mod.rs

// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
use std::collections::HashMap;
use std::sync::OnceLock;

mod base_parser;
mod gpt_oss_parser;
mod granite_parser;
mod minimax_append_think_parser;

// Re-export main types and functions for convenience
pub use base_parser::BasicReasoningParser;
pub use gpt_oss_parser::GptOssReasoningParser;
pub use granite_parser::GraniteReasoningParser;
pub use minimax_append_think_parser::MiniMaxAppendThinkParser;

static REASONING_PARSER_MAP: OnceLock<HashMap<&'static str, ReasoningParserType>> = OnceLock::new();

/// Initialize the global reasoning parser map
fn get_reasoning_parser_map() -> &'static HashMap<&'static str, ReasoningParserType> {
    REASONING_PARSER_MAP.get_or_init(|| {
        let mut map = HashMap::new();
        map.insert("deepseek_r1", ReasoningParserType::DeepseekR1);
        map.insert("basic", ReasoningParserType::Basic);
        map.insert("gpt_oss", ReasoningParserType::GptOss);
        map.insert("qwen3", ReasoningParserType::Qwen);
        map.insert("nemotron_deci", ReasoningParserType::NemotronDeci);
        map.insert("kimi", ReasoningParserType::Kimi);
        map.insert("kimi_k25", ReasoningParserType::KimiK25);
        map.insert("step3", ReasoningParserType::Step3);
        map.insert("mistral", ReasoningParserType::Mistral);
        map.insert("granite", ReasoningParserType::Granite);
        map.insert("nemotron_nano", ReasoningParserType::DeepseekR1); // nemotron nano is ...</think>
        map.insert("nemotron3", ReasoningParserType::DeepseekR1);
        map.insert("glm45", ReasoningParserType::NemotronDeci); // GLM-4.5/5 is <think>...</think>, no force_reasoning
        map.insert(
            "minimax_append_think",
            ReasoningParserType::MiniMaxAppendThink,
        );
        map
    })
}

/// Get all available reasoning parser names
pub fn get_available_reasoning_parsers() -> Vec<&'static str> {
    get_reasoning_parser_map().keys().copied().collect()
}

#[derive(Debug, Clone, Default)]
pub struct ParserResult {
    /// The normal text outside of reasoning blocks.
    pub normal_text: String,

    /// The extracted reasoning text from within reasoning blocks.
    pub reasoning_text: String,
}

impl ParserResult {
    pub fn get_some_reasoning(&self) -> Option<String> {
        if self.reasoning_text.is_empty() {
            None
        } else {
            Some(self.reasoning_text.clone())
        }
    }

    pub fn get_some_normal_text(&self) -> Option<String> {
        if self.normal_text.is_empty() {
            None
        } else {
            Some(self.normal_text.clone())
        }
    }
}

pub trait ReasoningParser: Send + std::fmt::Debug {
    /// Parses a standalone, non-streaming input chunk. Implementations may reset or ignore
    /// internal streaming state and should return the split of normal vs reasoning text for
    /// this complete input. Marker tokens must not be included in either output.
    fn detect_and_parse_reasoning(&mut self, text: &str, token_ids: &[u32]) -> ParserResult;

    /// Parses a streaming chunk and updates internal state. The return value should be the
    /// delta: only the newly discovered normal and reasoning text attributable to this chunk
    /// (not the cumulative totals). Marker tokens must not be included in either output.
    fn parse_reasoning_streaming_incremental(
        &mut self,
        text: &str,
        token_ids: &[u32],
    ) -> ParserResult;

    /// Override the parser's initial reasoning state. When called with `true`, the parser
    /// starts in reasoning mode without waiting for the start token in the completion stream.
    /// Use this when the chat template already injected the start token (e.g., `<think>`)
    /// into the prompt, so it won't appear in the model's output.
    fn set_in_reasoning(&mut self, _in_reasoning: bool) {
        // Default no-op for parsers that don't support per-request overrides.
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[non_exhaustive]
pub enum ReasoningParserType {
    DeepseekR1,
    Step3,
    Basic,
    GptOss,
    Qwen,
    NemotronDeci,
    Kimi,
    KimiK25,
    Mistral,
    Granite,
    MiniMaxAppendThink,
}

