simplify renderer and parser, add tests

This commit is contained in:
ParthSareen 2025-12-15 17:10:42 -08:00
parent 49d415eb2f
commit 383b028911
5 changed files with 1025 additions and 190 deletions

View File

@ -12,6 +12,7 @@ type Nemotron3NanoParserState int
const (
Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
Nemotron3NanoSkipWhitespaceAfterThinking
Nemotron3NanoCollectingContent
Nemotron3NanoCollectingToolCalls
)
@ -26,16 +27,11 @@ type Nemotron3NanoParser struct {
state Nemotron3NanoParserState
buffer strings.Builder
tools []api.Tool
hasThinkingSupport bool
HasThinking bool
}
func (p *Nemotron3NanoParser) HasToolSupport() bool {
return true
}
func (p *Nemotron3NanoParser) HasThinkingSupport() bool {
return p.hasThinkingSupport
}
func (p *Nemotron3NanoParser) HasToolSupport() bool { return true }
func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return p.HasThinking }
func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
p.tools = tools
@ -115,133 +111,103 @@ func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
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) {
var events []nemotronEvent
bufStr := p.buffer.String()
if bufStr == "" {
return events, false
return nil, false
}
switch p.state {
case Nemotron3NanoCollectingThinking:
if strings.Contains(bufStr, nemotronThinkClose) {
split := strings.SplitN(bufStr, nemotronThinkClose, 2)
thinking := split[0]
thinking = strings.TrimRightFunc(thinking, unicode.IsSpace)
remaining := split[1]
remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(remaining)
p.state = Nemotron3NanoCollectingContent
if len(thinking) > 0 {
events = append(events, nemotronEventThinkingContent{content: thinking})
}
return events, true
} else if overlapLen := overlap(bufStr, nemotronThinkClose); overlapLen > 0 {
beforePartialTag := bufStr[:len(bufStr)-overlapLen]
trailingLen := trailingWhitespaceLen(beforePartialTag)
ambiguousStart := len(beforePartialTag) - trailingLen
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, nemotronEventThinkingContent{content: unambiguous})
}
return events, false
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 {
whitespaceLen := trailingWhitespaceLen(bufStr)
ambiguousStart := len(bufStr) - whitespaceLen
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
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(ambiguous)
if len(unambiguous) > 0 {
events = append(events, nemotronEventThinkingContent{content: unambiguous})
p.buffer.WriteString(ambig)
if unambig != "" {
return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
}
return events, 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
}
p.state = Nemotron3NanoCollectingContent
return nil, true
case Nemotron3NanoCollectingContent:
switch {
case strings.Contains(bufStr, nemotronToolCallOpen):
if strings.Contains(bufStr, nemotronToolCallOpen) {
split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
remaining := split[1]
content := strings.TrimRightFunc(split[0], unicode.IsSpace)
p.buffer.Reset()
p.buffer.WriteString(remaining)
p.buffer.WriteString(split[1])
p.state = Nemotron3NanoCollectingToolCalls
if len(contentBefore) > 0 {
events = append(events, nemotronEventContent{content: contentBefore})
if content != "" {
return []nemotronEvent{nemotronEventContent{content: content}}, true
}
return events, true
default:
// Check for partial tool call tag
if overlapLen := overlap(bufStr, nemotronToolCallOpen); overlapLen > 0 {
beforePartialTag := bufStr[:len(bufStr)-overlapLen]
trailingLen := trailingWhitespaceLen(beforePartialTag)
ambiguousStart := len(beforePartialTag) - trailingLen
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
return nil, true
}
unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, nemotronEventContent{content: unambiguous})
}
return events, false
}
// Otherwise emit content, withholding trailing whitespace
whitespaceLen := trailingWhitespaceLen(bufStr)
ambiguousStart := len(bufStr) - whitespaceLen
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, nemotronEventContent{content: unambiguous})
}
return events, false
p.buffer.WriteString(ambig)
if unambig != "" {
return []nemotronEvent{nemotronEventContent{content: unambig}}, false
}
return nil, false
case Nemotron3NanoCollectingToolCalls:
// Look for complete tool call: <function=name>...</function>
if strings.Contains(bufStr, nemotronToolCallClose) {
// We have a complete tool call block
split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
toolContent := split[0]
remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
// Parse the tool call
if toolCall, err := p.parseToolCall(toolContent); err == nil {
events = append(events, nemotronEventToolCall{toolCall: toolCall})
}
p.buffer.Reset()
p.buffer.WriteString(remaining)
// Check if there are more tool calls
if strings.Contains(remaining, nemotronToolCallOpen) {
