modelfusion 0.1.1 → 0.3.0

package/README.md

@@ -377,6 +377,22 @@ const { chunks } = await retrieveTextChunks(
 - [Memory](https://modelfusion.dev/integration/vector-index/memory)
 - [Pinecone](https://modelfusion.dev/integration/vector-index/pinecone)
 
+### Observability
+
+- [Helicone](https://modelfusion.dev/integration/observability/helicone)
+
+### Prompt Formats
+
+Use higher level prompts that are mapped into model specific prompt formats.
+
+| Prompt Format | Instruction Prompt | Chat Prompt |
+| ------------- | ------------------ | ----------- |
+| OpenAI Chat   | ✅                 | ✅          |
+| Llama 2       | ✅                 | ✅          |
+| Alpaca        | ✅                 | ❌          |
+| Vicuna        | ❌                 | ✅          |
+| Generic Text  | ✅                 | ✅          |
+
 ## Documentation
 
 - [Guide](https://modelfusion.dev/guide)

package/model-provider/cohere/CohereError.cjs

@@ -25,6 +25,17 @@ class CohereError extends ApiCallError_js_1.ApiCallError {
 exports.CohereError = CohereError;
 const failedCohereCallResponseHandler = async ({ response, url, requestBodyValues }) => {
     const responseBody = await response.text();
+    // For some errors, the body of Cohere responses is empty:
+    if (responseBody.trim() === "") {
+        return new CohereError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            data: {
+                message: response.statusText,
+            },
+        });
+    }
     const parsedError = exports.cohereErrorDataSchema.parse(secure_json_parse_1.default.parse(responseBody));
     return new CohereError({
         url,

package/model-provider/cohere/CohereError.js

@@ -18,6 +18,17 @@ export class CohereError extends ApiCallError {
 }
 export const failedCohereCallResponseHandler = async ({ response, url, requestBodyValues }) => {
     const responseBody = await response.text();
+    // For some errors, the body of Cohere responses is empty:
+    if (responseBody.trim() === "") {
+        return new CohereError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            data: {
+                message: response.statusText,
+            },
+        });
+    }
     const parsedError = cohereErrorDataSchema.parse(SecureJSON.parse(responseBody));
     return new CohereError({
         url,

package/model-provider/cohere/CohereTextEmbeddingModel.cjs

@@ -159,7 +159,9 @@ const cohereTextEmbeddingResponseSchema = zod_1.default.object({
 async function callCohereEmbeddingAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, texts, truncate, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/embed`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             texts,

package/model-provider/cohere/CohereTextEmbeddingModel.js

@@ -152,7 +152,9 @@ const cohereTextEmbeddingResponseSchema = z.object({
 async function callCohereEmbeddingAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, texts, truncate, }) {
     return postJsonToApi({
         url: `${baseUrl}/embed`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             texts,

package/model-provider/cohere/CohereTextGenerationModel.cjs

@@ -179,7 +179,9 @@ const cohereTextGenerationResponseSchema = zod_1.z.object({
 async function callCohereTextGenerationAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, responseFormat, apiKey, model, prompt, numGenerations, maxTokens, temperature, k, p, frequencyPenalty, presencePenalty, endSequences, stopSequences, returnLikelihoods, logitBias, truncate, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/generate`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/cohere/CohereTextGenerationModel.js

@@ -172,7 +172,9 @@ const cohereTextGenerationResponseSchema = z.object({
 async function callCohereTextGenerationAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, responseFormat, apiKey, model, prompt, numGenerations, maxTokens, temperature, k, p, frequencyPenalty, presencePenalty, endSequences, stopSequences, returnLikelihoods, logitBias, truncate, }) {
     return postJsonToApi({
         url: `${baseUrl}/generate`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/cohere/CohereTokenizer.cjs

@@ -97,7 +97,9 @@ const cohereDetokenizationResponseSchema = zod_1.default.object({
 async function callCohereDetokenizeAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, tokens, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/detokenize`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             tokens,
@@ -124,7 +126,9 @@ const cohereTokenizationResponseSchema = zod_1.default.object({
 async function callCohereTokenizeAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, text, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/tokenize`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             text,

package/model-provider/cohere/CohereTokenizer.js

@@ -90,7 +90,9 @@ const cohereDetokenizationResponseSchema = z.object({
 async function callCohereDetokenizeAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, tokens, }) {
     return postJsonToApi({
         url: `${baseUrl}/detokenize`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             tokens,
@@ -117,7 +119,9 @@ const cohereTokenizationResponseSchema = z.object({
 async function callCohereTokenizeAPI({ baseUrl = "https://api.cohere.ai/v1", abortSignal, apiKey, model, text, }) {
     return postJsonToApi({
         url: `${baseUrl}/tokenize`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             text,