#[derive(std::fmt::Debug)]
pub struct ReasoningParserWrapper {
    parser: Box<dyn ReasoningParser>,
}

impl ReasoningParser for ReasoningParserWrapper {
    fn detect_and_parse_reasoning(&mut self, text: &str, token_ids: &[u32]) -> ParserResult {
        self.parser.detect_and_parse_reasoning(text, token_ids)
    }

    fn parse_reasoning_streaming_incremental(
        &mut self,
        text: &str,
        token_ids: &[u32],
    ) -> ParserResult {
        self.parser
            .parse_reasoning_streaming_incremental(text, token_ids)
    }

    fn set_in_reasoning(&mut self, in_reasoning: bool) {
        self.parser.set_in_reasoning(in_reasoning)
    }
}

impl ReasoningParserType {
    pub fn get_reasoning_parser(self) -> ReasoningParserWrapper {
        let basic_parser =
            BasicReasoningParser::new("<think>".into(), "</think>".into(), false, true);
        let force_reasoning_basic_parser =
            BasicReasoningParser::new("<think>".into(), "</think>".into(), true, true);
        match self {
            ReasoningParserType::DeepseekR1 => ReasoningParserWrapper {
                parser: Box::new(force_reasoning_basic_parser),
            },
            ReasoningParserType::Step3 => ReasoningParserWrapper {
                parser: Box::new(force_reasoning_basic_parser),
            },
            ReasoningParserType::Basic => ReasoningParserWrapper {
                parser: Box::new(basic_parser),
            },
            ReasoningParserType::Qwen => ReasoningParserWrapper {
                parser: Box::new(basic_parser),
            },
            ReasoningParserType::NemotronDeci => ReasoningParserWrapper {
                parser: Box::new(basic_parser),
            },
            ReasoningParserType::Kimi => ReasoningParserWrapper {
                parser: Box::new(BasicReasoningParser::new(
                    "◁think▷".into(),
                    "◁/think▷".into(),
                    false,
                    true,
                )),
            },
            ReasoningParserType::KimiK25 => ReasoningParserWrapper {
                parser: Box::new(BasicReasoningParser::new(
                    "<think>".into(),
                    "</think>".into(),
                    true,
                    true,
                )),
            },
            ReasoningParserType::Mistral => ReasoningParserWrapper {
                parser: Box::new(BasicReasoningParser::new(
                    "[THINK]".into(),
                    "[/THINK]".into(),
                    true,
                    true,
                )),
            },
            ReasoningParserType::GptOss => match GptOssReasoningParser::new() {
                Ok(parser) => ReasoningParserWrapper {
                    parser: Box::new(parser),
                },
                Err(e) => {
                    tracing::warn!(
                        "GptOssReasoningParser could not be initialized, falling back to Basic Reasoning Parser: {e}"
                    );
                    ReasoningParserWrapper {
                        parser: Box::new(BasicReasoningParser::new(
                            "<think>".into(),
                            "</think>".into(),
                            false,
                            true,
                        )),
                    }
                }
            },
            ReasoningParserType::Granite => ReasoningParserWrapper {
                parser: Box::new(GraniteReasoningParser::new()),
            },
            ReasoningParserType::MiniMaxAppendThink => ReasoningParserWrapper {
                parser: Box::new(MiniMaxAppendThinkParser::new()),
            },
        }
    }

    pub fn get_reasoning_parser_from_name(name: &str) -> ReasoningParserWrapper {
        tracing::debug!("Selected reasoning parser: {}", name);

        let parser_map = get_reasoning_parser_map();
        let normalized_name = name.to_lowercase();

        match parser_map.get(normalized_name.as_str()) {
            Some(parser_type) => parser_type.get_reasoning_parser(),
            None => {
                tracing::warn!(
                    parser_name = name,
                    "Unknown reasoning parser type, falling back to Basic Reasoning Parser",
                );
                Self::Basic.get_reasoning_parser()
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_get_available_reasoning_parsers() {
        let parsers = get_available_reasoning_parsers();
        assert!(!parsers.is_empty());
        // Update this list when adding a new parser
        let available_parsers = [
            "deepseek_r1",
            "basic",
            "gpt_oss",
            "qwen3",
            "nemotron_deci",
            "kimi",
            "kimi_k25",
            "step3",
            "mistral",
            "granite",
            "nemotron_nano",
            "nemotron3",
            "glm45",
            "minimax_append_think",
        ];
        for parser in available_parsers {
            assert!(parsers.contains(&parser));
        }
    }