// Stay in tool call state
return events, true
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 events, false
return nil, false
}
return events, false
return nil, false
}
var (

View File

@ -0,0 +1,486 @@
package parsers
import (
"testing"
"github.com/google/go-cmp/cmp"
"github.com/ollama/ollama/api"
)
func TestNemotron3NanoParser(t *testing.T) {
tests := []struct {
name string
input string
thinkValue *api.ThinkValue
expectedContent string
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.",
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "Let me think about this...",
expectedContent: "Here is my answer.",
},
{
name: "thinking with newlines",
input: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude</think>\nThe answer is 42.",
thinkValue: &api.ThinkValue{Value: true},
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: 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: 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: 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: map[string]any{"city": "San Francisco"},
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: 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>",
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "I should check the weather...",
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "Paris"},
},
},
},
},
{
name: "thinking content then tool call",
input: "Let me think...</think>\nI'll check for you.\n<tool_call>\n<function=search>\n<parameter=query>\ntest\n</parameter>\n</function>\n</tool_call>",
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "Let me think...",
expectedContent: "I'll check for you.",
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "search",
Arguments: map[string]any{"query": "test"},
},
},
},
},
{
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: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
},
},
},
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: tt.thinkValue != nil && tt.thinkValue.Bool()}
p.Init(nil, nil, tt.thinkValue)
content, thinking, calls, err := p.Add(tt.input, false)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
// Drain remaining content
finalContent, finalThinking, finalCalls, err := p.Add("", true)
if err != nil {
t.Fatalf("unexpected error on done: %v", err)
}
content += finalContent
thinking += finalThinking
calls = append(calls, finalCalls...)
if diff := cmp.Diff(content, tt.expectedContent); diff != "" {
t.Errorf("content mismatch (-got +want):\n%s", diff)
}
if diff := cmp.Diff(thinking, tt.expectedThinking); diff != "" {
t.Errorf("thinking mismatch (-got +want):\n%s", diff)
}
if diff := cmp.Diff(calls, tt.expectedCalls); diff != "" {
t.Errorf("calls mismatch (-got +want):\n%s", diff)
}
})
}
}
func TestNemotron3NanoParser_Streaming(t *testing.T) {
tests := []struct {
name string
chunks []string
thinkValue *api.ThinkValue
expectedContent string
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", "."},
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "Let me think about this...",
expectedContent: "Here is my answer.",
},
{
name: "streaming thinking with newlines - granular",
chunks: []string{"Step", " 1", ":", " Ana", "lyze\n", "Step", " 2", ":", " Pro", "cess", "</", "thi", "nk>", "\n", "The", " ans", "wer."},
thinkValue: &api.ThinkValue{Value: true},
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: 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: 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: map[string]any{},
},
},
},
},
{
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", "!"},
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "I'm thinking...",
expectedContent: "Done!",
},
{
name: "multiple whitespace after think tag - separate chunks",
chunks: []string{"Thinking...", "</think>", "\n", "\n", " ", "Content here."},
thinkValue: &api.ThinkValue{Value: true},
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: 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", ">"},
thinkValue: &api.ThinkValue{Value: true},
expectedThinking: "Analyzing your request...",
expectedContent: "I'll check that for you.",
expectedCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "search",
Arguments: map[string]any{"query": "test query"},
},
},
},
},
{
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: map[string]any{"city": "San Francisco"},
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: 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: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
},
},
},
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: tt.thinkValue != nil && tt.thinkValue.Bool()}
p.Init(nil, nil, tt.thinkValue)
var allContent string
var allThinking string
var allCalls []api.ToolCall
for _, chunk := range tt.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...)
