Compare commits

...

1 Commits

Author SHA1 Message Date
jmorganca
bb9c1b2561 parsers: refactor Nemotron parser to reuse Qwen3Coder for tool calls
Simplify Nemotron3NanoParser by delegating tool call parsing to
Qwen3CoderParser instead of duplicating the parsing logic. The
Nemotron parser now only handles the thinking state machine and
transitions to Qwen3CoderParser for content and tool call parsing.

This also fixes an issue where tool calls without </think> would
cause the parser to get stuck in thinking mode.
2026-01-17 12:36:23 -08:00
3 changed files with 170 additions and 453 deletions

View File

@@ -1,7 +1,6 @@
package parsers
import (
"regexp"
"strings"
"unicode"
@@ -14,243 +13,114 @@ const (
Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
Nemotron3NanoSkipWhitespaceAfterThinking
Nemotron3NanoCollectingContent
Nemotron3NanoCollectingToolCalls
)
const (
nemotronThinkClose = "</think>"
nemotronToolCallOpen = "<tool_call>"
nemotronToolCallClose = "</tool_call>"
nemotronThinkClose = "</think>"
nemotronToolCallOpen = "<tool_call>"
)
type Nemotron3NanoParser struct {
state Nemotron3NanoParserState
buffer strings.Builder
tools []api.Tool
state Nemotron3NanoParserState
buffer strings.Builder
toolParser *Qwen3CoderParser
}
func (p *Nemotron3NanoParser) HasToolSupport() bool { return true }
func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return true }
func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
p.tools = tools
p.toolParser = &Qwen3CoderParser{}
p.toolParser.Init(tools, nil, nil)
// thinking is enabled if user requests it
thinkingEnabled := thinkValue != nil && thinkValue.Bool()
prefill := lastMessage != nil && lastMessage.Role == "assistant"
if !thinkingEnabled {
if !thinkingEnabled || (prefill && lastMessage.Content != "") {
p.state = Nemotron3NanoCollectingContent
return tools
} else {
p.state = Nemotron3NanoCollectingThinking
}
if prefill && lastMessage.Content != "" {
p.state = Nemotron3NanoCollectingContent
return tools
}
p.state = Nemotron3NanoCollectingThinking
return tools
}
type nemotronEvent interface {
isNemotronEvent()
}
type nemotronEventThinkingContent struct {
content string
}
type nemotronEventContent struct {
content string
}
type nemotronEventToolCall struct {
toolCall api.ToolCall
}
func (nemotronEventThinkingContent) isNemotronEvent() {}
func (nemotronEventContent) isNemotronEvent() {}
func (nemotronEventToolCall) isNemotronEvent() {}
func (p *Nemotron3NanoParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
p.buffer.WriteString(s)
events := p.parseEvents()
var toolCalls []api.ToolCall
var contentSb strings.Builder
var thinkingSb strings.Builder
for _, event := range events {
switch event := event.(type) {
case nemotronEventToolCall:
toolCalls = append(toolCalls, event.toolCall)
case nemotronEventThinkingContent:
thinkingSb.WriteString(event.content)
case nemotronEventContent:
contentSb.WriteString(event.content)
}
if p.state == Nemotron3NanoCollectingContent {
return p.toolParser.Add(s, done)
}
return contentSb.String(), thinkingSb.String(), toolCalls, nil
}
func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
var all []nemotronEvent
keepLooping := true
for keepLooping {
var events []nemotronEvent
events, keepLooping = p.eat()
if len(events) > 0 {
all = append(all, events...)
