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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
|
// Package llm defines LLM provider interfaces, request options, configuration, and factory to build a client from config.
package llm
import (
"context"
"fmt"
"sort"
"strings"
"sync"
)
// 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)
// WithModel sets the model name for a request.
func WithModel(model string) RequestOption { return func(o *Options) { o.Model = model } }
// WithTemperature sets the sampling temperature for a request.
func WithTemperature(t float64) RequestOption { return func(o *Options) { o.Temperature = t } }
// WithMaxTokens sets the maximum number of tokens to generate.
func WithMaxTokens(n int) RequestOption { return func(o *Options) { o.MaxTokens = n } }
// WithStop sets custom stop sequences for a request.
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
RequestTimeout int // seconds; 0 means use default (30s)
// OpenAI options
OpenAIBaseURL string
OpenAIModel string
OpenAITemperature *float64
// OpenRouter options
OpenRouterBaseURL string
OpenRouterModel string
OpenRouterTemperature *float64
// Ollama options
OllamaBaseURL string
OllamaModel string
OllamaTemperature *float64
// Anthropic options
AnthropicBaseURL string
AnthropicModel string
AnthropicTemperature *float64
}
// ProviderKeys contains API credentials used by provider factories.
type ProviderKeys struct {
OpenAIAPIKey string
OpenRouterAPIKey string
AnthropicAPIKey string
}
// ProviderFactory builds an LLM client for a named provider.
type ProviderFactory func(cfg Config, keys ProviderKeys) (Client, error)
// providerRegistry is a package-level singleton populated by init() calls in
// each provider file (anthropic.go, openai.go, etc.). It must be a
// package-level var — rather than a constructor argument — because Go's
// init() mechanism runs before any application code, and the alternative
// (an explicit RegisterAll() in main) would require every binary that uses
// the llm package to manually enumerate all providers. The RWMutex makes the
// map safe for the rare case where RegisterProvider is called from a test
// goroutine after init() has completed.
var (
providerRegistryMu sync.RWMutex
providerRegistry = map[string]ProviderFactory{}
)
// RegisterProvider registers a provider factory by normalized name.
func RegisterProvider(name string, factory ProviderFactory) {
normalized := normalizeProvider(name)
if normalized == "" {
panic("llm: provider name cannot be empty")
}
if factory == nil {
panic("llm: provider factory cannot be nil")
}
providerRegistryMu.Lock()
defer providerRegistryMu.Unlock()
if _, exists := providerRegistry[normalized]; exists {
panic("llm: provider already registered: " + normalized)
}
providerRegistry[normalized] = factory
}
// 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, openRouterAPIKey, anthropicAPIKey string) (Client, error) {
provider := normalizeProvider(cfg.Provider)
if provider == "" {
provider = "openai"
}
factory, ok := lookupProviderFactory(provider)
if !ok {
return nil, unknownProviderError(provider)
}
return factory(cfg, ProviderKeys{
OpenAIAPIKey: openAIAPIKey,
OpenRouterAPIKey: openRouterAPIKey,
AnthropicAPIKey: anthropicAPIKey,
})
}
func normalizeProvider(provider string) string {
return strings.ToLower(strings.TrimSpace(provider))
}
func lookupProviderFactory(provider string) (ProviderFactory, bool) {
providerRegistryMu.RLock()
defer providerRegistryMu.RUnlock()
factory, ok := providerRegistry[provider]
return factory, ok
}
func withDefaultTemperature(configured *float64, fallback float64) *float64 {
if configured != nil {
return configured
}
v := fallback
return &v
}
func missingAPIKeyError(provider string, envVars ...string) error {
name := providerDisplayName(provider)
if len(envVars) == 0 {
return fmt.Errorf("missing %s API key", name)
}
return fmt.Errorf("missing %s API key for provider %s; set %s", name, normalizeProvider(provider), joinEnvVars(envVars))
}
func unknownProviderError(provider string) error {
return fmt.Errorf("unknown LLM provider %q; supported providers: %s", provider, strings.Join(supportedProviders(), ", "))
}
func providerDisplayName(provider string) string {
switch normalizeProvider(provider) {
case "openai":
return "OpenAI"
case "openrouter":
return "OpenRouter"
case "anthropic":
return "Anthropic"
default:
return provider
}
}
func joinEnvVars(envVars []string) string {
switch len(envVars) {
case 0:
return ""
case 1:
return envVars[0]
case 2:
return envVars[0] + " or " + envVars[1]
default:
return strings.Join(envVars[:len(envVars)-1], ", ") + ", or " + envVars[len(envVars)-1]
}
}
func supportedProviders() []string {
providerRegistryMu.RLock()
defer providerRegistryMu.RUnlock()
names := make([]string, 0, len(providerRegistry))
for name := range providerRegistry {
names = append(names, name)
}
sort.Strings(names)
return names
}
|