if diff := cmp.Diff(allContent, tt.expectedContent); diff != "" {
t.Errorf("content mismatch (-got +want):\n%s", diff)
}
if diff := cmp.Diff(allThinking, tt.expectedThinking); diff != "" {
t.Errorf("thinking mismatch (-got +want):\n%s", diff)
}
if diff := cmp.Diff(allCalls, tt.expectedCalls); diff != "" {
t.Errorf("calls mismatch (-got +want):\n%s", diff)
}
})
}
}
func TestNemotron3NanoParser_HasToolSupport(t *testing.T) {
p := &Nemotron3NanoParser{}
if !p.HasToolSupport() {
t.Error("expected HasToolSupport to return true")
}
}
func TestNemotron3NanoParser_HasThinkingSupport(t *testing.T) {
t.Run("with thinking enabled", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: true}
if !p.HasThinkingSupport() {
t.Error("expected HasThinkingSupport to return true")
}
})
t.Run("with thinking disabled", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: false}
if p.HasThinkingSupport() {
t.Error("expected HasThinkingSupport to return false")
}
})
}
func TestNemotron3NanoParser_Init(t *testing.T) {
t.Run("starts in thinking state when enabled", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: true}
p.Init(nil, nil, &api.ThinkValue{Value: true})
if p.state != Nemotron3NanoCollectingThinking {
t.Errorf("expected state Nemotron3NanoCollectingThinking, got %v", p.state)
}
})
t.Run("starts in content state when thinking disabled", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: true}
p.Init(nil, nil, &api.ThinkValue{Value: false})
if p.state != Nemotron3NanoCollectingContent {
t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
}
})
t.Run("starts in content state when nil thinkValue", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: true}
p.Init(nil, nil, nil)
if p.state != Nemotron3NanoCollectingContent {
t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
}
})
t.Run("starts in content state with assistant prefill", func(t *testing.T) {
p := &Nemotron3NanoParser{HasThinking: true}
prefill := &api.Message{Role: "assistant", Content: "Starting..."}
p.Init(nil, prefill, &api.ThinkValue{Value: true})
if p.state != Nemotron3NanoCollectingContent {
t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
}
})
}
func TestNemotron3NanoParser_WithTools(t *testing.T) {
tools := []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}},
},
},
},
},
}
p := &Nemotron3NanoParser{}
returnedTools := p.Init(tools, nil, nil)
if diff := cmp.Diff(returnedTools, tools); diff != "" {
t.Errorf("tools mismatch (-got +want):\n%s", diff)
}
// Parse a tool call
input := "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>"
_, _, calls, err := p.Add(input, true)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
expectedCalls := []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "Paris"},
},
},
}
if diff := cmp.Diff(calls, expectedCalls); diff != "" {
t.Errorf("calls mismatch (-got +want):\n%s", diff)
}
}

View File

@ -63,9 +63,9 @@ func ParserForName(name string) Parser {
case "olmo3-think":
return &Olmo3ThinkParser{}
case "nemotron-3-nano":
return &Nemotron3NanoParser{hasThinkingSupport: false}
return &Nemotron3NanoParser{HasThinking: false}
case "nemotron-3-nano-thinking":
return &Nemotron3NanoParser{hasThinkingSupport: true}
return &Nemotron3NanoParser{HasThinking: true}
default:
return nil
}

View File

@ -18,17 +18,6 @@ func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool,
// thinking is enabled: model must support it AND user must request it
enableThinking := r.IsThinking && (thinkValue != nil && thinkValue.Bool())
// truncate_history_thinking: drop thinking from historical assistant messages
truncateHistoryThinking := true
// Find the last user message index
lastUserIdx := -1
for i, msg := range messages {
if msg.Role == "user" {
lastUserIdx = i
}
}
// Extract system message if present
var systemMessage string
var loopMessages []api.Message
@ -39,16 +28,14 @@ func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool,
loopMessages = messages
}
// Recalculate lastUserIdx for loopMessages
lastUserIdxInLoop := -1
// Find last user message index for thinking truncation
