@llumiverse/common 1.2.0 → 1.4.0-dev.20260629.090753Z

package/src/options/context-windows.ts

@@ -10,8 +10,8 @@ export function getMaxOutputTokens(model: string): number {
         if (model.includes('opus-4-6')) return 128_000;
         if (model.includes('opus-4-5')) return 64_000;
         if (model.includes('opus-')) return 32_768; // Opus 4.0, 4.1
-        if (model.includes('-4-')) return 64_000;   // Sonnet 4.x, Haiku 4.5
-        if (model.includes('-3-7')) return 64_000;  // 128K with beta header, default 64K
+        if (model.includes('-4-')) return 64_000; // Sonnet 4.x, Haiku 4.5
+        if (model.includes('-3-7')) return 64_000; // 128K with beta header, default 64K
         if (model.includes('-3-5')) return 8_192;
         return 4_096; // Claude 3 older
     }

package/src/options/embedding.ts

@@ -0,0 +1,92 @@
+import type { EmbeddingInput, Providers } from '../types.js';
+
+/** Default embedding model ids, one per provider. */
+
+// ── Vertex AI ──────────────────────────────────────────────────────────────
+/** Default Vertex AI text/multimodal embedding model (supports text + image/video/audio via embedContent). */
+export const VERTEX_DEFAULT_EMBEDDING_MODEL = 'gemini-embedding-2';
+/** Legacy Vertex AI multimodal predict API model (joint text+image+video in one vector). */
+export const VERTEX_MULTIMODAL_EMBEDDING_MODEL = 'multimodalembedding@001';
+
+// ── AWS Bedrock ────────────────────────────────────────────────────────────
+/** Default Bedrock embedding model — Nova 2 multimodal (text, image, video, audio). */
+export const BEDROCK_DEFAULT_EMBEDDING_MODEL = 'amazon.nova-2-multimodal-embeddings-v1:0';
+/** Bedrock Titan text embedding model (text only). */
+export const BEDROCK_TITAN_TEXT_EMBEDDING_MODEL = 'amazon.titan-embed-text-v2:0';
+/** Bedrock Titan image embedding model (text + image). */
+export const BEDROCK_TITAN_IMAGE_EMBEDDING_MODEL = 'amazon.titan-embed-image-v1';
+/** Bedrock Cohere English embedding model (text + image, native batch). */
+export const BEDROCK_COHERE_ENGLISH_EMBEDDING_MODEL = 'cohere.embed-english-v3';
+/** Bedrock Cohere multilingual embedding model (text + image, native batch). */
+export const BEDROCK_COHERE_MULTILINGUAL_EMBEDDING_MODEL = 'cohere.embed-multilingual-v3';
+
+// ── OpenAI ─────────────────────────────────────────────────────────────────
+/** Default OpenAI embedding model (text only, supports dimensions). */
+export const OPENAI_DEFAULT_EMBEDDING_MODEL = 'text-embedding-3-small';
+/** OpenAI large embedding model (text only, higher capacity). */
+export const OPENAI_LARGE_EMBEDDING_MODEL = 'text-embedding-3-large';
+
+// ── Mistral AI ─────────────────────────────────────────────────────────────
+/** Default Mistral AI embedding model (text only). */
+export const MISTRAL_DEFAULT_EMBEDDING_MODEL = 'mistral-embed';
+
+// ── IBM Watsonx ────────────────────────────────────────────────────────────
+/** Default Watsonx embedding model (text only). */
+export const WATSONX_DEFAULT_EMBEDDING_MODEL = 'ibm/slate-125m-english-rtrvr';
+
+// ── Azure Foundry ──────────────────────────────────────────────────────────
+/**
+ * Azure Foundry has no global default — the model is deployment-specific and
+ * must always be specified explicitly.
+ */
+export const AZURE_FOUNDRY_DEFAULT_EMBEDDING_MODEL: undefined = undefined;
+
+// ── Provider → default model map ───────────────────────────────────────────
+
+type EmbeddingModality = EmbeddingInput['type'];
+
+/**
+ * Per-provider default embedding model, keyed first by provider then by the
+ * primary modality of the input. When no modality-specific default exists for
+ * a provider, the `text` entry is used as the fallback.
+ *
+ * Providers that have a single multimodal model (Bedrock Nova, Vertex) return
+ * the same model for all modalities.
+ */
+const DEFAULT_EMBEDDING_MODELS: Partial<Record<Providers, Partial<Record<EmbeddingModality, string>>>> = {
+    vertexai: {
+        text: VERTEX_DEFAULT_EMBEDDING_MODEL,
+        image: VERTEX_DEFAULT_EMBEDDING_MODEL,
+        video: VERTEX_DEFAULT_EMBEDDING_MODEL,
+        audio: VERTEX_DEFAULT_EMBEDDING_MODEL,
+    },
+    bedrock: {
+        text: BEDROCK_DEFAULT_EMBEDDING_MODEL,
+        image: BEDROCK_DEFAULT_EMBEDDING_MODEL,
+        video: BEDROCK_DEFAULT_EMBEDDING_MODEL,
+        audio: BEDROCK_DEFAULT_EMBEDDING_MODEL,
+    },
+    openai: {
+        text: OPENAI_DEFAULT_EMBEDDING_MODEL,
+    },
+    mistralai: {
+        text: MISTRAL_DEFAULT_EMBEDDING_MODEL,
+    },
+    watsonx: {
+        text: WATSONX_DEFAULT_EMBEDDING_MODEL,
+    },
+};
+
+/**
+ * Returns the default embedding model for a provider and modality.
+ * Falls back to the provider's `text` default when no modality-specific entry exists.
+ * Returns `undefined` for providers without a global default (e.g. Azure Foundry).
+ */
+export function getDefaultEmbeddingModel(
+    provider: Providers,
+    modality: EmbeddingModality = 'text',
+): string | undefined {
+    const byProvider = DEFAULT_EMBEDDING_MODELS[provider];
+    if (!byProvider) return undefined;
+    return byProvider[modality] ?? byProvider.text;
+}