    #[test]
    fn test_kimi_k25_detect_and_parse() {
        // (description, input, expected_reasoning, expected_normal)
        let cases = [
            (
                "force reasoning: no think tags",
                "no think tags here",
                "no think tags here",
                "",
            ),
            (
                "standard think tags",
                "<think>Let me reason about this.</think>Hello!",
                "Let me reason about this.",
                "Hello!",
            ),
            (
                "empty think block (instant mode)",
                "<think></think>Hello from instant mode!",
                "",
                "Hello from instant mode!",
            ),
            (
                "empty think block with newline",
                "<think>\n</think>Hello from instant mode!",
                "",
                "Hello from instant mode!",
            ),
        ];

        for (desc, input, expected_reasoning, expected_normal) in cases {
            let mut parser = ReasoningParserType::KimiK25.get_reasoning_parser();
            let result = parser.detect_and_parse_reasoning(input, &[]);
            assert_eq!(
                result.reasoning_text, expected_reasoning,
                "FAILED reasoning: {desc}"
            );
            assert_eq!(result.normal_text, expected_normal, "FAILED normal: {desc}");
        }
    }

    #[test]
    fn test_kimi_k25_streaming_force_reasoning() {
        // Streaming: force_reasoning means tokens before <think> are treated as reasoning
        let mut parser = ReasoningParserType::KimiK25.get_reasoning_parser();

        // First chunk: partial think tag — buffered because it's a prefix of "<think>"
        let r1 = parser.parse_reasoning_streaming_incremental("<thi", &[]);
        assert_eq!(r1.reasoning_text, "");
        assert_eq!(r1.normal_text, "");

        // Second chunk: completes the think tag + reasoning content
        let r2 = parser.parse_reasoning_streaming_incremental("nk>reasoning here", &[]);
        assert_eq!(r2.reasoning_text, "reasoning here");
        assert_eq!(r2.normal_text, "");

        // Third chunk: close tag + normal content
        let r3 = parser.parse_reasoning_streaming_incremental("</think>Hello!", &[]);
        assert_eq!(r3.reasoning_text, "");
        assert_eq!(r3.normal_text, "Hello!");
    }

    #[test]
    fn test_kimi_k25_streaming() {
        // (description, tokens, expected_reasoning, expected_content)
        let cases: Vec<(&str, &[&str], &str, &str)> = vec![
            (
                "complete response",
                &[
                    "<think>",
                    "I need to",
                    " think about",
                    " this carefully.",
                    "</think>",
                    "Bonjour",
                    "!",
                ],
                "I need to think about this carefully.",
                "Bonjour!",
            ),
            (
                "empty think (instant mode)",
                &["<think>", "</think>", "Direct answer."],
                "",
                "Direct answer.",
            ),
        ];

        for (desc, tokens, expected_reasoning, expected_content) in cases {
            let mut parser = ReasoningParserType::KimiK25.get_reasoning_parser();
            let mut all_reasoning = String::new();
            let mut all_content = String::new();
            for token in tokens {
                let r = parser.parse_reasoning_streaming_incremental(token, &[]);
                all_reasoning.push_str(&r.reasoning_text);
                all_content.push_str(&r.normal_text);
            }
            assert_eq!(
                all_reasoning, expected_reasoning,
                "FAILED reasoning: {desc}"
            );
            assert_eq!(all_content, expected_content, "FAILED content: {desc}");
        }
    }

    #[test]
    fn test_kimi_k25_parser_lookup_by_name() {
        // Verify the parser can be looked up by name
        let mut parser = ReasoningParserType::get_reasoning_parser_from_name("kimi_k25");
        let result = parser.detect_and_parse_reasoning("<think>thinking</think>answer", &[]);
        assert_eq!(result.reasoning_text, "thinking");
        assert_eq!(result.normal_text, "answer");
    }

