summaryrefslogtreecommitdiff
path: root/internal/llm/provider.go
blob: 84efaf9f2f49f91a0cc3dec23dbd3bc6cb2b227e (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
// Summary: LLM provider interfaces, request options, configuration, and factory to build a client from config.
package llm

import (
	"context"
	"errors"
	"strings"
)

// Message represents a chat-style prompt message.
type Message struct {
	Role    string
	Content string
}

// Client is a minimal LLM provider interface.
// Future providers (Ollama, etc.) should implement this.
type Client interface {
	// Chat sends chat messages and returns the assistant text.
	Chat(ctx context.Context, messages []Message, opts ...RequestOption) (string, error)
	// Name returns the provider's short name (e.g., "openai", "ollama").
	Name() string
	// DefaultModel returns the configured default model name.
	DefaultModel() string
}

// Streamer is an optional interface that providers may implement to support
// token-by-token streaming responses. Callers can type-assert to Streamer and
// fall back to Client.Chat when not implemented.
type Streamer interface {
	// ChatStream sends chat messages and invokes onDelta with incremental text
	// chunks as they are produced by the model. Implementations should call
	// onDelta with empty strings sparingly (prefer only non-empty chunks).
	ChatStream(ctx context.Context, messages []Message, onDelta func(string), opts ...RequestOption) error
}

// CodeCompleter is an optional interface for providers that support a
// prompt/suffix code-completion API (e.g., Copilot Codex endpoint). Clients
// can type-assert to this and prefer it over chat when available.
type CodeCompleter interface {
	// CodeCompletion requests up to n suggestions given a left-hand prompt and
	// right-hand suffix around the cursor. Language is advisory and may be
	// ignored. Temperature applies when provider supports it.
	CodeCompletion(ctx context.Context, prompt string, suffix string, n int, language string, temperature float64) ([]string, error)
}

// Options for a request. Providers may ignore unsupported fields.
type Options struct {
	Model       string
	Temperature float64
	MaxTokens   int
	Stop        []string
}

// RequestOption mutates Options.
type RequestOption func(*Options)

func WithModel(model string) RequestOption    { return func(o *Options) { o.Model = model } }
func WithTemperature(t float64) RequestOption { return func(o *Options) { o.Temperature = t } }
func WithMaxTokens(n int) RequestOption       { return func(o *Options) { o.MaxTokens = n } }
func WithStop(stop ...string) RequestOption {
	return func(o *Options) { o.Stop = append([]string{}, stop...) }
}

// Config defines provider configuration read from the Hexai config file.
type Config struct {
	Provider string
	// OpenAI options
	OpenAIBaseURL     string
	OpenAIModel       string
	OpenAITemperature *float64
	// Ollama options
	OllamaBaseURL     string
	OllamaModel       string
	OllamaTemperature *float64
	// Copilot options
	CopilotBaseURL     string
	CopilotModel       string
	CopilotTemperature *float64
}

// NewFromConfig creates an LLM client using only the supplied configuration.
// The OpenAI API key is supplied separately and may be read from the environment
// by the caller; other environment-based configuration is not used.
func NewFromConfig(cfg Config, openAIAPIKey, copilotAPIKey string) (Client, error) {
	p := strings.ToLower(strings.TrimSpace(cfg.Provider))
	if p == "" {
		p = "openai"
	}
	switch p {
	case "openai":
		if strings.TrimSpace(openAIAPIKey) == "" {
			return nil, errors.New("missing OPENAI_API_KEY for provider openai")
		}
		// Default temperature selection:
		// - When model is gpt-5*, prefer 1.0 by default (more exploratory).
		// - Otherwise, prefer 0.2 by default (coding friendly).
		// The app-wide defaults currently set provider temps to 0.2.
		// If the user hasn't explicitly overridden and the model is gpt-5*,
		// upgrade 0.2 → 1.0 to satisfy the requested default for gpt-5.
		model := strings.ToLower(strings.TrimSpace(cfg.OpenAIModel))
		if strings.HasPrefix(model, "gpt-5") {
			if cfg.OpenAITemperature == nil {
				v := 1.0
				cfg.OpenAITemperature = &v
			} else if *cfg.OpenAITemperature == 0.2 {
				v := 1.0
				cfg.OpenAITemperature = &v
			}
		} else if cfg.OpenAITemperature == nil {
			v := 0.2
			cfg.OpenAITemperature = &v
		}
		return newOpenAI(cfg.OpenAIBaseURL, cfg.OpenAIModel, openAIAPIKey, cfg.OpenAITemperature), nil
	case "ollama":
		if cfg.OllamaTemperature == nil {
			t := 0.2
			cfg.OllamaTemperature = &t
		}
		return newOllama(cfg.OllamaBaseURL, cfg.OllamaModel, cfg.OllamaTemperature), nil
	case "copilot":
		if strings.TrimSpace(copilotAPIKey) == "" {
			return nil, errors.New("missing COPILOT_API_KEY for provider copilot")
		}
		if cfg.CopilotTemperature == nil {
			t := 0.2
			cfg.CopilotTemperature = &t
		}
		return newCopilot(cfg.CopilotBaseURL, cfg.CopilotModel, copilotAPIKey, cfg.CopilotTemperature), nil
	default:
		return nil, errors.New("unknown LLM provider: " + p)
	}
}