package/model-provider/huggingface/HuggingFaceTextGenerationModel.cjs

@@ -149,7 +149,9 @@ const huggingFaceTextGenerationResponseSchema = zod_1.default.array(zod_1.defaul
 async function callHuggingFaceTextGenerationAPI({ baseUrl = "https://api-inference.huggingface.co/models", abortSignal, apiKey, model, inputs, topK, topP, temperature, repetitionPenalty, maxNewTokens, maxTime, numReturnSequences, doSample, options, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/${model}`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             inputs,
             top_k: topK,

package/model-provider/huggingface/HuggingFaceTextGenerationModel.js

@@ -142,7 +142,9 @@ const huggingFaceTextGenerationResponseSchema = z.array(z.object({
 async function callHuggingFaceTextGenerationAPI({ baseUrl = "https://api-inference.huggingface.co/models", abortSignal, apiKey, model, inputs, topK, topP, temperature, repetitionPenalty, maxNewTokens, maxTime, numReturnSequences, doSample, options, }) {
     return postJsonToApi({
         url: `${baseUrl}/${model}`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             inputs,
             top_k: topK,

package/model-provider/openai/OpenAIError.cjs

@@ -16,7 +16,7 @@ exports.openAIErrorDataSchema = zod_1.z.object({
     }),
 });
 class OpenAIError extends ApiCallError_js_1.ApiCallError {
-    constructor({ data, statusCode, url, requestBodyValues, message = data.error.message, }) {
+    constructor({ data, statusCode, url, requestBodyValues, message, }) {
         super({
             message,
             statusCode,
@@ -24,7 +24,7 @@ class OpenAIError extends ApiCallError_js_1.ApiCallError {
             url,
             isRetryable: (statusCode === 429 &&
                 // insufficient_quota is also reported as a 429, but it's not retryable:
-                data.error.type !== "insufficient_quota") ||
+                data?.error.type !== "insufficient_quota") ||
                 statusCode >= 500,
         });
         Object.defineProperty(this, "data", {
@@ -39,12 +39,24 @@ class OpenAIError extends ApiCallError_js_1.ApiCallError {
 exports.OpenAIError = OpenAIError;
 const failedOpenAICallResponseHandler = async ({ response, url, requestBodyValues }) => {
     const responseBody = await response.text();
-    const parsedError = exports.openAIErrorDataSchema.parse(secure_json_parse_1.default.parse(responseBody));
-    return new OpenAIError({
-        url,
-        requestBodyValues,
-        statusCode: response.status,
-        data: parsedError,
-    });
+    // resilient parsing in case the response is not JSON or does not match the schema:
+    try {
+        const parsedError = exports.openAIErrorDataSchema.parse(secure_json_parse_1.default.parse(responseBody));
+        return new OpenAIError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            message: parsedError.error.message,
+            data: parsedError,
+        });
+    }
+    catch (parseError) {
+        return new OpenAIError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            message: responseBody.trim() !== "" ? responseBody : response.statusText,
+        });
+    }
 };
 exports.failedOpenAICallResponseHandler = failedOpenAICallResponseHandler;

package/model-provider/openai/OpenAIError.d.ts

@@ -35,13 +35,13 @@ export declare const openAIErrorDataSchema: z.ZodObject<{
 }>;
 export type OpenAIErrorData = z.infer<typeof openAIErrorDataSchema>;
 export declare class OpenAIError extends ApiCallError {
-    readonly data: OpenAIErrorData;
+    readonly data?: OpenAIErrorData;
     constructor({ data, statusCode, url, requestBodyValues, message, }: {
-        message?: string;
+        message: string;
         statusCode: number;
         url: string;
         requestBodyValues: unknown;
-        data: OpenAIErrorData;
+        data?: OpenAIErrorData;
     });
 }
 export declare const failedOpenAICallResponseHandler: ResponseHandler<ApiCallError>;