lastUserIdx := -1
for i, msg := range loopMessages {
if msg.Role == "user" {
lastUserIdxInLoop = i
lastUserIdx = i
}
}
_ = lastUserIdx // silence unused variable warning
// Write system message - always include the system block
sb.WriteString("<|im_start|>system\n")
if systemMessage != "" {
sb.WriteString(systemMessage)
@ -65,79 +52,18 @@ func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool,
for i, message := range loopMessages {
switch message.Role {
case "assistant":
// Build content with reasoning handling
var content string
if message.Thinking != "" {
content = "<think>\n" + message.Thinking + "\n</think>\n" + message.Content
} else {
content = message.Content
// Allow downstream logic to handle broken thought, only handle coherent reasoning here
if !strings.Contains(content, "<think>") && !strings.Contains(content, "</think>") {
content = "<think></think>" + content
}
}
// Build content with thinking tags
content := r.buildContent(message)
shouldTruncate := i < lastUserIdx
if len(message.ToolCalls) > 0 {
// Assistant message with tool calls
sb.WriteString("<|im_start|>assistant\n")
includeContent := !(truncateHistoryThinking && i < lastUserIdxInLoop)
if content != "" {
if includeContent {
sb.WriteString(strings.TrimSpace(content) + "\n")
} else {
// Truncate thinking
c := content
if strings.Contains(c, "</think>") {
// Keep only content after the last closing think
parts := strings.Split(c, "</think>")
c = parts[len(parts)-1]
} else if strings.Contains(c, "<think>") {
// If <think> was opened but never closed, drop the trailing think segment
parts := strings.Split(c, "<think>")
c = parts[0]
}
c = "<think></think>" + strings.TrimSpace(c)
if len(c) > len("<think></think>") {
sb.WriteString(c + "\n")
} else {
sb.WriteString("<think></think>")
}
}
} else {
sb.WriteString("<think></think>")
}
// Write tool calls
for _, toolCall := range message.ToolCalls {
sb.WriteString("<tool_call>\n<function=" + toolCall.Function.Name + ">\n")
for argName, argValue := range toolCall.Function.Arguments {
sb.WriteString("<parameter=" + argName + ">\n")
valueStr := r.formatArgValue(argValue)
sb.WriteString(valueStr + "\n</parameter>\n")
}
sb.WriteString("</function>\n</tool_call>\n")
}
sb.WriteString(r.formatContent(content, shouldTruncate, true))
r.writeToolCalls(&sb, message.ToolCalls)
sb.WriteString("<|im_end|>\n")
} else {
// Assistant message without tool calls
if !(truncateHistoryThinking && i < lastUserIdxInLoop) {
sb.WriteString("<|im_start|>assistant\n" + strings.TrimSpace(content) + "<|im_end|>\n")
} else {
// Truncate thinking - keep only content after </think>
c := content
if strings.Contains(c, "<think>") && strings.Contains(c, "</think>") {
parts := strings.Split(c, "</think>")
// Trim the content after </think> before concatenating
c = "<think></think>" + strings.TrimSpace(parts[len(parts)-1])
}
c = strings.TrimSpace(c)
if c != "" {
sb.WriteString("<|im_start|>assistant\n" + c + "<|im_end|>\n")
} else {
sb.WriteString("<|im_start|>assistant\n<|im_end|>\n")
}
}
formatted := r.formatContent(content, shouldTruncate, false)
sb.WriteString("<|im_start|>assistant\n" + formatted + "<|im_end|>\n")
}
case "user", "system":
@ -234,6 +160,59 @@ func (r *Nemotron3NanoRenderer) renderTools(tools []api.Tool) string {
return sb.String()
}
func (r *Nemotron3NanoRenderer) buildContent(message api.Message) string {
if message.Thinking != "" {
return "<think>\n" + message.Thinking + "\n</think>\n" + message.Content
}
content := message.Content
if !strings.Contains(content, "<think>") && !strings.Contains(content, "</think>") {
return "<think></think>" + content
}
return content
}
func (r *Nemotron3NanoRenderer) formatContent(content string, truncate bool, addNewline bool) string {
if content == "" {
return "<think></think>"
}
if !truncate {
if addNewline {
return strings.TrimSpace(content) + "\n"
}
return strings.TrimSpace(content)
}
// Truncate thinking - keep only content after </think>
c := content
if strings.Contains(c, "</think>") {