}
}
return all
}
// emitWithPartialCheck extracts unambiguous content before a potential partial tag
func (p *Nemotron3NanoParser) emitWithPartialCheck(bufStr, tag string) (unambiguous, ambiguous string) {
if overlapLen := overlap(bufStr, tag); overlapLen > 0 {
beforePartialTag := bufStr[:len(bufStr)-overlapLen]
trailingLen := trailingWhitespaceLen(beforePartialTag)
return bufStr[:len(beforePartialTag)-trailingLen], bufStr[len(beforePartialTag)-trailingLen:]
}
wsLen := trailingWhitespaceLen(bufStr)
return bufStr[:len(bufStr)-wsLen], bufStr[len(bufStr)-wsLen:]
}
func (p *Nemotron3NanoParser) eat() ([]nemotronEvent, bool) {
bufStr := p.buffer.String()
if bufStr == "" {
return nil, false
}
switch p.state {
case Nemotron3NanoCollectingThinking:
if strings.Contains(bufStr, nemotronThinkClose) {
split := strings.SplitN(bufStr, nemotronThinkClose, 2)
thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
p.buffer.Reset()
remainder := strings.TrimLeftFunc(split[1], unicode.IsSpace)
p.buffer.WriteString(remainder)
// Transition to whitespace-skipping state if buffer is empty,
// otherwise go directly to content collection
if remainder == "" {
p.state = Nemotron3NanoSkipWhitespaceAfterThinking
} else {
p.state = Nemotron3NanoCollectingContent
}
if thinking != "" {
return []nemotronEvent{nemotronEventThinkingContent{content: thinking}}, true
}
return nil, true
}
unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronThinkClose)
p.buffer.Reset()
p.buffer.WriteString(ambig)
if unambig != "" {
return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
}
return nil, false
// We only want to skip whitespace between thinking and content
case Nemotron3NanoSkipWhitespaceAfterThinking:
bufStr = strings.TrimLeftFunc(bufStr, unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(bufStr)
if bufStr == "" {
return nil, false
if p.state == Nemotron3NanoSkipWhitespaceAfterThinking {
s = strings.TrimLeftFunc(s, unicode.IsSpace)
if s == "" {
return "", "", nil, nil
}
p.state = Nemotron3NanoCollectingContent
return nil, true
return p.toolParser.Add(s, done)
}
case Nemotron3NanoCollectingContent:
if strings.Contains(bufStr, nemotronToolCallOpen) {
split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
content := strings.TrimRightFunc(split[0], unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(split[1])
p.state = Nemotron3NanoCollectingToolCalls
if content != "" {
return []nemotronEvent{nemotronEventContent{content: content}}, true
}
return nil, true
}
unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
// Nemotron3NanoCollectingThinking - buffer and look for end markers
p.buffer.WriteString(s)
bufStr := p.buffer.String()
// Look for end of thinking: </think> or <tool_call> (model may skip </think>)
thinkIdx := strings.Index(bufStr, nemotronThinkClose)
toolIdx := strings.Index(bufStr, nemotronToolCallOpen)
var endIdx int = -1
var remainder string
if thinkIdx != -1 && (toolIdx == -1 || thinkIdx < toolIdx) {
endIdx = thinkIdx
remainder = strings.TrimLeftFunc(bufStr[thinkIdx+len(nemotronThinkClose):], unicode.IsSpace)
} else if toolIdx != -1 {
endIdx = toolIdx
remainder = bufStr[toolIdx:] // Include <tool_call> tag
}
if endIdx != -1 {
thinking = strings.TrimRightFunc(bufStr[:endIdx], unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(ambig)
if unambig != "" {
