renderers

package
v0.40.13 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Oct 28, 2026 License: MIT Imports: 20 Imported by: 1

Documentation

Index

Constants

View Source
const (
	// LFM3SpeechChunkTokens bounds one generation; longer input is read in sentence
	// groups of about 20 seconds each.
	LFM3SpeechChunkTokens = 80
)

Variables

View Source
var RenderImgTags bool

RenderImgTags is a global flag that tells renderers to use [img] tags for images. This is set by the Ollama server package on init, or left as false for other environments where renderers are used

Functions

func AddJSONSpaces added in v0.40.0

func AddJSONSpaces(b []byte) []byte

AddJSONSpaces adds Python-style separator spaces to compact JSON without changing string values or escaping.

func LFM3EscapeSpeechText added in v0.40.13

func LFM3EscapeSpeechText(text string) string

LFM3EscapeSpeechText swaps the brackets of control-token spellings for full-width ones, so text to read aloud cannot end its turn or switch modes.

func LFM3SpeechChunks added in v0.40.13

func LFM3SpeechChunks(text string, limit int, count func(string) int) []string

LFM3SpeechChunks splits text into groups of whole sentences of at most limit tokens; sentences over the limit split between words, and words between characters.

func LFM3SpeechFrameBudget added in v0.40.13

func LFM3SpeechFrameBudget(textTokens int) int

LFM3SpeechFrameBudget is the most codec frames one chunk of textTokens may generate.

func LFM3SpeechVoices added in v0.40.13

func LFM3SpeechVoices() []string

LFM3SpeechVoices lists the voices to advertise: the default, then a female and a male voice for each language under its full name ("male-british-english").

func LanguageForName added in v0.40.12

func LanguageForName(s string) (language.Tag, bool)

LanguageForName finds the ISO 639-1 language with the English name s.

func LanguageName added in v0.40.12

func LanguageName(tag language.Tag) string

LanguageName returns the English name of tag's base language, the form speech models take in instructions ("Chinese" for zh-TW), or "" when the tag names no known language.

func LeadingBOSForRenderer added in v0.30.0

func LeadingBOSForRenderer(name string) string

func PythonJSON added in v0.40.4

func PythonJSON(v any, indent int) (string, error)

PythonJSON renders v like Python's json.dumps(v, ensure_ascii=False), indented by indent spaces if indent > 0.

func PythonJSONCompact added in v0.40.4

func PythonJSONCompact(v any) (string, error)

PythonJSONCompact renders v like json.dumps(v, ensure_ascii=False, separators=(",", ":")).

func Register added in v0.12.6

func Register(name string, renderer RendererConstructor)

func RenderWithRenderer

func RenderWithRenderer(name string, msgs []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func ResolveThinking added in v0.34.3

func ResolveThinking(requestedThink *api.ThinkValue, thinking *model.Thinking) *api.ThinkValue

ResolveThinking preserves explicit booleans and supported names. Omission or an unsupported name uses the default. Unknown metadata preserves legacy handling.

func ThinkingForRenderer added in v0.34.3

func ThinkingForRenderer(name string) *model.Thinking

ThinkingForRenderer returns the controls of the selected renderer variant. Unknown or invalid descriptors are omitted.

Types

type AudioPrompt added in v0.40.12

type AudioPrompt struct {
	// Messages carry the clip in Images; an "[img]" marker in Content
	// positions it within the text.
	Messages []api.Message
	// Think must be set for renderers that think by default; nil keeps the
	// renderer's default.
	Think *api.ThinkValue
}

AudioPrompt is the chat a renderer asks for to serve an AudioRequest.

type AudioPrompter added in v0.40.12

type AudioPrompter interface {
	// AudioClipLimit is the longest clip one generation should take; longer
	// audio is split and transcribed in pieces. Zero means no limit.
	AudioClipLimit() time.Duration
	// AudioPrompt returns an error for requests the model can't serve, such
	// as a translation target it doesn't produce; the client gets a 400.
	AudioPrompt(AudioRequest) (AudioPrompt, error)
	// AudioText extracts the requested text from a generation's output.
	AudioText(req AudioRequest, output string) string
}

AudioPrompter builds speech-to-text prompts for the models a renderer serves, following the publisher's documented instructions.