package/model-provider/openai/OpenAIError.js

@@ -10,7 +10,7 @@ export const openAIErrorDataSchema = z.object({
     }),
 });
 export class OpenAIError extends ApiCallError {
-    constructor({ data, statusCode, url, requestBodyValues, message = data.error.message, }) {
+    constructor({ data, statusCode, url, requestBodyValues, message, }) {
         super({
             message,
             statusCode,
@@ -18,7 +18,7 @@ export class OpenAIError extends ApiCallError {
             url,
             isRetryable: (statusCode === 429 &&
                 // insufficient_quota is also reported as a 429, but it's not retryable:
-                data.error.type !== "insufficient_quota") ||
+                data?.error.type !== "insufficient_quota") ||
                 statusCode >= 500,
         });
         Object.defineProperty(this, "data", {
@@ -32,11 +32,23 @@ export class OpenAIError extends ApiCallError {
 }
 export const failedOpenAICallResponseHandler = async ({ response, url, requestBodyValues }) => {
     const responseBody = await response.text();
-    const parsedError = openAIErrorDataSchema.parse(SecureJSON.parse(responseBody));
-    return new OpenAIError({
-        url,
-        requestBodyValues,
-        statusCode: response.status,
-        data: parsedError,
-    });
+    // resilient parsing in case the response is not JSON or does not match the schema:
+    try {
+        const parsedError = openAIErrorDataSchema.parse(SecureJSON.parse(responseBody));
+        return new OpenAIError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            message: parsedError.error.message,
+            data: parsedError,
+        });
+    }
+    catch (parseError) {
+        return new OpenAIError({
+            url,
+            requestBodyValues,
+            statusCode: response.status,
+            message: responseBody.trim() !== "" ? responseBody : response.statusText,
+        });
+    }
 };

package/model-provider/openai/OpenAIImageGenerationModel.cjs

@@ -109,7 +109,9 @@ exports.OpenAIImageGenerationResponseFormat = {
 async function callOpenAIImageGenerationAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, apiKey, prompt, n, size, responseFormat, user, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/images/generations`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             prompt,
             n,

package/model-provider/openai/OpenAIImageGenerationModel.js

@@ -104,7 +104,9 @@ export const OpenAIImageGenerationResponseFormat = {
 async function callOpenAIImageGenerationAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, apiKey, prompt, n, size, responseFormat, user, }) {
     return postJsonToApi({
         url: `${baseUrl}/images/generations`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             prompt,
             n,

package/model-provider/openai/OpenAITextEmbeddingModel.cjs

@@ -158,7 +158,9 @@ const openAITextEmbeddingResponseSchema = zod_1.default.object({
 async function callOpenAITextEmbeddingAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, apiKey, model, input, user, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/embeddings`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             input,

package/model-provider/openai/OpenAITextEmbeddingModel.js

@@ -149,7 +149,9 @@ const openAITextEmbeddingResponseSchema = z.object({
 async function callOpenAITextEmbeddingAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, apiKey, model, input, user, }) {
     return postJsonToApi({
         url: `${baseUrl}/embeddings`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             model,
             input,

package/model-provider/openai/OpenAITextGenerationModel.cjs

@@ -209,10 +209,13 @@ const openAITextGenerationResponseSchema = zod_1.default.object({
  *
  * console.log(response.choices[0].text);
  */
-async function callOpenAITextGenerationAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, responseFormat, apiKey, model, prompt, suffix, maxTokens, temperature, topP, n, logprobs, echo, stop, presencePenalty, frequencyPenalty, bestOf, user, }) {
+async function callOpenAITextGenerationAPI({ baseUrl = "https://api.openai.com/v1", headers, abortSignal, responseFormat, apiKey, model, prompt, suffix, maxTokens, temperature, topP, n, logprobs, echo, stop, presencePenalty, frequencyPenalty, bestOf, user, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/completions`,
-        apiKey,
+        headers: {
+            ...headers,
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/openai/OpenAITextGenerationModel.d.ts

@@ -65,6 +65,7 @@ export declare const calculateOpenAITextGenerationCostInMillicents: ({ model, re
 }) => number;
 export interface OpenAITextGenerationModelSettings extends TextGenerationModelSettings {
     model: OpenAITextGenerationModelType;
+    headers?: Record<string, string>;
     baseUrl?: string;
     apiKey?: string;
     retry?: RetryFunction;

package/model-provider/openai/OpenAITextGenerationModel.js

@@ -200,10 +200,13 @@ const openAITextGenerationResponseSchema = z.object({
  *
  * console.log(response.choices[0].text);
  */
-async function callOpenAITextGenerationAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, responseFormat, apiKey, model, prompt, suffix, maxTokens, temperature, topP, n, logprobs, echo, stop, presencePenalty, frequencyPenalty, bestOf, user, }) {
+async function callOpenAITextGenerationAPI({ baseUrl = "https://api.openai.com/v1", headers, abortSignal, responseFormat, apiKey, model, prompt, suffix, maxTokens, temperature, topP, n, logprobs, echo, stop, presencePenalty, frequencyPenalty, bestOf, user, }) {
     return postJsonToApi({
         url: `${baseUrl}/completions`,
-        apiKey,
+        headers: {
+            ...headers,
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/openai/OpenAITranscriptionModel.cjs