return []nemotronEvent{nemotronEventContent{content: unambig}}, false
if remainder == "" {
p.state = Nemotron3NanoSkipWhitespaceAfterThinking
} else {
p.state = Nemotron3NanoCollectingContent
content, _, calls, err = p.toolParser.Add(remainder, done)
}
return nil, false
case Nemotron3NanoCollectingToolCalls:
if strings.Contains(bufStr, nemotronToolCallClose) {
split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(remaining)
var events []nemotronEvent
if tc, err := p.parseToolCall(split[0]); err == nil {
events = append(events, nemotronEventToolCall{toolCall: tc})
}
if !strings.Contains(remaining, nemotronToolCallOpen) {
p.state = Nemotron3NanoCollectingContent
}
return events, true
}
return nil, false
return content, thinking, calls, err
}
return nil, false
// No end marker - emit unambiguous thinking
thinking = p.emitThinking(bufStr)
return "", thinking, nil, nil
}
var (
nemotronFunctionRegex = regexp.MustCompile(`<function=([^>]+)>`)
nemotronParameterRegex = regexp.MustCompile(`<parameter=([^>]+)>\n?([\s\S]*?)\n?</parameter>`)
)
// emitThinking returns unambiguous thinking content, keeping potential partial tags in buffer
func (p *Nemotron3NanoParser) emitThinking(bufStr string) string {
// Check for partial </think> or <tool_call> at end
thinkOverlap := overlap(bufStr, nemotronThinkClose)
toolOverlap := overlap(bufStr, nemotronToolCallOpen)
maxOverlap := max(thinkOverlap, toolOverlap)
func (p *Nemotron3NanoParser) parseToolCall(content string) (api.ToolCall, error) {
toolCall := api.ToolCall{}
// Extract function name
fnMatch := nemotronFunctionRegex.FindStringSubmatch(content)
if len(fnMatch) < 2 {
return toolCall, nil
}
toolCall.Function.Name = fnMatch[1]
// Extract parameters
toolCall.Function.Arguments = api.NewToolCallFunctionArguments()
paramMatches := nemotronParameterRegex.FindAllStringSubmatch(content, -1)
for _, match := range paramMatches {
if len(match) >= 3 {
paramName := match[1]
paramValue := strings.TrimSpace(match[2])
// Try to parse as typed value based on tool definition
toolCall.Function.Arguments.Set(paramName, p.parseParamValue(paramName, paramValue))
}
if maxOverlap > 0 {
unambiguous := bufStr[:len(bufStr)-maxOverlap]
unambiguous = strings.TrimRightFunc(unambiguous, unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(bufStr[len(bufStr)-maxOverlap:])
return unambiguous
}
return toolCall, nil
}
func (p *Nemotron3NanoParser) parseParamValue(paramName string, raw string) any {
// Find the matching tool to get parameter type
var paramType api.PropertyType
for _, tool := range p.tools {
if tool.Function.Parameters.Properties != nil {
if prop, ok := tool.Function.Parameters.Properties.Get(paramName); ok {
paramType = prop.Type
break
}
}
}
return parseValue(raw, paramType)
// No partial tags - emit all but trailing whitespace
wsLen := trailingWhitespaceLen(bufStr)
if wsLen > 0 {
unambiguous := bufStr[:len(bufStr)-wsLen]
p.buffer.Reset()
p.buffer.WriteString(bufStr[len(bufStr)-wsLen:])
return unambiguous
}
// Nothing to hold back
p.buffer.Reset()
return bufStr
}

View File

@@ -8,6 +8,8 @@ import (
"github.com/ollama/ollama/api"
)
// TestNemotron3NanoParser tests Nemotron-specific behavior (thinking support).
// Tool call parsing is tested in qwen3coder_test.go since Nemotron delegates to Qwen3CoderParser.
