Implementation:Ollama Ollama Chat Prompt

Knowledge Sources	Ollama
Domains	NLP, Prompt_Engineering
Last Updated	2026-02-14 00:00 GMT

Overview

Concrete tool for constructing model-specific prompts from chat messages provided by the server package.

Description

The chatPrompt function accepts a list of chat messages and produces the final prompt string and image data for the inference engine. It performs context window truncation by iteratively removing older messages until the rendered prompt fits within the model's num_ctx limit. System messages are always preserved regardless of truncation.

The companion Template.Execute method renders messages using Go's text/template engine with model-specific format strings. Parse parses the template string from model metadata.

Usage

Called internally by GenerateHandler and ChatHandler before dispatching to the inference engine. Not directly accessible to API clients.

Code Reference

Source Location

• Repository: ollama
• File: server/prompt.go (chatPrompt), template/template.go (Template.Execute, Parse)
• Lines: prompt.go:L23-131 (chatPrompt), template.go:L257-290 (Execute), template.go:L145-256 (Parse)

Signature

func chatPrompt(
    ctx context.Context,
    m *Model,
    tokenize tokenizeFunc,
    opts *api.Options,
    msgs []api.Message,
    tools []api.Tool,
    think *api.ThinkValue,
    truncate bool,
) (prompt string, images []llm.ImageData, _ error)

func (t *Template) Execute(w io.Writer, v Values) error

func Parse(s string) (*Template, error)

Import

import "github.com/ollama/ollama/server"
import "github.com/ollama/ollama/template"

I/O Contract

Inputs

Name	Type	Required	Description
ctx	context.Context	Yes	Request context
m	*Model	Yes	Model with template, config, projector paths
tokenize	tokenizeFunc	Yes	Function to tokenize text (from the loaded runner)
opts	*api.Options	Yes	Options including NumCtx (context window size)
msgs	[]api.Message	Yes	Chat messages with Role, Content, Images fields
tools	[]api.Tool	No	Tool/function definitions for function calling
think	*api.ThinkValue	No	Chain-of-thought thinking mode setting
truncate	bool	Yes	Whether to truncate messages to fit context window

Outputs

Name	Type	Description
prompt	string	Rendered prompt string in model-specific format
images	[]llm.ImageData	Extracted image data for multimodal models (ID + bytes)
error	error	Non-nil if template execution or tokenization fails

Usage Examples

Internal Usage

// From ChatHandler in server/routes.go
prompt, images, err := chatPrompt(
    c.Request.Context(),
    model,
    runner.Tokenize,
    &opts,
    msgs,
    req.Tools,
    req.Think,
    true, // truncate to fit context
)
if err != nil {
    // handle error
}
// Pass prompt and images to inference engine

Related Pages

Implements Principle

• Principle:Ollama_Ollama_Prompt_Construction

Requires Environment

• Environment:Ollama_Ollama_Go_Runtime