parts := strings.Split(c, "</think>")
c = parts[len(parts)-1]
} else if strings.Contains(c, "<think>") {
parts := strings.Split(c, "<think>")
c = parts[0]
}
c = "<think></think>" + strings.TrimSpace(c)
if addNewline && len(c) > len("<think></think>") {
return c + "\n"
}
if c == "<think></think>" {
return c
}
return strings.TrimSpace(c)
}
func (r *Nemotron3NanoRenderer) writeToolCalls(sb *strings.Builder, toolCalls []api.ToolCall) {
for _, tc := range toolCalls {
sb.WriteString("<tool_call>\n<function=" + tc.Function.Name + ">\n")
for name, value := range tc.Function.Arguments {
sb.WriteString("<parameter=" + name + ">\n" + r.formatArgValue(value) + "\n</parameter>\n")
}
sb.WriteString("</function>\n</tool_call>\n")
}
}
func (r *Nemotron3NanoRenderer) formatArgValue(value any) string {
switch v := value.(type) {
case map[string]any, []any:

View File

@ -0,0 +1,404 @@
package renderers
import (
"testing"
"github.com/google/go-cmp/cmp"
"github.com/ollama/ollama/api"
)
func TestNemotron3NanoRenderer(t *testing.T) {
tests := []struct {
name string
msgs []api.Message
tools []api.Tool
thinkValue *api.ThinkValue
isThinking bool
expected string
}{
{
name: "basic user message - thinking mode",
msgs: []api.Message{
{Role: "user", Content: "Hello!"},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n<|im_end|>\n" +
"<|im_start|>user\nHello!<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "basic user message - no thinking",
msgs: []api.Message{
{Role: "user", Content: "Hello!"},
},
isThinking: false,
expected: "<|im_start|>system\n<|im_end|>\n" +
"<|im_start|>user\nHello!<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>",
},
{
name: "with system message",
msgs: []api.Message{
{Role: "system", Content: "You are a helpful assistant."},
{Role: "user", Content: "Hello!"},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n" +
"<|im_start|>user\nHello!<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "multi-turn conversation",
msgs: []api.Message{
{Role: "user", Content: "Hi"},
{Role: "assistant", Content: "Hello! How can I help?"},
{Role: "user", Content: "Tell me a joke"},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n<|im_end|>\n" +
"<|im_start|>user\nHi<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>Hello! How can I help?<|im_end|>\n" +
"<|im_start|>user\nTell me a joke<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "with tools",
msgs: []api.Message{
{Role: "user", Content: "What's the weather in Paris?"},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Description: "Get the current weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Required: []string{"city"},
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}, Description: "The city name"},
},
},
},
},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n" +
"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
"<function>\n<name>get_weather</name>\n" +
"<description>Get the current weather</description>\n" +
"<parameters>\n" +
"<parameter>\n<name>city</name>\n<type>string</type>\n<description>The city name</description>\n</parameter>\n" +
"<required>[\"city\"]</required>\n" +
"</parameters>\n</function>\n</tools>\n\n" +
"If 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>\n\n<IMPORTANT>\nReminder:\n" +
"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
"- Required parameters MUST be specified\n" +
"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
"</IMPORTANT><|im_end|>\n" +
"<|im_start|>user\nWhat's the weather in Paris?<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "tool call with response",
msgs: []api.Message{
{Role: "user", Content: "What's the weather in Paris?"},
{
Role: "assistant",
ToolCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "Paris"},
},
},
},
},
{Role: "tool", Content: "Sunny, 72F"},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Description: "Get the current weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Required: []string{"city"},
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}, Description: "The city name"},
},
},
},