func TestNemotron3NanoParser(t *testing.T) {
tests := []struct {
name string
@@ -17,18 +19,6 @@ func TestNemotron3NanoParser(t *testing.T) {
expectedThinking string
expectedCalls []api.ToolCall
}{
{
name: "simple content - no thinking",
input: "Hello, how can I help you?",
thinkValue: nil,
expectedContent: "Hello, how can I help you?",
},
{
name: "simple content - thinking disabled",
input: "Hello, how can I help you?",
thinkValue: &api.ThinkValue{Value: false},
expectedContent: "Hello, how can I help you?",
},
{
name: "thinking then content",
input: "Let me think about this...</think>\nHere is my answer.",
@@ -43,69 +33,6 @@ func TestNemotron3NanoParser(t *testing.T) {
expectedThinking: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude",
expectedContent: "The answer is 42.",
},
{
name: "simple tool call",
input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "Paris"}),
},
},
},
},
{
name: "content then tool call",
input: "Let me check the weather.\n<tool_call>\n<function=get_weather>\n<parameter=city>\nNYC\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedContent: "Let me check the weather.",
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "NYC"}),
},
},
},
},
{
name: "tool call with multiple parameters",
input: "<tool_call>\n<function=book_flight>\n<parameter=from>\nSFO\n</parameter>\n<parameter=to>\nNYC\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "book_flight",
Arguments: testArgs(map[string]any{
"from": "SFO",
"to": "NYC",
}),
},
},
},
},
{
name: "multiple tool calls",
input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nSan Francisco\n</parameter>\n</function>\n</tool_call>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nNew York\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "San Francisco"}),
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "New York"}),
},
},
},
},
{
name: "thinking then tool call",
input: "I should check the weather...</think>\n<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
@@ -135,19 +62,6 @@ func TestNemotron3NanoParser(t *testing.T) {
},
},
},
{
name: "tool call with multiline parameter value",
input: "<tool_call>\n<function=create_note>\n<parameter=content>\nLine 1\nLine 2\nLine 3\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "create_note",
Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
},
},
},
},
{
name: "empty thinking block - immediate close",
input: "</think>\nHere is my answer.",
@@ -161,18 +75,6 @@ func TestNemotron3NanoParser(t *testing.T) {
thinkValue: &api.ThinkValue{Value: false},
expectedContent: "</think>\nSome content after spurious tag.",
},
{
name: "tool call with no function name - returns empty tool call",
input: "<tool_call>\n<function=>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "", Arguments: api.NewToolCallFunctionArguments()}}},
},
{
name: "content with newlines preserved",
input: "Line 1\n\nLine 2\n\n\nLine 3",
thinkValue: nil,
expectedContent: "Line 1\n\nLine 2\n\n\nLine 3",
},
{
name: "thinking with only whitespace after close tag",
input: "My thoughts...</think> \n\t\n Content here.",
@@ -180,25 +82,6 @@ func TestNemotron3NanoParser(t *testing.T) {
expectedThinking: "My thoughts...",
expectedContent: "Content here.",
},
{
name: "unicode content",
input: "Hello 世界! 🌍 Ñoño",
thinkValue: nil,
expectedContent: "Hello 世界! 🌍 Ñoño",
},
{
name: "tool call with numeric parameter",
input: "<tool_call>\n<function=set_temp>\n<parameter=value>\n42\n</parameter>\n</function>\n</tool_call>",
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "set_temp",
Arguments: testArgs(map[string]any{"value": "42"}),
},
},
},
},
}
for _, tt := range tests {
@@ -233,6 +116,8 @@ func TestNemotron3NanoParser(t *testing.T) {
}
}
// TestNemotron3NanoParser_Streaming tests streaming behavior for thinking support.
// Tool call streaming is tested in qwen3coder_test.go.
func TestNemotron3NanoParser_Streaming(t *testing.T) {
tests := []struct {
name string
@@ -242,18 +127,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
expectedThinking string
expectedCalls []api.ToolCall
}{
{
name: "streaming content character by character",
chunks: []string{"H", "e", "l", "l", "o", ",", " ", "w", "o", "r", "l", "d", "!"},
thinkValue: nil,
expectedContent: "Hello, world!",
},
{
name: "streaming content small tokens",
chunks: []string{"Hel", "lo", ", ", "how ", "can", " I", " help", " you", " today", "?"},
thinkValue: nil,
expectedContent: "Hello, how can I help you today?",
},
{
name: "streaming thinking then content - granular",
chunks: []string{"Let", " me", " th", "ink", " about", " this", "...", "<", "/", "think", ">", "\n", "Here", " is", " my", " answer", "."},
@@ -268,45 +141,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
expectedThinking: "Step 1: Analyze\nStep 2: Process",
expectedContent: "The answer.",
},
{
name: "streaming tool call - highly granular",