func AudioPrompterForRenderer added in v0.40.12

func AudioPrompterForRenderer(name string) AudioPrompter

AudioPrompterForRenderer returns the named renderer's AudioPrompter, or a generic one for renderers without speech-specific instructions.

type AudioRequest added in v0.40.12

type AudioRequest struct {
	Task AudioTask
	// Source is the spoken language; language.Und asks the model to detect it.
	Source language.Tag
	// Target is the output language of a translation.
	Target language.Tag
	// Prompt is optional client text describing the audio.
	Prompt string
	Audio  api.ImageData
}

AudioRequest is one speech-to-text generation over a clip.

type AudioTask added in v0.40.12

type AudioTask int

AudioTask is the speech-to-text task a request asks for.

const (
	AudioTranscribe AudioTask = iota
	AudioTranslate
)

type CogitoRenderer added in v0.13.0

type CogitoRenderer struct {
	// contains filtered or unexported fields
}

func (*CogitoRenderer) LeadingBOS added in v0.30.0

func (r *CogitoRenderer) LeadingBOS() string

func (*CogitoRenderer) Render added in v0.13.0

func (r *CogitoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*CogitoRenderer) Thinking added in v0.34.3

func (r *CogitoRenderer) Thinking() *model.Thinking

type CohereRenderer added in v0.30.10

type CohereRenderer struct{}

CohereRenderer renders the Cohere North / Command A 2026 chat template (Cohere2 MoE models such as CohereLabs/North-Mini-Code-1.0): a platform system turn with an Available Tools section, <|START_TEXT|>-wrapped message bodies, <|START_THINKING|> reasoning, <|START_ACTION|> tool calls, and <|START_TOOL_RESULT|> tool results.

The chat template's model-specific platform instructions (identity, default policies) belong in the model's Modelfile SYSTEM prompt, which arrives here as the first system message.

func (*CohereRenderer) LeadingBOS added in v0.30.10

func (r *CohereRenderer) LeadingBOS() string

func (*CohereRenderer) Render added in v0.30.10

func (r *CohereRenderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*CohereRenderer) Thinking added in v0.34.3

func (r *CohereRenderer) Thinking() *model.Thinking

type DeepSeek3Renderer added in v0.13.4

type DeepSeek3Renderer struct {
	IsThinking bool
	Variant    DeepSeek3Variant
}

func (*DeepSeek3Renderer) LeadingBOS added in v0.30.0

func (r *DeepSeek3Renderer) LeadingBOS() string

func (*DeepSeek3Renderer) Render added in v0.13.4

func (r *DeepSeek3Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*DeepSeek3Renderer) Thinking added in v0.34.3

func (r *DeepSeek3Renderer) Thinking() *model.Thinking

type DeepSeek3Variant added in v0.13.4

type DeepSeek3Variant int
const (
	Deepseek31 DeepSeek3Variant = iota
)

type FlintRenderer added in v0.40.8

type FlintRenderer struct{}

func (*FlintRenderer) LeadingBOS added in v0.40.8

func (*FlintRenderer) LeadingBOS() string

func (*FlintRenderer) Render added in v0.40.8

func (*FlintRenderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*FlintRenderer) Thinking added in v0.40.8

func (*FlintRenderer) Thinking() *model.Thinking

The publisher's bare assistant prompt lets the model choose whether to reason. Explicit thinking controls add a prefill below.

type FunctionGemmaRenderer added in v0.13.5

type FunctionGemmaRenderer struct{}

func (*FunctionGemmaRenderer) LeadingBOS added in v0.30.0

func (r *FunctionGemmaRenderer) LeadingBOS() string

func (*FunctionGemmaRenderer) Render added in v0.13.5

func (r *FunctionGemmaRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*FunctionGemmaRenderer) Thinking added in v0.34.3

func (r *FunctionGemmaRenderer) Thinking() *model.Thinking

type GLM46Renderer added in v0.14.3

type GLM46Renderer struct{}

func (*GLM46Renderer) LeadingBOS added in v0.30.0

func (r *GLM46Renderer) LeadingBOS() string

func (*GLM46Renderer) Render added in v0.14.3

func (r *GLM46Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

type GLM47Renderer added in v0.14.3

type GLM47Renderer struct{}

GLM47Renderer renders messages for GLM-4.7 models.

