_text, " done");
}
#[test]
fn test_interleaved_partial_think_tag_between_blocks() {
// After first reasoning block, partial <think> tag arrives across chunks
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>first</think> normal", &[]);
assert_eq!(r1.reasoning_text, "first");
assert_eq!(r1.normal_text, " normal");
// Partial <think> prefix: "<th" (2 chars, meets threshold)
let r2 = parser.parse_reasoning_streaming_incremental("<th", &[]);
assert_eq!(r2.normal_text, "");
assert_eq!(r2.reasoning_text, "");
// Complete the tag
let r3 = parser.parse_reasoning_streaming_incremental("ink>second</think> end", &[]);
assert_eq!(r3.reasoning_text, "second");
assert_eq!(r3.normal_text, " end");
}
#[test]
fn test_lone_angle_bracket_between_reasoning_blocks() {
// A lone `<` between reasoning blocks should pass through (not buffer)
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>thought</think>", &[]);
assert_eq!(r1.reasoning_text, "thought");
// Lone `<` must not be buffered — could be a tool call
let r2 = parser.parse_reasoning_streaming_incremental("<", &[]);
assert_eq!(r2.normal_text, "<");
assert_eq!(r2.reasoning_text, "");
let r3 = parser.parse_reasoning_streaming_incremental("tool_call>", &[]);
assert_eq!(r3.normal_text, "tool_call>");
assert_eq!(r3.reasoning_text, "");
// But a real <think> should still work after
let r4 =
parser.parse_reasoning_streaming_incremental("<think>more thought</think> done", &[]);
assert_eq!(r4.reasoning_text, "more thought");
assert_eq!(r4.normal_text, " done");
}
#[test]
fn test_force_reasoning_stream_false_buffers_until_end_token() {
// force_reasoning=true, stream_reasoning=false: content is buffered until </think>
// arrives, then returned as a single chunk. This is the expected behavior.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), true, false);
// No <think> — forced into reasoning, stream_reasoning=false means buffer silently
let r1 = parser.parse_reasoning_streaming_incremental("chunk one", &[]);
assert_eq!(r1.reasoning_text, "");
assert_eq!(r1.normal_text, "");
let r2 = parser.parse_reasoning_streaming_incremental(" chunk two", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, "");
// </think> arrives — entire buffered reasoning is flushed
let r3 = parser.parse_reasoning_streaming_incremental("</think> answer", &[]);
assert_eq!(r3.reasoning_text, "chunk one chunk two");
assert_eq!(r3.normal_text, " answer");
}
#[test]
fn test_multiple_full_blocks_in_single_streaming_chunk() {
// Two complete <think>...</think> blocks arrive in one chunk.
// The loop exhausts all transitions in a single call — both blocks are fully
// processed and no follow-up call is needed to flush buffered content.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental(
"<think>A</think> mid <think>B</think> end",
&[],
);
assert_eq!(r1.reasoning_text, "AB");
assert_eq!(r1.normal_text, " mid end");
// Buffer is fully drained; empty follow-up returns nothing
let r2 = parser.parse_reasoning_streaming_incremental("", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, "");
}
#[test]
fn test_partial_end_token_stream_reasoning_true() {
// Partial </think> split across chunks with stream_reasoning=true.
// The partial-end-token buffer check only fires when the parser is ALREADY in
// reasoning mode from a prior call. If <think> and </th arrive in the same chunk,
// stream_reasoning=true emits the reasoning content immediately (including </th).
// So <think> must arrive as its own chunk first.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>reasoning", &[]);
assert_eq!(r1.reasoning_text, "reasoning");
assert_eq!(r1.normal_text, "");
// Partial end token while already in reasoning — buffered, nothing emitted
let r2 = parser.parse_reasoning_streaming_incremental("</th", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, "");
// Complete the end token
let r3 = parser.parse_reasoning_streaming_incremental("ink> normal", &[]);
assert_eq!(r3.reasoning_text, "");
assert_eq!(r3.normal_text, " normal");
}
#[test]
fn test_empty_string_input_various_states() {
// Empty string input should always return empty results without changing state
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
// State: idle
let r1 = parser.parse_reasoning_streaming_incremental("", &[]);
assert_eq!(r1.reasoning_text, "");
assert_eq!(r1.normal_text, "");
// Enter reasoning
parser.parse_reasoning_streaming_incremental("<think>content", &[]);
// State: in reasoning
let r2 = parser.parse_reasoning_streaming_incremental("", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, "");
// Complete and exit reasoning
parser.parse_reasoning_streaming_incremental("</think>", &[]);
// State: post-reasoning (normal text)
let r3 = parser.parse_reasoning_streaming_incremental("", &[]);
assert_eq!(r3.reasoning_text, "");
assert_eq!(r3.normal_text, "");
}
#[test]
fn test_force_reasoning_stream_false_multiple_blocks() {
// force_reasoning=true (deepseek_r1 mode), stream_reasoning=false.