chunks: []string{"<", "tool", "_", "call", ">", "\n", "<", "func", "tion", "=", "get", "_", "weather", ">", "\n", "<", "param", "eter", "=", "city", ">", "\n", "Par", "is", "\n", "</", "param", "eter", ">", "\n", "</", "func", "tion", ">", "\n", "</", "tool", "_", "call", ">"},
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "Paris"}),
},
},
},
},
{
name: "streaming content then tool call - granular",
chunks: []string{"Let", " me", " check", " the", " weather", ".", "\n<", "tool_call", ">", "\n", "<function=", "get_weather", ">", "\n", "<parameter=", "city", ">", "\n", "NYC", "\n", "</parameter>", "\n", "</function>", "\n", "</tool_call>"},
thinkValue: nil,
expectedContent: "Let me check the weather.",
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "NYC"}),
},
},
},
},
{
name: "tool call tag split character by character",
chunks: []string{"<", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">", "\n", "<", "f", "u", "n", "c", "t", "i", "o", "n", "=", "t", "e", "s", "t", ">", "\n", "<", "/", "f", "u", "n", "c", "t", "i", "o", "n", ">", "\n", "<", "/", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">"},
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "test",
Arguments: api.NewToolCallFunctionArguments(),
},
},
},
},
{
name: "thinking close tag split character by character",
chunks: []string{"I", "'", "m", " ", "t", "h", "i", "n", "k", "i", "n", "g", ".", ".", ".", "<", "/", "t", "h", "i", "n", "k", ">", "\n", "D", "o", "n", "e", "!"},
@@ -321,22 +155,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
expectedThinking: "Thinking...",
expectedContent: "Content here.",
},
{
name: "tool call with multiple parameters - streaming",
chunks: []string{"<tool_", "call>\n", "<function", "=book_", "flight>", "\n<para", "meter=", "from>\n", "SFO\n", "</param", "eter>", "\n<param", "eter=to", ">\nNYC", "\n</para", "meter>", "\n</func", "tion>\n", "</tool_", "call>"},
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "book_flight",
Arguments: testArgs(map[string]any{
"from": "SFO",
"to": "NYC",
}),
},
},
},
},
{
name: "thinking then content then tool call - streaming",
chunks: []string{"Ana", "lyzing", " your", " request", "...", "</", "think", ">\n", "I'll", " check", " that", " for", " you", ".", "\n", "<tool", "_call", ">\n", "<function", "=search", ">\n", "<parameter", "=query", ">\n", "test", " query", "\n</", "parameter", ">\n", "</function", ">\n", "</tool", "_call", ">"},
@@ -352,45 +170,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
},
},
},
{
name: "multiple tool calls - streaming",
chunks: []string{
"<tool_call>", "\n", "<function=", "get_weather>", "\n",
"<parameter=", "city>\n", "San Fran", "cisco\n", "</parameter>", "\n",
"</function>", "\n", "</tool_call>", "\n",
"<tool_", "call>\n", "<function", "=get_weather", ">\n",
"<param", "eter=city", ">\nNew", " York\n", "</parameter>\n",
"</function>\n", "</tool_call>",
},
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "San Francisco"}),
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "New York"}),
},
},
},
},
{
name: "tool call with multiline parameter - streaming",
chunks: []string{"<tool_call>\n", "<function=", "create_note>\n", "<parameter=", "content>\n", "Line 1", "\nLine", " 2\n", "Line 3", "\n</parameter>\n", "</function>\n", "</tool_call>"},
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "create_note",
Arguments: testArgs(map[string]any{"content": "Line 1\nLine 2\nLine 3"}),
},
},
},
},
{
name: "empty thinking block",
chunks: []string{"</think>", "\n", "Just content."},
@@ -398,12 +177,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
expectedThinking: "",
expectedContent: "Just content.",
},
{
name: "empty input chunks interspersed",
chunks: []string{"Hello", "", " ", "", "world", "", "!"},
thinkValue: nil,
expectedContent: "Hello world!",
},
{
name: "tool call immediately after think close - no content",
chunks: []string{"Analyzing...", "</think>", "\n", "<tool_call>", "\n<function=test>\n</function>\n", "</tool_call>"},
@@ -418,25 +191,6 @@ func TestNemotron3NanoParser_Streaming(t *testing.T) {
},
},
},
{
name: "tool call with empty parameter value",
chunks: []string{"<tool_call>\n<function=test>\n<parameter=name>\n", "\n</parameter>\n</function>\n</tool_call>"},
thinkValue: nil,
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "test",
Arguments: testArgs(map[string]any{"name": ""}),
},
},
},
},
{
name: "partial tool call tag at end - buffered",
chunks: []string{"Here's some content", "<tool"},
thinkValue: nil,
expectedContent: "Here's some content",
},
}
for _, tt := range tests {
@@ -572,3 +326,65 @@ func TestNemotron3NanoParser_WithTools(t *testing.T) {
t.Errorf("calls mismatch (-got +want):\n%s", diff)
}
}
// TestNemotron3NanoParser_ToolCallWithoutThinkClose tests the case where thinking is enabled
// but the model outputs content + tool call WITHOUT the </think> tag.