},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n" +
"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
"<function>\n<name>get_weather</name>\n" +
"<description>Get the current weather</description>\n" +
"<parameters>\n" +
"<parameter>\n<name>city</name>\n<type>string</type>\n<description>The city name</description>\n</parameter>\n" +
"<required>[\"city\"]</required>\n" +
"</parameters>\n</function>\n</tools>\n\n" +
"If 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>\n\n<IMPORTANT>\nReminder:\n" +
"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
"- Required parameters MUST be specified\n" +
"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
"</IMPORTANT><|im_end|>\n" +
"<|im_start|>user\nWhat's the weather in Paris?<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
"<|im_start|>user\n<tool_response>\nSunny, 72F\n</tool_response>\n<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "assistant with content and tool call",
msgs: []api.Message{
{Role: "user", Content: "What's the weather?"},
{
Role: "assistant",
Content: "Let me check that for you.",
ToolCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "Paris"},
},
},
},
},
{Role: "tool", Content: "Sunny"},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}},
},
},
},
},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n" +
"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
"<function>\n<name>get_weather</name>\n" +
"<parameters>\n" +
"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
"</parameters>\n</function>\n</tools>\n\n" +
"If 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>\n\n<IMPORTANT>\nReminder:\n" +
"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
"- Required parameters MUST be specified\n" +
"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
"</IMPORTANT><|im_end|>\n" +
"<|im_start|>user\nWhat's the weather?<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>Let me check that for you.\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
"<|im_start|>user\n<tool_response>\nSunny\n</tool_response>\n<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "thinking in history is truncated",
msgs: []api.Message{
{Role: "user", Content: "Hi"},
{Role: "assistant", Content: "Hello!", Thinking: "Let me think about this..."},
{Role: "user", Content: "How are you?"},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n<|im_end|>\n" +
"<|im_start|>user\nHi<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>Hello!<|im_end|>\n" +
"<|im_start|>user\nHow are you?<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "parallel tool calls",
msgs: []api.Message{
{Role: "user", Content: "Weather in Paris and London?"},
{
Role: "assistant",
ToolCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "Paris"},
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{"city": "London"},
},
},
},
},
{Role: "tool", Content: "Sunny"},
{Role: "tool", Content: "Rainy"},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}},
},
},
},
},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n" +
"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
"<function>\n<name>get_weather</name>\n" +
"<parameters>\n" +
"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
"</parameters>\n</function>\n</tools>\n\n" +
"If 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>\n\n<IMPORTANT>\nReminder:\n" +
"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
"- Required parameters MUST be specified\n" +
"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
"</IMPORTANT><|im_end|>\n" +
"<|im_start|>user\nWeather in Paris and London?<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nLondon\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
"<|im_start|>user\n<tool_response>\nSunny\n</tool_response>\n<tool_response>\nRainy\n</tool_response>\n<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
{
name: "thinking disabled even when model supports it",
msgs: []api.Message{
{Role: "user", Content: "Hello!"},
},
isThinking: true, // model supports thinking