// First block uses forced-reasoning (no explicit <think>); subsequent blocks
// use explicit tags.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), true, false);
// Forced reasoning without open tag, flushed on </think>
let r1 =
parser.parse_reasoning_streaming_incremental("initial reasoning</think> normal1 ", &[]);
assert_eq!(r1.reasoning_text, "initial reasoning");
assert_eq!(r1.normal_text, " normal1 ");
// Subsequent explicit <think> block works correctly
let r2 = parser
.parse_reasoning_streaming_incremental("<think>second block</think> normal2", &[]);
assert_eq!(r2.reasoning_text, "second block");
assert_eq!(r2.normal_text, " normal2");
}
#[test]
fn test_glm5_pattern_a_burst_single_chunk() {
// GLM-5 Pattern A: the entire completion arrives in one SSE event.
// Format: <think>T1</think><tool_call>A</tool_call><think>T2</think><tool_call>B</tool_call>
//
// Both reasoning blocks must be extracted into reasoning_text; both tool calls
// must land in normal_text for the downstream tool call parser. No follow-up
// call should be needed — the loop fully drains the buffer in a single call.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental(
"<think>T1</think><tool_call>A</tool_call><think>T2</think><tool_call>B</tool_call>",
&[],
);
assert_eq!(r1.reasoning_text, "T1T2");
assert_eq!(
r1.normal_text,
"<tool_call>A</tool_call><tool_call>B</tool_call>"
);
// Buffer is fully drained; stream can end here with no content loss
let r2 = parser.parse_reasoning_streaming_incremental("", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, "");
}
#[test]
fn test_tool_call_xml_between_reasoning_blocks_streaming() {
// GLM-5 Pattern A chunk-by-chunk: verifies that tool call XML between reasoning
// blocks lands in normal_text, not reasoning_text, across separate SSE events.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>T1</think>", &[]);
assert_eq!(r1.reasoning_text, "T1");
assert_eq!(r1.normal_text, "");
let r2 = parser.parse_reasoning_streaming_incremental("<tool_call>A</tool_call>", &[]);
assert_eq!(r2.normal_text, "<tool_call>A</tool_call>");
assert_eq!(r2.reasoning_text, "");
let r3 = parser.parse_reasoning_streaming_incremental("<think>T2</think>", &[]);
assert_eq!(r3.reasoning_text, "T2");
assert_eq!(r3.normal_text, "");
let r4 = parser.parse_reasoning_streaming_incremental("<tool_call>B</tool_call>", &[]);
assert_eq!(r4.normal_text, "<tool_call>B</tool_call>");
assert_eq!(r4.reasoning_text, "");
}
// =========================================================================
// Mid-string partial tag tests (overlap-based buffering)
//
// These test scenarios where a <think> or </think> tag is split mid-string
// (not at the start of the buffer). Backends that batch multiple forward-pass
// tokens into a single chunked response can produce these patterns.
//
// Ported from PR #6448 (ryanolson) with additional fakeout tests.
// =========================================================================
#[test]
fn test_mid_string_partial_opening_tag_batched() {
// Backend batches tokens: "Hello world <th" arrives as one chunk
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("Hello world <th", &[]);
// "Hello world " emitted as normal, "<th" held in buffer
assert_eq!(r1.normal_text, "Hello world ");
assert_eq!(r1.reasoning_text, "");
let r2 = parser
.parse_reasoning_streaming_incremental("ink>reasoning content</think> answer", &[]);
assert_eq!(r2.reasoning_text, "reasoning content");
assert_eq!(r2.normal_text, " answer");
}
#[test]
fn test_batched_tag_boundary_split() {
// Aggressive batching: <think> tag split with normal text prefix
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("The answer is <thi", &[]);
assert_eq!(r1.normal_text, "The answer is ");
assert_eq!(r1.reasoning_text, "");
let r2 = parser.parse_reasoning_streaming_incremental("nk>let me think</think>42", &[]);
assert_eq!(r2.reasoning_text, "let me think");
assert_eq!(r2.normal_text, "42");
}
#[test]
fn test_mid_string_partial_closing_tag_stream_reasoning_false() {
// With stream_reasoning=false, content stays buffered until </think>.
// Partial </think> split mid-string while in reasoning mode.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, false);
let r1 =
parser.parse_reasoning_streaming_incremental("<think>reasoning content and </th", &[]);
assert_eq!(r1.normal_text, "");
assert_eq!(r1.reasoning_text, "");
let r2 = parser.parse_reasoning_streaming_incremental("ink> normal text", &[]);
assert_eq!(r2.reasoning_text, "reasoning content and ");
assert_eq!(r2.normal_text, " normal text");
}
#[test]
fn test_mid_string_partial_closing_tag_stream_reasoning_true() {
// With stream_reasoning=true, reasoning content is emitted incrementally.
// The partial "</th" at the end must NOT be emitted as reasoning text.
