go-llm/v2/ollama/native.go

// Package ollama implements the go-llm v2 provider interface for Ollama,
// targeting Ollama's native /api/chat endpoint. Supports both local Ollama
// instances (no API key) and Ollama Cloud (https://ollama.com, requires an
// API key).
package ollama

import (
	"context"
	"encoding/json"
	"errors"
	"net/http"

	"gitea.stevedudenhoeffer.com/steve/go-llm/v2/provider"
)

// DefaultLocalBaseURL is the default base URL for a locally-running Ollama
// instance.
const DefaultLocalBaseURL = "http://localhost:11434"

// DefaultCloudBaseURL is the default base URL for Ollama Cloud.
const DefaultCloudBaseURL = "https://ollama.com"

// Provider implements provider.Provider over Ollama's native /api/chat
// endpoint. An empty apiKey means local-mode (no Authorization header sent);
// a non-empty apiKey is sent as a Bearer token (cloud-mode).
type Provider struct {
	apiKey  string
	baseURL string
	client  *http.Client
}

// newNative constructs a native Ollama provider. Callers should use the
// package-level New() constructor or the v2 llm.Ollama() / llm.OllamaCloud()
// helpers.
func newNative(apiKey, baseURL string) *Provider {
	return &Provider{
		apiKey:  apiKey,
		baseURL: baseURL,
		client:  &http.Client{},
	}
}

// nativeChatRequest is the JSON body POSTed to /api/chat.
type nativeChatRequest struct {
	Model    string              `json:"model"`
	Messages []nativeChatMessage `json:"messages"`
	Tools    []nativeToolDef     `json:"tools,omitempty"`
	Stream   bool                `json:"stream"`
	// Think is polymorphic — Ollama accepts true/false or "low"/"medium"/"high".
	Think   json.RawMessage `json:"think,omitempty"`
	Options map[string]any  `json:"options,omitempty"`
}

// nativeChatMessage is one entry in the messages array on the wire. It also
// carries assistant tool calls and tool-role responses.
type nativeChatMessage struct {
	Role       string           `json:"role"`
	Content    string           `json:"content,omitempty"`
	Images     []string         `json:"images,omitempty"`
	ToolCalls  []nativeToolCall `json:"tool_calls,omitempty"`
	ToolCallID string           `json:"tool_call_id,omitempty"`
	Thinking   string           `json:"thinking,omitempty"`
}

// nativeToolCall mirrors Ollama's tool-call wire shape: a function with name
// and JSON-encoded arguments. Ollama's spec doesn't require an id, but some
// builds and some streaming chunks include one — we accept it on both wire and
// internal sides.
type nativeToolCall struct {
	ID       string             `json:"id,omitempty"`
	Function nativeFunctionCall `json:"function"`
}

type nativeFunctionCall struct {
	Index     *int            `json:"index,omitempty"`
	Name      string          `json:"name,omitempty"`
	Arguments json.RawMessage `json:"arguments,omitempty"`
}

// nativeChatResponse is the JSON body returned from a non-streaming /api/chat
// call (and is also the per-line shape during streaming).
type nativeChatResponse struct {
	Model           string            `json:"model,omitempty"`
	Message         nativeChatMessage `json:"message"`
	Done            bool              `json:"done"`
	DoneReason      string            `json:"done_reason,omitempty"`
	PromptEvalCount int               `json:"prompt_eval_count,omitempty"`
	EvalCount       int               `json:"eval_count,omitempty"`
	TotalDuration   int64             `json:"total_duration,omitempty"`
}

// nativeToolDef is the wire shape of a tool definition sent to Ollama.
type nativeToolDef struct {
	Type     string            `json:"type"`
	Function nativeFunctionDef `json:"function"`
}

type nativeFunctionDef struct {
	Name        string         `json:"name"`
	Description string         `json:"description,omitempty"`
	Parameters  map[string]any `json:"parameters,omitempty"`
}

// encodeThink converts a go-llm Reasoning string ("", "low", "medium",
// "high", or the literal strings "true"/"false") into Ollama's polymorphic
// `think` field. Returns nil for the empty string so the field is omitted.
func encodeThink(reasoning string) json.RawMessage {
	switch reasoning {
	case "":
		return nil
	case "true":
		return json.RawMessage(`true`)
	case "false":
		return json.RawMessage(`false`)
	default:
		// "low" / "medium" / "high" — encode as a JSON string.
		b, _ := json.Marshal(reasoning)
		return b
	}
}

var errNotImplemented = errors.New("ollama native provider: not implemented")

// Complete performs a non-streaming chat completion via /api/chat.
func (p *Provider) Complete(ctx context.Context, req provider.Request) (provider.Response, error) {
	return provider.Response{}, errNotImplemented
}

// Stream performs a streaming chat completion via /api/chat with
// `stream: true`, parsing NDJSON line-by-line.
func (p *Provider) Stream(ctx context.Context, req provider.Request, events chan<- provider.StreamEvent) error {
	return errNotImplemented
}