    #[test]
    fn test_kimi_vs_kimi_k25_different_tags() {
        // Kimi (original) uses ◁think▷/◁/think▷, KimiK25 uses <think>/</think>
        let mut kimi = ReasoningParserType::Kimi.get_reasoning_parser();
        let mut kimi_k25 = ReasoningParserType::KimiK25.get_reasoning_parser();

        // Kimi original does NOT parse <think> tags
        let r_kimi = kimi.detect_and_parse_reasoning("<think>reasoning</think>answer", &[]);
        assert_eq!(r_kimi.normal_text, "<think>reasoning</think>answer");
        assert_eq!(r_kimi.reasoning_text, "");

        // KimiK25 does parse <think> tags
        let r_k25 = kimi_k25.detect_and_parse_reasoning("<think>reasoning</think>answer", &[]);
        assert_eq!(r_k25.reasoning_text, "reasoning");
        assert_eq!(r_k25.normal_text, "answer");
    }

    // Scenario 1: Normal streaming flow with force_reasoning + set_in_reasoning.
    // Simulates the OpenAI path where the preprocessor detects prompt-injected
    // reasoning and calls set_in_reasoning(true). The parser should correctly
    // transition from reasoning to content when </think> arrives.
    #[test]
    fn test_nemotron_streaming_with_set_in_reasoning() {
        let mut parser = ReasoningParserType::DeepseekR1.get_reasoning_parser();
        parser.set_in_reasoning(true); // OpenAI path calls this

        let tokens = &["Think", "ing about", " this", ".\n\n", "</think>", "Four"];

        let mut all_reasoning = String::new();
        let mut all_content = String::new();
        for token in tokens {
            let r = parser.parse_reasoning_streaming_incremental(token, &[]);
            all_reasoning.push_str(&r.reasoning_text);
            all_content.push_str(&r.normal_text);
        }
        assert_eq!(all_reasoning, "Thinking about this.\n\n");
        assert_eq!(all_content, "Four");
    }

    // Scenario 2: Streaming with force_reasoning but WITHOUT set_in_reasoning.
    // Simulates the Anthropic path bug where thinking_enabled=false and
    // set_in_reasoning is never called. The parser still starts in reasoning
    // mode (force_reasoning=true) but stripped_think_start=false. The </think>
    // boundary must still be detected correctly.
    #[test]
    fn test_nemotron_streaming_force_reasoning_without_set_in_reasoning() {
        // DeepseekR1 has force_reasoning=true but we do NOT call set_in_reasoning
        let mut parser = ReasoningParserType::DeepseekR1.get_reasoning_parser();

        let tokens = &["Think", "ing about", " this", ".\n\n", "</think>", "Four"];

        let mut all_reasoning = String::new();
        let mut all_content = String::new();
        for token in tokens {
            let r = parser.parse_reasoning_streaming_incremental(token, &[]);
            all_reasoning.push_str(&r.reasoning_text);
            all_content.push_str(&r.normal_text);
        }
        assert_eq!(all_reasoning, "Thinking about this.\n\n");
        assert_eq!(all_content, "Four");
    }

    // Scenario 3: Token-by-token </think> split across chunks.
    // The '<' in '</think>' is a prefix of '<think>'. When stripped_think_start
    // is false, the parser's prefix-check could buffer '<' and interfere with
    // </think> detection. This test verifies the boundary is detected even when
    // </think> arrives as individual characters.
    #[test]
    fn test_nemotron_streaming_split_end_think_tokens() {
        let mut parser = ReasoningParserType::DeepseekR1.get_reasoning_parser();
        parser.set_in_reasoning(true);

        // Simulate token-by-token arrival including </think> split across chunks
        let tokens = &[
            "reason", "ing", " done", ".", "</", "think", ">", "Hello", " world",
        ];

        let mut all_reasoning = String::new();
        let mut all_content = String::new();
        for token in tokens {
            let r = parser.parse_reasoning_streaming_incremental(token, &[]);
            all_reasoning.push_str(&r.reasoning_text);
            all_content.push_str(&r.normal_text);
        }
        assert_eq!(all_reasoning, "reasoning done.");
        assert_eq!(all_content, "Hello world");
    }
}