GLM-4.7 Thinking Modes (ref: https://docs.z.ai/guides/capabilities/thinking-mode):

  1. INTERLEAVED THINKING The model thinks between tool calls and after receiving tool results. This enables complex step-by-step reasoning: interpreting each tool output before deciding what to do next. Thinking blocks are preserved and returned with tool results to maintain reasoning continuity.

  2. PRESERVED THINKING The model retains reasoning content from previous assistant turns in context. This preserves reasoning continuity across multi-turn conversations. The upstream API has a "clear_thinking" parameter to control this: - clear_thinking=true: clears reasoning from previous turns (outputs </think>) - clear_thinking=false: preserves <think>...</think> blocks from previous turns

  3. TURN-LEVEL THINKING Controls whether the model should reason on each turn. The upstream API uses "enable_thinking" parameter: - enable_thinking=true: outputs <think> to start reasoning - enable_thinking=false: outputs </think> to skip reasoning

OLLAMA DEFAULTS:

  • Thinking is ENABLED by default (thinkValue=nil or true outputs <think>)
  • Thinking is PRESERVED by default (reasoning content from previous turns is always included in <think>...</think> blocks, equivalent to clear_thinking=false)
  • Users can disable thinking per-turn via thinkValue=false

func (*GLM47Renderer) LeadingBOS added in v0.30.0

func (r *GLM47Renderer) LeadingBOS() string

func (*GLM47Renderer) Render added in v0.14.3

func (r *GLM47Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*GLM47Renderer) Thinking added in v0.34.3

func (r *GLM47Renderer) Thinking() *model.Thinking

type Gemma4Renderer added in v0.20.0

type Gemma4Renderer struct {
	// contains filtered or unexported fields
}

Gemma4Renderer renders prompts using Gemma 4's chat format with <|turn>/<turn|> markers, <|"|> string delimiters, and <|tool>/ <|tool_call>/<|tool_response> tags for function calling.

func (*Gemma4Renderer) AudioClipLimit added in v0.40.12

func (*Gemma4Renderer) AudioClipLimit() time.Duration

Gemma 4 takes up to 30 s of audio, placed after the instruction in a single user turn.

func (*Gemma4Renderer) AudioPrompt added in v0.40.12

func (r *Gemma4Renderer) AudioPrompt(req AudioRequest) (AudioPrompt, error)

func (*Gemma4Renderer) AudioText added in v0.40.12

func (r *Gemma4Renderer) AudioText(req AudioRequest, output string) string

AudioText keeps only the translation from a transcribe-then-translate answer.

func (*Gemma4Renderer) LeadingBOS added in v0.30.0

func (r *Gemma4Renderer) LeadingBOS() string

func (*Gemma4Renderer) Render added in v0.20.0

func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*Gemma4Renderer) Thinking added in v0.34.3

func (r *Gemma4Renderer) Thinking() *model.Thinking

type GlimmerRenderer added in v0.32.7

type GlimmerRenderer struct {
	// contains filtered or unexported fields
}

func (*GlimmerRenderer) LeadingBOS added in v0.32.7

func (r *GlimmerRenderer) LeadingBOS() string

func (*GlimmerRenderer) Render added in v0.32.7

func (r *GlimmerRenderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*GlimmerRenderer) Thinking added in v0.34.3

func (r *GlimmerRenderer) Thinking() *model.Thinking

type GlmOcrRenderer added in v0.15.5

type GlmOcrRenderer struct {
	// contains filtered or unexported fields
}

func (*GlmOcrRenderer) LeadingBOS added in v0.30.0

func (r *GlmOcrRenderer) LeadingBOS() string

func (*GlmOcrRenderer) Render added in v0.15.5

func (r *GlmOcrRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*GlmOcrRenderer) Thinking added in v0.34.3

func (r *GlmOcrRenderer) Thinking() *model.Thinking

type LFM2Renderer added in v0.14.3

type LFM2Renderer struct {
	IsThinking bool
	// contains filtered or unexported fields
}

func (*LFM2Renderer) LeadingBOS added in v0.30.0

func (r *LFM2Renderer) LeadingBOS() string

func (*LFM2Renderer) Render added in v0.14.3

func (r *LFM2Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*LFM2Renderer) Thinking added in v0.34.3

func (r *LFM2Renderer) Thinking() *model.Thinking

type LFM3Renderer added in v0.40.12

type LFM3Renderer struct {
	// contains filtered or unexported fields
}

LFM3Renderer covers three publisher template branches: text and VL (thinking by default), Omni (non-reasoning), and Audio (an older template without merged tool results; an explicit think selects the README's prefills).