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 =
parser.parse_reasoning_streaming_incremental("<think>reasoning content and </th", &[]);
// "reasoning content and " emitted as reasoning, "</th" held
assert_eq!(r1.reasoning_text, "reasoning content and ");
assert_eq!(r1.normal_text, "");
let r2 = parser.parse_reasoning_streaming_incremental("ink> normal text", &[]);
assert_eq!(r2.reasoning_text, "");
assert_eq!(r2.normal_text, " normal text");
}
#[test]
fn test_batched_interleaved_with_mid_string_partial() {
// First block complete in chunk 1, second block's <think> split at boundary
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 =
parser.parse_reasoning_streaming_incremental("<think>thought1</think>answer1<thi", &[]);
assert_eq!(r1.reasoning_text, "thought1");
assert_eq!(r1.normal_text, "answer1");
let r2 = parser.parse_reasoning_streaming_incremental("nk>thought2</think>answer2", &[]);
assert_eq!(r2.reasoning_text, "thought2");
assert_eq!(r2.normal_text, "answer2");
}
#[test]
fn test_partial_tag_false_positive() {
// "<th" looks like partial <think> but "thesis" is not <think>
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("value <thesis on", &[]);
// No suffix of "value <thesis on" is a prefix of "<think>" — all emitted
let r2 = parser.parse_reasoning_streaming_incremental(" AI> is great", &[]);
let combined_normal = format!("{}{}", r1.normal_text, r2.normal_text);
assert_eq!(combined_normal, "value <thesis on AI> is great");
assert_eq!(r1.reasoning_text, "");
assert_eq!(r2.reasoning_text, "");
}
#[test]
fn test_partial_closing_tag_fakeout() {
// Ollama-style fakeout: "</th" buffered, but "ing>" completes "</thing>" not "</think>"
let mut parser =
BasicReasoningParser::new("<think>".to_string(), "</think>".to_string(), false, true);
let r1 = parser.parse_reasoning_streaming_incremental("<think>abc</th", &[]);
assert_eq!(r1.reasoning_text, "abc");
assert_eq!(r1.normal_text, "");
// "ing>def" completes the partial as "</thing>def" — not a closing tag
let r2 = parser.parse_reasoning_streaming_incremental("ing>def", &[]);
assert_eq!(r2.reasoning_text, "</thing>def");
assert_eq!(r2.normal_text, "");
// Real closing tag arrives
let r3 = parser.parse_reasoning_streaming_incremental("</think>done", &[]);
assert_eq!(r3.reasoning_text, "");
assert_eq!(r3.normal_text, "done");
}
#[test]
fn test_overlap_helper_function() {
// Direct tests for the overlap utility
assert_eq!(overlap("abc</th", "</think>"), 4);
assert_eq!(overlap("abc</thing>def", "</think>"), 0);
assert_eq!(overlap("<", "<think>"), 1);
assert_eq!(overlap("<th", "<think>"), 3);
assert_eq!(overlap("<think>", "<think>"), 7); // full match
assert_eq!(overlap("no match", "<think>"), 0);
assert_eq!(overlap("", "<think>"), 0);
assert_eq!(overlap("Hello world <thi", "<think>"), 4);
// Multi-byte delimiters (Kimi parser uses ◁think▷ / ◁/think▷)
assert_eq!(overlap("text◁", "◁think▷"), 3); // ◁ is 3 bytes
assert_eq!(overlap("text◁th", "◁think▷"), 5);
assert_eq!(overlap("text◁/thi", "◁/think▷"), 7);
assert_eq!(overlap("no match", "◁think▷"), 0);
}
} }
lib/parsers/src/reasoning/mod.rs
View file @ d82b0050
...@@ -30,6 +30,7 @@ fn get_reasoning_parser_map() -> &'static HashMap<&'static str, ReasoningParserT ...@@ -30,6 +30,7 @@ fn get_reasoning_parser_map() -> &'static HashMap<&'static str, ReasoningParserT
map.insert("mistral", ReasoningParserType::Mistral); map.insert("mistral", ReasoningParserType::Mistral);
map.insert("granite", ReasoningParserType::Granite); map.insert("granite", ReasoningParserType::Granite);
map.insert("nemotron_nano", ReasoningParserType::NemotronDeci); // nemotron nano is <think>...</think> map.insert("nemotron_nano", ReasoningParserType::NemotronDeci); // nemotron nano is <think>...</think>
map.insert("glm45", ReasoningParserType::NemotronDeci); // GLM-4.5/5 is <think>...</think>, no force_reasoning
map.insert( map.insert(
"minimax_append_think", "minimax_append_think",
ReasoningParserType::MiniMaxAppendThink, ReasoningParserType::MiniMaxAppendThink,
...@@ -225,6 +226,7 @@ mod tests { ...@@ -225,6 +226,7 @@ mod tests {
"mistral", "mistral",
"granite", "granite",
"nemotron_nano", "nemotron_nano",
"glm45",
"minimax_append_think", "minimax_append_think",
]; ];
for parser in available_parsers { for parser in available_parsers {
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment