sarvam-ai-sdk 0.0.5 → 0.0.7

package/README.md

@@ -69,9 +69,24 @@ const { text } = await transcribe({
 console.log(text); // പാചകം തുടരും സുഹൃത്തുക്കളെ
 ```
 
+## Speech-to-Text-Translate
+
+```ts
+import { sarvam } from "sarvam-ai-sdk";
+import { experimental_transcribe as transcribe } from "ai";
+import { readFile } from "fs/promises";
+
+const result = await transcribe({
+    model: sarvam.speechTranslation("saaras:v2"),
+    audio: await readFile("./src/transcript-test.wav"),
+});
+
+console.log(result.text); // Cooking continues, my friends
+```
+
 ## Translation
 
-> Only transliterates `prompt` and `role:user` messages, not `system` not `assistant`.
+> NB: Only transliterates `prompt` and `role:user` messages, not `system` not `assistant`.
 
 ```ts
 import { sarvam } from "sarvam-ai-sdk";
@@ -90,7 +105,7 @@ console.log(result.text); // Shouldn't we be careful about this, Ambane?
 
 ## Transliterate
 
-> Only transliterates `prompt` and `role:user` messages, not `system` not `assistant`.
+> NB: Only transliterates `prompt` and `role:user` messages, not `system` not `assistant`.
 
 ```ts
 import { sarvam } from "sarvam-ai-sdk";
@@ -107,6 +122,22 @@ const result = await generateText({
 console.log(result.text); // എടാ മോനെ, ഹാപ്പി അല്ലേ?
 ```
 
+## Language Identification
+
+> NB: Only identifies `prompt` and `role:user` messages, not `system` not `assistant`.
+
+```ts
+import { sarvam } from "sarvam-ai-sdk";
+import { generateText } from "ai";
+
+const result = await generateText({
+    model: sarvam.languageIdentification(),
+    prompt: "ബുദ്ധിയാണ് സാറേ ഇവൻ്റെ മെയിൻ",
+});
+
+console.log(result.text); // ml-IN
+```
+
 ## Tool Calling
 
 > [!WARNING]
@@ -120,7 +151,7 @@ import { sarvam } from "sarvam-ai-sdk";
 
 const result = await generateText({
   model: sarvam("sarvam-m", {
-    simulateToolCalling: true, // ⚠️ important
+    simulate: "tool-calling" // ⚠️ important
   }),
   tools: {
     weather: tool({
@@ -140,6 +171,32 @@ const result = await generateText({
 
 console.log(result.toolResults);
 ```
+## Generate JSON object
+
+> [!WARNING]
+> Latest `sarvam-m` model isn't trained on native JSON object generation. So we simulate this with prompt engineering technique.
+
+```ts
+import { z } from "zod";
+import { sarvam } from "sarvam-ai-sdk";
+import { generateObject } from 'ai';
+
+const { object } = await generateObject({
+  model: sarvam("sarvam-m", {
+    simulate: "json-object" // ⚠️ important
+  }),
+  schema: z.object({
+    recipe: z.object({
+      name: z.string(),
+      ingredients: z.array(z.string()),
+      steps: z.array(z.string()),
+    }),
+  }),
+  prompt: 'Generate a South Indian recipe, in Malayalam',
+});
+
+console.log(object);
+```
 
 ## Documentation
 

package/dist/index.cjs

@@ -55,7 +55,7 @@ __export(index_exports, {
 module.exports = __toCommonJS(index_exports);
 
 // src/sarvam-provider.ts
-var import_provider_utils9 = require("@ai-sdk/provider-utils");
+var import_provider_utils11 = require("@ai-sdk/provider-utils");
 
 // src/sarvam-chat-language-model.ts
 var import_provider3 = require("@ai-sdk/provider");
@@ -300,24 +300,29 @@ const myChoice: YourToolChoices = {
 }`;
   return text;
 };
-var extractToolCallData = (text) => {
+var extractToolCallData = (jsonObject) => {
+  const toolFunction = jsonObject;
+  if (!("toolName" in toolFunction)) return;
+  if (!("toolData" in toolFunction)) return;
+  return {
+    args: JSON.stringify(toolFunction.toolData),
+    toolCallId: (0, import_provider_utils3.generateId)(),
+    toolCallType: "function",
+    toolName: toolFunction.toolName
+  };
+};
+var parseJSON = (text) => {
   const jsonRegex = /\{(?:[^{}]*|\{[^{}]*\})*\}/g;
   const jsonMatches = text.match(jsonRegex);
   if (jsonMatches && jsonMatches[0]) {
     try {
-      const toolFunction = JSON.parse(jsonMatches[0]);
-      if (!("toolName" in toolFunction)) return;
-      if (!("toolData" in toolFunction)) return;
-      return {
-        args: JSON.stringify(toolFunction.toolData),
-        toolCallId: (0, import_provider_utils3.generateId)(),
-        toolCallType: "function",
-        toolName: toolFunction.toolName
-      };
+      const jsonObject = JSON.parse(jsonMatches[0]);
+      return jsonObject;
     } catch (error) {
     }
   }
 };
+var simulateJsonSchema = () => "If user doen't specify, make sure to translate json data content into pure English.";
 
 // src/sarvam-chat-language-model.ts
 var SarvamChatLanguageModel = class {
@@ -351,7 +356,18 @@ var SarvamChatLanguageModel = class {
     providerMetadata
   }) {
     const type = mode.type;
+    const simulate = this.settings.simulate;
+    if (type === "object-json" && simulate === "tool-calling")
+      throw new Error('Use { simulate: "json-object" } with generateObject()');
+    if (type === "regular" && simulate === "json-object")
+      throw new Error('Use { simulate: "tool-calling" } with generateText()');
     const warnings = [];
+    if (stream) {
+      warnings.push({
+        type: "other",
+        message: "Streaming is still experimental for Sarvam"
+      });
+    }
     if (topK != null) {
       warnings.push({
         type: "unsupported-setting",
@@ -372,7 +388,7 @@ var SarvamChatLanguageModel = class {
         reasoningFormat: import_zod2.z.enum(["parsed", "raw", "hidden"]).nullish()
       })
     });
-    const baseArgs = (prompt2, fakeToolSystemPrompt) => ({
+    const baseArgs = (prompt2, extraSystemPrompt) => ({
       // model id:
       model: this.modelId,
       // model specific settings:
@@ -394,16 +410,16 @@ var SarvamChatLanguageModel = class {
       // provider options:
       reasoning_format: sarvamOptions == null ? void 0 : sarvamOptions.reasoningFormat,
       // messages:
-      messages: convertToSarvamChatMessages(prompt2, fakeToolSystemPrompt)
+      messages: convertToSarvamChatMessages(prompt2, extraSystemPrompt)
     });
     switch (type) {
       case "regular": {
         const { tools, tool_choice, toolWarnings } = prepareTools({
           mode
         });
-        const fakeSystemPrompt = tools && this.settings.simulateToolCalling ? await simulateToolCalling(tools) : void 0;
+        const extraSystemPrompt = tools && simulate === "tool-calling" ? await simulateToolCalling(tools) : void 0;
         return {
-          args: __spreadProps(__spreadValues({}, baseArgs(prompt, fakeSystemPrompt)), {
+          args: __spreadProps(__spreadValues({}, baseArgs(prompt, extraSystemPrompt)), {
             tools,
             tool_choice
           }),
@@ -411,8 +427,9 @@ var SarvamChatLanguageModel = class {
         };
       }
       case "object-json": {
+        const extraSystemPrompt = simulate === "json-object" ? simulateJsonSchema() : void 0;
         return {
-          args: __spreadProps(__spreadValues({}, baseArgs(prompt)), {
+          args: __spreadProps(__spreadValues({}, baseArgs(prompt, extraSystemPrompt)), {
             response_format: (
               // json object response format is not supported for streaming:
               stream === false ? { type: "json_object" } : void 0
@@ -484,12 +501,24 @@ var SarvamChatLanguageModel = class {
         args: toolCall.function.arguments
       };
     });
-    if (this.settings.simulateToolCalling) {
-      if (text && text.length !== 0 && (!toolCalls || (toolCalls == null ? void 0 : toolCalls.length) === 0)) {
-        const newTools = extractToolCallData(text);
-        if (newTools) {
-          toolCalls = [newTools];
-          text = void 0;
+    if (this.settings.simulate === "tool-calling") {
+      if (text && text.length !== 0) {
+        const jsonObject = parseJSON(text);
+        if (jsonObject) {
+          const newTools = extractToolCallData(jsonObject);
+          if (newTools) {
+            toolCalls = [newTools];
+            text = void 0;
+          }
+        }
+      }
+    }
+    if (this.settings.simulate === "json-object") {
+      if (text && text.length !== 0) {
+        const jsonObject = parseJSON(text);
+        if (jsonObject) {
+          const newTools = extractToolCallData(jsonObject);
+          text = JSON.stringify(jsonObject);
         }
       }
     }
@@ -1064,6 +1093,18 @@ var SarvamLanguageCodeSchema = import_zod7.z.enum([
   "en-IN",
   "gu-IN"
 ]);
+var SarvamScriptCodeSchema = import_zod7.z.enum([
+  "Latn",
+  "Deva",
+  "Beng",
+  "Gujr",
+  "Knda",
+  "Mlym",
+  "Orya",
+  "Guru",
+  "Taml",
+  "Telu"
+]);
 
 // src/sarvam-translation-model.ts
 var SarvamTranslationModel = class {
@@ -1267,11 +1308,185 @@ var sarvamTransliterateResponseSchema = import_zod9.z.object({
   request_id: import_zod9.z.string().nullish()
 });
 
+// src/sarvam-lid-model.ts
+var import_provider_utils9 = require("@ai-sdk/provider-utils");
+var import_zod10 = require("zod");
+var SarvamLidModel = class {
+  constructor(config) {
+    this.specificationVersion = "v1";
+    this.supportsStructuredOutputs = false;
+    this.defaultObjectGenerationMode = "json";
+    this.modelId = "unknown";
+    this.config = config;
+  }
+  get provider() {
+    return this.config.provider;
+  }
+  get supportsImageUrls() {
+    return false;
+  }
+  getArgs({
+    mode,
+    prompt
+  }) {
+    const type = mode.type;
+    const warnings = [];
+    if (type !== "regular") {
+      const _exhaustiveCheck = type;
+      throw new Error(`Unsupported type: ${_exhaustiveCheck}`);
+    }
+    const messages = convertToSarvamChatMessages(prompt);
+    return {
+      messages,
+      args: {
+        input: messages.filter((m) => m.role === "user").map((m) => m.content).join("\n")
+      },
+      warnings
+    };
+  }
+  async doGenerate(options) {
+    var _b;
+    const { args, warnings, messages } = this.getArgs(__spreadProps(__spreadValues({}, options), {
+      stream: false
+    }));
+    const body = JSON.stringify(args);
+    const {
+      responseHeaders,
+      value: response,
+      rawValue: rawResponse
+    } = await (0, import_provider_utils9.postJsonToApi)({
+      url: this.config.url({
+        path: "/text-lid"
+      }),
+      headers: (0, import_provider_utils9.combineHeaders)(this.config.headers(), options.headers),
+      body: args,
+      failedResponseHandler: sarvamFailedResponseHandler,
+      successfulResponseHandler: (0, import_provider_utils9.createJsonResponseHandler)(
+        sarvamLidResponseSchema
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch
+    });
+    const _a = args, { input: rawPrompt } = _a, rawSettings = __objRest(_a, ["input"]);
+    const text = (_b = response.language_code) != null ? _b : void 0;
+    return {
+      text,
+      toolCalls: void 0,
+      reasoning: void 0,
+      finishReason: "unknown",
+      usage: {
+        promptTokens: NaN,
+        completionTokens: NaN
+      },
+      rawCall: { rawPrompt, rawSettings },
+      rawResponse: { headers: responseHeaders, body: rawResponse },
+      response: void 0,
+      warnings,
+      request: { body }
+    };
+  }
+  async doStream(options) {
+    throw new Error("Language Identification feature doesn't streaming yet");
+  }
+};
+var sarvamLidResponseSchema = import_zod10.z.object({
+  script_code: SarvamScriptCodeSchema.nullish(),
+  language_code: SarvamLanguageCodeSchema.nullable(),
+  request_id: import_zod10.z.string().nullish()
+});
+
+// src/sarvam-speech-translation-model.ts
+var import_provider_utils10 = require("@ai-sdk/provider-utils");
+var import_zod11 = require("zod");
+var SarvamSpeechTranslationModel = class {
+  constructor(modelId, config) {
+    this.modelId = modelId;
+    this.config = config;
+    this.specificationVersion = "v1";
+  }
+  get provider() {
+    return this.config.provider;
+  }
+  getArgs({
+    audio,
+    mediaType,
+    providerOptions
+  }) {
+    const warnings = [];
+    const formData = new FormData();
+    const blob = audio instanceof Blob ? audio : new Blob([audio], { type: mediaType });
+    formData.append("file", blob);
+    formData.append("model", this.modelId);
+    return {
+      formData,
+      warnings
+    };
+  }
+  async doGenerate(options) {
+    var _a, _b, _c;
+    const currentDate = (_c = (_b = (_a = this.config._internal) == null ? void 0 : _a.currentDate) == null ? void 0 : _b.call(_a)) != null ? _c : /* @__PURE__ */ new Date();
+    const { formData, warnings } = this.getArgs(options);
+    const {
+      value: response,
+      responseHeaders,
+      rawValue: rawResponse
+    } = await (0, import_provider_utils10.postFormDataToApi)({
+      url: this.config.url({
+        path: "/speech-to-text-translate",
+        modelId: this.modelId
+      }),
+      headers: (0, import_provider_utils10.combineHeaders)(this.config.headers(), options.headers),
+      formData,
+      failedResponseHandler: sarvamFailedResponseHandler,
+      successfulResponseHandler: (0, import_provider_utils10.createJsonResponseHandler)(
+        sarvamTranscriptionResponseSchema2
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch
+    });
+    return {
+      text: response.transcript,
+      segments: [],
+      language: response.language_code ? response.language_code : void 0,
+      durationInSeconds: void 0,
+      warnings,
+      response: {
+        timestamp: currentDate,
+        modelId: this.modelId,
+        headers: responseHeaders,
+        body: rawResponse
+      }
+    };
+  }
+};
+var sarvamTranscriptionResponseSchema2 = import_zod11.z.object({
+  request_id: import_zod11.z.string().nullable(),
+  transcript: import_zod11.z.string(),
+  language_code: import_zod11.z.string().nullable(),
+  // timestamps: z
+  //   .object({
+  //     end_time_seconds: z.array(z.number()),
+  //     start_time_seconds: z.array(z.number()),
+  //     words: z.array(z.string()),
+  //   })
+  //   .optional(),
+  diarized_transcript: import_zod11.z.object({
+    entries: import_zod11.z.array(
+      import_zod11.z.object({
+        end_time_seconds: import_zod11.z.array(import_zod11.z.number()),
+        start_time_seconds: import_zod11.z.array(import_zod11.z.number()),
+        transcript: import_zod11.z.string(),
+        speaker_id: import_zod11.z.string()
+      })
+    )
+  }).nullable().optional()
+});
+
 // src/sarvam-provider.ts
 function createSarvam(options = {}) {
   var _a;
-  const baseURL = (_a = (0, import_provider_utils9.withoutTrailingSlash)(options.baseURL)) != null ? _a : "https://api.sarvam.ai";
-  const ApiKey = (0, import_provider_utils9.loadApiKey)({
+  const baseURL = (_a = (0, import_provider_utils11.withoutTrailingSlash)(options.baseURL)) != null ? _a : "https://api.sarvam.ai";
+  const ApiKey = (0, import_provider_utils11.loadApiKey)({
     apiKey: options.apiKey,
     environmentVariableName: "SARVAM_API_KEY",
     description: "Sarvam"
@@ -1294,15 +1509,19 @@ function createSarvam(options = {}) {
     }
     return createChatModel(modelId, settings);
   };
-  const createTranscriptionModel = (modelId, languageCode = "unknown", settings) => {
-    return new SarvamTranscriptionModel(modelId, languageCode, {
-      provider: "sarvam.transcription",
-      url: ({ path }) => `${baseURL}${path}`,
-      headers: getHeaders,
-      fetch: options.fetch,
-      transcription: settings
-    });
-  };
+  const createTranscriptionModel = (modelId, languageCode = "unknown", settings) => new SarvamTranscriptionModel(modelId, languageCode, {
+    provider: "sarvam.transcription",
+    url: ({ path }) => `${baseURL}${path}`,
+    headers: getHeaders,
+    fetch: options.fetch,
+    transcription: settings
+  });
+  const createSpeechTranslation = (modelId) => new SarvamSpeechTranslationModel(modelId, {
+    provider: "sarvam.transcription",
+    url: ({ path }) => `${baseURL}${path}`,
+    headers: getHeaders,
+    fetch: options.fetch
+  });
   const createSpeechModel = (modelId, languageCode, settings) => new SarvamSpeechModel(modelId, languageCode, {
     provider: "sarvam.speech",
     url: ({ path }) => `${baseURL}${path}`,
@@ -1334,13 +1553,23 @@ function createSarvam(options = {}) {
       fetch: options.fetch
     }
   );
+  const createLidModel = () => new SarvamLidModel(
+    {
+      provider: "sarvam.lid",
+      url: ({ path }) => `${baseURL}${path}`,
+      headers: getHeaders,
+      fetch: options.fetch
+    }
+  );
   const provider = (modelId, settings) => createLanguageModel(modelId, settings);
   provider.languageModel = createLanguageModel;
   provider.chat = createChatModel;
   provider.transcription = createTranscriptionModel;
+  provider.speechTranslation = createSpeechTranslation;
   provider.speech = createSpeechModel;
   provider.transliterate = createTransliterateModel;
   provider.translation = createTranslationModel;
+  provider.languageIdentification = createLidModel;
   return provider;
 }
 var sarvam = createSarvam();