func (*LFM3Renderer) AudioClipLimit added in v0.40.12

func (*LFM3Renderer) AudioClipLimit() time.Duration

The publisher documents no clip limit; 30 s matches the other speech models and keeps each generation short.

func (*LFM3Renderer) AudioPrompt added in v0.40.12

func (r *LFM3Renderer) AudioPrompt(req AudioRequest) (AudioPrompt, error)

AudioPrompt follows the publisher's recipes: the instruction is the system turn and the user turn holds only the audio. Client prompt text is left out; any user text pulls the model toward describing or answering the clip.

func (*LFM3Renderer) AudioText added in v0.40.12

func (*LFM3Renderer) AudioText(_ AudioRequest, output string) string

func (*LFM3Renderer) LeadingBOS added in v0.40.12

func (*LFM3Renderer) LeadingBOS() string

func (*LFM3Renderer) Render added in v0.40.12

func (r *LFM3Renderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*LFM3Renderer) SpeechChunks added in v0.40.13

func (*LFM3Renderer) SpeechChunks(input string, count func(string) int) []SpeechChunk

func (*LFM3Renderer) SpeechVoice added in v0.40.13

func (*LFM3Renderer) SpeechVoice(name string) (SpeechVoice, error)

func (*LFM3Renderer) SpeechVoices added in v0.40.13

func (r *LFM3Renderer) SpeechVoices() []string

func (*LFM3Renderer) Thinking added in v0.40.12

func (r *LFM3Renderer) Thinking() *model.Thinking

type LFM3Voice added in v0.40.13

type LFM3Voice struct{ Gender, Language string }

LFM3Voice is the speaker the TTS prompt asks for.

func ParseLFM3Voice added in v0.40.13

func ParseLFM3Voice(name string) (LFM3Voice, error)

ParseLFM3Voice reads "female" or "male", optionally followed by a language ("male-french", "female-en-gb"), or a language alone. The default, also named "default" as with other speech models, is a female American English voice.

type LagunaRenderer added in v0.22.0

type LagunaRenderer struct{}

func (*LagunaRenderer) LeadingBOS added in v0.30.0

func (r *LagunaRenderer) LeadingBOS() string

func (*LagunaRenderer) Render added in v0.22.0

func (r *LagunaRenderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*LagunaRenderer) Thinking added in v0.34.3

func (r *LagunaRenderer) Thinking() *model.Thinking

type LagunaV8Renderer added in v0.32.2

type LagunaV8Renderer struct{}

func (*LagunaV8Renderer) LeadingBOS added in v0.32.2

func (r *LagunaV8Renderer) LeadingBOS() string

func (*LagunaV8Renderer) Render added in v0.32.2

func (r *LagunaV8Renderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*LagunaV8Renderer) Thinking added in v0.34.3

func (r *LagunaV8Renderer) Thinking() *model.Thinking

type Nemotron3NanoRenderer added in v0.13.4

type Nemotron3NanoRenderer struct {
	// contains filtered or unexported fields
}

func (*Nemotron3NanoRenderer) LeadingBOS added in v0.30.0

func (r *Nemotron3NanoRenderer) LeadingBOS() string

func (*Nemotron3NanoRenderer) Render added in v0.13.4

func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error)

func (*Nemotron3NanoRenderer) Thinking added in v0.34.3

func (r *Nemotron3NanoRenderer) Thinking() *model.Thinking

type Olmo3Renderer added in v0.13.3

type Olmo3Renderer struct {
	UseExtendedSystemMessage bool
}

func (*Olmo3Renderer) LeadingBOS added in v0.30.0

func (r *Olmo3Renderer) LeadingBOS() string

func (*Olmo3Renderer) Render added in v0.13.3

func (r *Olmo3Renderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error)