package/src/options/fallback.ts

@@ -1,7 +1,7 @@
-import { ModelOptionsInfo, OptionType, SharedOptions } from "../types.js";
+import { type ModelOptionsInfo, OptionType, SharedOptions } from '../types.js';
 
 export interface TextFallbackOptions {
-    _option_id: "text-fallback";    //For specific models should be format as "provider-model"
+    _option_id: 'text-fallback'; //For specific models should be format as "provider-model"
     max_tokens?: number;
     temperature?: number;
     top_p?: number;
@@ -12,32 +12,65 @@ export interface TextFallbackOptions {
 }
 
 export const textOptionsFallback: ModelOptionsInfo = {
-    _option_id: "text-fallback",
+    _option_id: 'text-fallback',
     options: [
         {
-            name: SharedOptions.max_tokens, type: OptionType.numeric, min: 1,
-            integer: true, step: 200, description: "The maximum number of tokens to generate"
+            name: SharedOptions.max_tokens,
+            type: OptionType.numeric,
+            min: 1,
+            integer: true,
+            step: 200,
+            description: 'The maximum number of tokens to generate',
         },
         {
-            name: SharedOptions.temperature, type: OptionType.numeric, min: 0.0, default: 0.7,
-            integer: false, step: 0.1, description: "A higher temperature biases toward less likely tokens, making the model more creative"
+            name: SharedOptions.temperature,
+            type: OptionType.numeric,
+            min: 0.0,
+            default: 0.7,
+            integer: false,
+            step: 0.1,
+            description: 'A higher temperature biases toward less likely tokens, making the model more creative',
         },
         {
-            name: SharedOptions.top_p, type: OptionType.numeric, min: 0, max: 1,
-            integer: false, step: 0.1, description: "Limits token sampling to the cumulative probability of the top p tokens"
+            name: SharedOptions.top_p,
+            type: OptionType.numeric,
+            min: 0,
+            max: 1,
+            integer: false,
+            step: 0.1,
+            description: 'Limits token sampling to the cumulative probability of the top p tokens',
         },
         {
-            name: SharedOptions.top_k, type: OptionType.numeric, min: 1,
-            integer: true, step: 1, description: "Limits token sampling to the top k tokens"
+            name: SharedOptions.top_k,
+            type: OptionType.numeric,
+            min: 1,
+            integer: true,
+            step: 1,
+            description: 'Limits token sampling to the top k tokens',
         },
         {
-            name: SharedOptions.presence_penalty, type: OptionType.numeric, min: -2.0, max: 2.0,
-            integer: false, step: 0.1, description: "Penalise tokens if they appear at least once in the text"
+            name: SharedOptions.presence_penalty,
+            type: OptionType.numeric,
+            min: -2.0,
+            max: 2.0,
+            integer: false,
+            step: 0.1,
+            description: 'Penalise tokens if they appear at least once in the text',
         },
         {
-            name: SharedOptions.frequency_penalty, type: OptionType.numeric, min: -2.0, max: 2.0,
-            integer: false, step: 0.1, description: "Penalise tokens based on their frequency in the text"
+            name: SharedOptions.frequency_penalty,
+            type: OptionType.numeric,
+            min: -2.0,
+            max: 2.0,
+            integer: false,
+            step: 0.1,
+            description: 'Penalise tokens based on their frequency in the text',
         },
-        { name: SharedOptions.stop_sequence, type: OptionType.string_list, value: [], description: "The generation will halt if one of the stop sequences is output" },
-    ]
+        {
+            name: SharedOptions.stop_sequence,
+            type: OptionType.string_list,
+            value: [],
+            description: 'The generation will halt if one of the stop sequences is output',
+        },
+    ],
 };

package/src/options/groq.ts

@@ -1,47 +1,74 @@
-import { ModelOptionsInfo, ModelOptionInfoItem, ModelOptions, OptionType, SharedOptions } from "../types.js";
-import { textOptionsFallback } from "./fallback.js";
+import {
+    type ModelOptionInfoItem,
+    type ModelOptions,
+    type ModelOptionsInfo,
+    OptionType,
+    SharedOptions,
+} from '../types.js';
+import { textOptionsFallback } from './fallback.js';
 
 // Union type of all Bedrock options
 export type GroqOptions = GroqDeepseekThinkingOptions;
 
 export interface GroqDeepseekThinkingOptions {
-    _option_id: "groq-deepseek-thinking",
-    max_tokens?: number,
-    temperature?: number,
-    top_p?: number,
-    stop_sequence?: string[],
-    reasoning_format: 'parsed' | 'raw' | 'hidden',
+    _option_id: 'groq-deepseek-thinking';
+    max_tokens?: number;
+    temperature?: number;
+    top_p?: number;
+    stop_sequence?: string[];
+    reasoning_format: 'parsed' | 'raw' | 'hidden';
 }
 
 export function getGroqOptions(model: string, _option?: ModelOptions): ModelOptionsInfo {
-    if (model.includes("deepseek") && model.includes("r1")) {
+    if (model.includes('deepseek') && model.includes('r1')) {
         const commonOptions: ModelOptionInfoItem[] = [
             {
-                name: SharedOptions.max_tokens, type: OptionType.numeric, min: 1, max: 131072,
-                integer: true, description: "The maximum number of tokens to generate",
+                name: SharedOptions.max_tokens,
+                type: OptionType.numeric,
+                min: 1,
+                max: 131072,
+                integer: true,
+                description: 'The maximum number of tokens to generate',
             },
             {
-                name: SharedOptions.temperature, type: OptionType.numeric, min: 0.0, default: 0.7, max: 2.0,
-                integer: false, step: 0.1, description: "A higher temperature biases toward less likely tokens, making the model more creative. A lower temperature than other models is recommended for deepseek R1, 0.3-0.7 approximately.",
+                name: SharedOptions.temperature,
+                type: OptionType.numeric,
+                min: 0.0,
+                default: 0.7,
+                max: 2.0,
+                integer: false,
+                step: 0.1,
+                description:
+                    'A higher temperature biases toward less likely tokens, making the model more creative. A lower temperature than other models is recommended for deepseek R1, 0.3-0.7 approximately.',
             },
             {
-                name: SharedOptions.top_p, type: OptionType.numeric, min: 0, max: 1,
-                integer: false, step: 0.1, description: "Limits token sampling to the cumulative probability of the top p tokens",
+                name: SharedOptions.top_p,
+                type: OptionType.numeric,
+                min: 0,
+                max: 1,
+                integer: false,
+                step: 0.1,
+                description: 'Limits token sampling to the cumulative probability of the top p tokens',
             },
             {
-                name: SharedOptions.stop_sequence, type: OptionType.string_list, value: [],
-                description: "The generation will halt if one of the stop sequences is output",
+                name: SharedOptions.stop_sequence,
+                type: OptionType.string_list,
+                value: [],
+                description: 'The generation will halt if one of the stop sequences is output',
             },
             {
-                name: "reasoning_format", type: OptionType.enum, enum: { "Parsed": "parsed", "Raw": "raw", "Hidden": "hidden" },
-                default: "parsed", description: "Controls how the reasoning is returned.",
+                name: 'reasoning_format',
+                type: OptionType.enum,
+                enum: { Parsed: 'parsed', Raw: 'raw', Hidden: 'hidden' },
+                default: 'parsed',
+                description: 'Controls how the reasoning is returned.',
             },
         ];
 
         return {
-            _option_id: "groq-deepseek-thinking",
+            _option_id: 'groq-deepseek-thinking',
             options: commonOptions,
         };
     }
     return textOptionsFallback;
-}
+}