// The parser should still parse the tool call (content before is treated as thinking).
func TestNemotron3NanoParser_ToolCallWithoutThinkClose(t *testing.T) {
chunks := []string{
"Let", " me", " analyze", " this", ".", "\n",
"<tool_call>", "\n",
"<function=get_weather>", "\n",
"<parameter=city>", "Paris", "</parameter>", "\n",
"</function>", "\n",
"</tool_call>",
}
p := &Nemotron3NanoParser{}
p.Init(nil, nil, &api.ThinkValue{Value: true}) // thinking ENABLED but model doesn't output </think>
var allContent string
var allThinking string
var allCalls []api.ToolCall
for _, chunk := range chunks {
content, thinking, calls, err := p.Add(chunk, false)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
allContent += content
allThinking += thinking
allCalls = append(allCalls, calls...)
}
// Drain
content, thinking, calls, err := p.Add("", true)
if err != nil {
t.Fatalf("unexpected error on done: %v", err)
}
allContent += content
allThinking += thinking
allCalls = append(allCalls, calls...)
// The parser was in thinking mode, so text before <tool_call> is emitted as thinking.
expectedThinking := "Let me analyze this."
expectedCalls := []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: testArgs(map[string]any{"city": "Paris"}),
},
},
}
if allContent != "" {
t.Errorf("expected no content (text was streamed as thinking), got: %q", allContent)
}
if diff := cmp.Diff(allThinking, expectedThinking); diff != "" {
t.Errorf("thinking mismatch (-got +want):\n%s", diff)
}
if diff := cmp.Diff(allCalls, expectedCalls, argsComparer); diff != "" {
t.Errorf("calls mismatch (-got +want):\n%s", diff)
}
}

View File

@@ -91,6 +91,37 @@ func TestQwenParserStreaming(t *testing.T) {
},
},
},
{
desc: "tool call tags split character by character",
steps: []step{
{input: "<", wantEvents: []qwenEvent{}},
{input: "t", wantEvents: []qwenEvent{}},
{input: "o", wantEvents: []qwenEvent{}},
{input: "o", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: "_", wantEvents: []qwenEvent{}},
{input: "c", wantEvents: []qwenEvent{}},
{input: "a", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: ">", wantEvents: []qwenEvent{}},
{input: "a", wantEvents: []qwenEvent{}},
{input: "b", wantEvents: []qwenEvent{}},
{input: "c", wantEvents: []qwenEvent{}},
{input: "<", wantEvents: []qwenEvent{}},
{input: "/", wantEvents: []qwenEvent{}},
{input: "t", wantEvents: []qwenEvent{}},
{input: "o", wantEvents: []qwenEvent{}},
{input: "o", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: "_", wantEvents: []qwenEvent{}},
{input: "c", wantEvents: []qwenEvent{}},
{input: "a", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: "l", wantEvents: []qwenEvent{}},
{input: ">", wantEvents: []qwenEvent{qwenEventRawToolCall{raw: "abc"}}},
},
},
{
desc: "trailing whitespace between content and tool call",
steps: []step{