thinkValue: nil, // but user didn't request it
expected: "<|im_start|>system\n<|im_end|>\n" +
"<|im_start|>user\nHello!<|im_end|>\n" +
"<|im_start|>assistant\n<think></think>",
},
{
name: "complex message history with thinking, tools, tool calls, tool results and content",
msgs: []api.Message{
{Role: "user", Content: "What's the weather in Paris and London? Also, what's 2+2?"},
{Role: "assistant", Content: "", Thinking: "I need to check the weather for both cities and calculate 2+2. Let me start with the weather calls.", ToolCalls: []api.ToolCall{
{Function: api.ToolCallFunction{Name: "get_weather", Arguments: api.ToolCallFunctionArguments{"city": "Paris"}}},
{Function: api.ToolCallFunction{Name: "get_weather", Arguments: api.ToolCallFunctionArguments{"city": "London"}}},
}},
{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call1"},
{Role: "tool", Content: "Rainy, 15°C", ToolCallID: "call2"},
{Role: "assistant", Content: "", Thinking: "Now I have the weather data. Let me calculate 2+2.", ToolCalls: []api.ToolCall{
{Function: api.ToolCallFunction{Name: "calculate", Arguments: api.ToolCallFunctionArguments{"expression": "2+2"}}},
}},
{Role: "tool", Content: "4", ToolCallID: "call3"},
{Role: "assistant", Content: "Based on the weather data, Paris is sunny at 22°C and London is rainy at 15°C. Also, 2+2 equals 4.", Thinking: "Perfect! I have all the information needed to provide a complete answer."},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Parameters: api.ToolFunctionParameters{
Type: "object",
Properties: map[string]api.ToolProperty{
"city": {Type: api.PropertyType{"string"}},
},
},
},
},
{
Type: "function",
Function: api.ToolFunction{
Name: "calculate",
Parameters: api.ToolFunctionParameters{
Type: "object",
Properties: map[string]api.ToolProperty{
"expression": {Type: api.PropertyType{"string"}},
},
},
},
},
},
isThinking: true,
thinkValue: &api.ThinkValue{Value: true},
expected: "<|im_start|>system\n" +
"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
"<function>\n<name>get_weather</name>\n" +
"<parameters>\n" +
"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
"</parameters>\n</function>\n" +
"<function>\n<name>calculate</name>\n" +
"<parameters>\n" +
"<parameter>\n<name>expression</name>\n<type>string</type>\n</parameter>\n" +
"</parameters>\n</function>\n</tools>\n\n" +
"If 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>\n\n<IMPORTANT>\nReminder:\n" +
"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
"- Required parameters MUST be specified\n" +
"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
"</IMPORTANT><|im_end|>\n" +
"<|im_start|>user\nWhat's the weather in Paris and London? Also, what's 2+2?<|im_end|>\n" +
"<|im_start|>assistant\n" +
"<think>\nI need to check the weather for both cities and calculate 2+2. Let me start with the weather calls.\n</think>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n" +
"<tool_call>\n<function=get_weather>\n<parameter=city>\nLondon\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
"<|im_start|>user\n<tool_response>\nSunny, 22°C\n</tool_response>\n<tool_response>\nRainy, 15°C\n</tool_response>\n<|im_end|>\n" +
"<|im_start|>assistant\n" +
"<think>\nNow I have the weather data. Let me calculate 2+2.\n</think>\n" +
"<tool_call>\n<function=calculate>\n<parameter=expression>\n2+2\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
"<|im_start|>user\n<tool_response>\n4\n</tool_response>\n<|im_end|>\n" +
"<|im_start|>assistant\n" +
"<think>\nPerfect! I have all the information needed to provide a complete answer.\n</think>\n" +
"Based on the weather data, Paris is sunny at 22°C and London is rainy at 15°C. Also, 2+2 equals 4.<|im_end|>\n" +
"<|im_start|>assistant\n<think>\n",
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
renderer := &Nemotron3NanoRenderer{IsThinking: tt.isThinking}
rendered, err := renderer.Render(tt.msgs, tt.tools, tt.thinkValue)
if err != nil {
t.Fatal(err)
}
if diff := cmp.Diff(rendered, tt.expected); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
}
}