func (*Olmo3Renderer) Thinking added in v0.34.3

func (r *Olmo3Renderer) Thinking() *model.Thinking

type Olmo3ThinkRenderer added in v0.13.3

type Olmo3ThinkRenderer struct {
	Variant Olmo3ThinkVariant
}

func (*Olmo3ThinkRenderer) LeadingBOS added in v0.30.0

func (r *Olmo3ThinkRenderer) LeadingBOS() string

func (*Olmo3ThinkRenderer) Render added in v0.13.3

func (r *Olmo3ThinkRenderer) Render(messages []api.Message, _ []api.Tool, _ *api.ThinkValue) (string, error)

func (*Olmo3ThinkRenderer) Thinking added in v0.34.3

func (r *Olmo3ThinkRenderer) Thinking() *model.Thinking

type Olmo3ThinkVariant added in v0.13.4

type Olmo3ThinkVariant int
const (
	// Olmo3Think32B is for allenai/Olmo-3-32B-Think
	Olmo3Think32B Olmo3ThinkVariant = iota
	// Olmo31Think is for allenai/Olmo-3-7B-Think and allenai/Olmo-3.1-32B-Think (includes model info)
	Olmo31Think
)

type OrnithRenderer added in v0.30.11

type OrnithRenderer struct {
	Qwen35Renderer
}

type Qwen3CoderRenderer

type Qwen3CoderRenderer struct{}

func (*Qwen3CoderRenderer) LeadingBOS added in v0.30.0

func (r *Qwen3CoderRenderer) LeadingBOS() string

func (*Qwen3CoderRenderer) Render added in v0.12.6

func (r *Qwen3CoderRenderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error)

func (*Qwen3CoderRenderer) Thinking added in v0.34.3

func (r *Qwen3CoderRenderer) Thinking() *model.Thinking

type Qwen3VLRenderer added in v0.12.6

type Qwen3VLRenderer struct {
	// contains filtered or unexported fields
}

func (*Qwen3VLRenderer) LeadingBOS added in v0.30.0

func (r *Qwen3VLRenderer) LeadingBOS() string

func (*Qwen3VLRenderer) Render added in v0.12.6

func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*Qwen3VLRenderer) Thinking added in v0.34.3

func (r *Qwen3VLRenderer) Thinking() *model.Thinking

type Qwen35Renderer added in v0.17.7

type Qwen35Renderer struct {
	// contains filtered or unexported fields
}

func (*Qwen35Renderer) LeadingBOS added in v0.30.0

func (r *Qwen35Renderer) LeadingBOS() string

func (*Qwen35Renderer) Render added in v0.17.7

func (r *Qwen35Renderer) Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)

func (*Qwen35Renderer) Thinking added in v0.34.3

func (r *Qwen35Renderer) Thinking() *model.Thinking

type Renderer added in v0.12.6

type Renderer interface {
	Thinking() *model.Thinking
	Render(messages []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error)
	LeadingBOS() string
}

type RendererConstructor added in v0.12.6

type RendererConstructor func() Renderer

type RendererRegistry added in v0.12.6

type RendererRegistry struct {
	// contains filtered or unexported fields
}

func (*RendererRegistry) Register added in v0.12.6

func (r *RendererRegistry) Register(name string, renderer RendererConstructor)

type SpeechChunk added in v0.40.13

type SpeechChunk struct {
	// Text is read aloud, escaped for the model's tts template.
	Text string
	// MaxFrames caps the codec frames the generation may produce.
	MaxFrames int
}

SpeechChunk is one /tts generation of a speech request.

type SpeechPrompter added in v0.40.13

type SpeechPrompter interface {
	// SpeechVoices lists the voice names to advertise; renderers whose models
	// do not speak return none.
	SpeechVoices() []string
	// SpeechVoice resolves a voice name; the client gets a 400 on error.
	SpeechVoice(name string) (SpeechVoice, error)
	// SpeechChunks splits input into generations, or returns nil when there is
	// no text to speak; count returns the token count of a piece of text.
	SpeechChunks(input string, count func(string) int) []SpeechChunk
}

SpeechPrompter maps speech requests onto llama-server's /tts for the models a renderer serves; llama-server renders the model's own TTS template.

func SpeechPrompterForRenderer added in v0.40.13

func SpeechPrompterForRenderer(name string) (SpeechPrompter, bool)

SpeechPrompterForRenderer returns the named renderer's SpeechPrompter if its models read text aloud through llama-server.

type SpeechVoice added in v0.40.13

type SpeechVoice struct{ Voice, Language string }

SpeechVoice is the voice and language a llama-server /tts request asks for.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL