konnosDashboard
model.utils.ts167 lines · main
1export type ProviderName = 'bedrock' | 'openai'
2
3export type BedrockModel = 'anthropic.claude-3-7-sonnet-20250219-v1:0' | 'openai.gpt-oss-120b-1:0'
4
5export type OpenAIModelId = 'gpt-5.4-nano' | 'gpt-5.3-codex'
6
7// Source: https://developers.openai.com/api/docs/guides/reasoning + per-model pages
8export type ReasoningEffort = 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
9
10// Per-model reasoning effort compatibility.
11// Sources: https://developers.openai.com/api/docs/models/gpt-5.4-nano
12// https://developers.openai.com/api/docs/models/gpt-5.3-codex
13type ModelReasoningSupport = {
14 'gpt-5.4-nano': 'none' | 'low' | 'medium' | 'high' | 'xhigh'
15 'gpt-5.3-codex': 'low' | 'medium' | 'high' | 'xhigh'
16}
17
18type ReasoningEffortFor<ModelId extends OpenAIModelId> = ModelId extends keyof ModelReasoningSupport
19 ? ModelReasoningSupport[ModelId]
20 : never
21
22/** Type-safe factory for configuring OpenAI models with compatible reasoning efforts. */
23export function openaiModelEntry<
24 ModelId extends OpenAIModelId,
25 RequiresAdvance extends boolean = false,
26>(config: {
27 id: ModelId
28 /** When true, the model requires the `assistant.advance_model` entitlement (paid plans). Defaults to false. */
29 requiresAdvanceModelEntitlement?: RequiresAdvance
30 /**
31 * When omitted, OpenAI applies its own default reasoning effort for the model,
32 * which may not be zero. Use an explicit level to control cost and latency.
33 */
34 reasoningEffort?: ReasoningEffortFor<ModelId>
35}): {
36 id: ModelId
37 requiresAdvanceModelEntitlement: RequiresAdvance
38 reasoningEffort?: ReasoningEffortFor<ModelId>
39} {
40 return {
41 requiresAdvanceModelEntitlement: false as RequiresAdvance,
42 ...config,
43 }
44}
45
46export type OpenAIModelEntry = ReturnType<typeof openaiModelEntry>
47
48/** Default model entry for simple completion endpoints where latency is more important than reasoning. */
49export const DEFAULT_COMPLETION_MODEL = openaiModelEntry({
50 id: 'gpt-5.4-nano',
51 reasoningEffort: 'none',
52})
53
54// Single source of truth for all Assistant chat model variants and their reasoning levels.
55// Models with requiresAdvanceModelEntitlement false are available to all users; true requires the assistant.advance_model entitlement.
56export const ASSISTANT_MODELS = [
57 openaiModelEntry({
58 id: 'gpt-5.4-nano',
59 requiresAdvanceModelEntitlement: false,
60 reasoningEffort: 'low',
61 }),
62 openaiModelEntry({
63 id: 'gpt-5.3-codex',
64 requiresAdvanceModelEntitlement: true,
65 reasoningEffort: 'low',
66 }),
67] as const
68
69export type AssistantBaseModelId = Extract<
70 (typeof ASSISTANT_MODELS)[number],
71 { requiresAdvanceModelEntitlement: false }
72>['id']
73export type AssistantModelId = (typeof ASSISTANT_MODELS)[number]['id']
74
75const ASSISTANT_MODELS_MAP = Object.fromEntries(ASSISTANT_MODELS.map((m) => [m.id, m])) as Record<
76 AssistantModelId,
77 (typeof ASSISTANT_MODELS)[number]
78>
79
80export const DEFAULT_ASSISTANT_BASE_MODEL_ID = 'gpt-5.4-nano' satisfies AssistantBaseModelId
81
82export const DEFAULT_ASSISTANT_ADVANCE_MODEL_ID = 'gpt-5.3-codex' satisfies AssistantModelId
83
84export function defaultAssistantModelId(hasAccessToAdvanceModel: boolean): AssistantModelId {
85 return hasAccessToAdvanceModel
86 ? DEFAULT_ASSISTANT_ADVANCE_MODEL_ID
87 : DEFAULT_ASSISTANT_BASE_MODEL_ID
88}
89
90export function isKnownAssistantModelId(id: string): id is AssistantModelId {
91 return Object.hasOwn(ASSISTANT_MODELS_MAP, id)
92}
93
94export function isAssistantBaseModelId(id: string): id is AssistantBaseModelId {
95 return (
96 id in ASSISTANT_MODELS_MAP &&
97 !ASSISTANT_MODELS_MAP[id as AssistantModelId].requiresAdvanceModelEntitlement
98 )
99}
100
101export function isAdvanceOnlyModelId(id: string): boolean {
102 return (
103 id in ASSISTANT_MODELS_MAP &&
104 ASSISTANT_MODELS_MAP[id as AssistantModelId].requiresAdvanceModelEntitlement
105 )
106}
107
108export function getAssistantModelEntry(id: AssistantModelId): (typeof ASSISTANT_MODELS)[number] {
109 return ASSISTANT_MODELS_MAP[id]
110}
111
112export type Model = BedrockModel | OpenAIModelId
113
114export type ProviderModelConfig = {
115 /** Optional providerOptions to attach to the system message for this model */
116 systemProviderOptions?: Record<string, any>
117 /** The default model for this provider (used when limited or no preferred specified) */
118 default: boolean
119}
120
121export type ProviderRegistry = {
122 bedrock: {
123 models: Record<BedrockModel, ProviderModelConfig>
124 providerOptions?: Record<string, any>
125 }
126 openai: {
127 models: Record<OpenAIModelId, ProviderModelConfig>
128 providerOptions?: Record<string, any>
129 }
130}
131
132export const PROVIDERS: ProviderRegistry = {
133 bedrock: {
134 models: {
135 'anthropic.claude-3-7-sonnet-20250219-v1:0': {
136 systemProviderOptions: {
137 bedrock: {
138 // Always cache the system prompt (must not contain dynamic content)
139 cachePoint: { type: 'default' },
140 },
141 },
142 default: false,
143 },
144 'openai.gpt-oss-120b-1:0': {
145 default: true,
146 },
147 },
148 },
149 openai: {
150 models: {
151 'gpt-5.3-codex': { default: false },
152 'gpt-5.4-nano': { default: true },
153 },
154 providerOptions: {
155 openai: {
156 store: false,
157 },
158 },
159 },
160}
161
162export function getDefaultModelForProvider(provider: ProviderName): Model | undefined {
163 const models = PROVIDERS[provider]?.models as Record<Model, ProviderModelConfig>
164 if (!models) return undefined
165
166 return Object.keys(models).find((id) => models[id as Model]?.default) as Model | undefined
167}