@@ -132,8 +132,9 @@ async function callOpenAITranscriptionAPI({ baseUrl = "https://api.openai.com/v1
     }
     return (0, postToApi_js_1.postToApi)({
         url: `${baseUrl}/audio/transcriptions`,
-        apiKey,
-        contentType: null,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             content: formData,
             values: {

package/model-provider/openai/OpenAITranscriptionModel.js

@@ -124,8 +124,9 @@ async function callOpenAITranscriptionAPI({ baseUrl = "https://api.openai.com/v1
     }
     return postToApi({
         url: `${baseUrl}/audio/transcriptions`,
-        apiKey,
-        contentType: null,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             content: formData,
             values: {

package/model-provider/openai/chat/OpenAIChatModel.cjs

@@ -246,10 +246,13 @@ const openAIChatResponseSchema = zod_1.default.object({
         total_tokens: zod_1.default.number(),
     }),
 });
-async function callOpenAIChatCompletionAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, responseFormat, apiKey, model, messages, functions, functionCall, temperature, topP, n, stop, maxTokens, presencePenalty, frequencyPenalty, user, }) {
+async function callOpenAIChatCompletionAPI({ baseUrl = "https://api.openai.com/v1", headers, abortSignal, responseFormat, apiKey, model, messages, functions, functionCall, temperature, topP, n, stop, maxTokens, presencePenalty, frequencyPenalty, user, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/chat/completions`,
-        apiKey,
+        headers: {
+            ...headers,
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/openai/chat/OpenAIChatModel.d.ts

@@ -77,6 +77,7 @@ export declare const calculateOpenAIChatCostInMillicents: ({ model, response, }:
 }) => number;
 export interface OpenAIChatCallSettings {
     model: OpenAIChatModelType;
+    headers?: Record<string, string>;
     functions?: Array<{
         name: string;
         description?: string;

package/model-provider/openai/chat/OpenAIChatModel.js

@@ -237,10 +237,13 @@ const openAIChatResponseSchema = z.object({
         total_tokens: z.number(),
     }),
 });
-async function callOpenAIChatCompletionAPI({ baseUrl = "https://api.openai.com/v1", abortSignal, responseFormat, apiKey, model, messages, functions, functionCall, temperature, topP, n, stop, maxTokens, presencePenalty, frequencyPenalty, user, }) {
+async function callOpenAIChatCompletionAPI({ baseUrl = "https://api.openai.com/v1", headers, abortSignal, responseFormat, apiKey, model, messages, functions, functionCall, temperature, topP, n, stop, maxTokens, presencePenalty, frequencyPenalty, user, }) {
     return postJsonToApi({
         url: `${baseUrl}/chat/completions`,
-        apiKey,
+        headers: {
+            ...headers,
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             stream: responseFormat.stream,
             model,

package/model-provider/stability/StabilityImageGenerationModel.cjs

@@ -113,7 +113,9 @@ const stabilityImageGenerationResponseSchema = zod_1.z.object({
 async function callStabilityImageGenerationAPI({ baseUrl = "https://api.stability.ai/v1", abortSignal, apiKey, engineId, height, width, textPrompts, cfgScale, clipGuidancePreset, sampler, samples, seed, steps, stylePreset, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: `${baseUrl}/generation/${engineId}/text-to-image`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             height,
             width,

package/model-provider/stability/StabilityImageGenerationModel.js

@@ -109,7 +109,9 @@ const stabilityImageGenerationResponseSchema = z.object({
 async function callStabilityImageGenerationAPI({ baseUrl = "https://api.stability.ai/v1", abortSignal, apiKey, engineId, height, width, textPrompts, cfgScale, clipGuidancePreset, sampler, samples, seed, steps, stylePreset, }) {
     return postJsonToApi({
         url: `${baseUrl}/generation/${engineId}/text-to-image`,
-        apiKey,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
         body: {
             height,
             width,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "modelfusion",
   "description": "Build AI applications, chatbots, and agents with JavaScript and TypeScript.",
-  "version": "0.1.1",
+  "version": "0.3.0",
   "author": "Lars Grammel",
   "license": "MIT",
   "keywords": [
@@ -65,10 +65,10 @@
     "@typescript-eslint/parser": "^6.1.0",
     "copyfiles": "2.4.1",
     "eslint": "^8.45.0",
-    "eslint-config-prettier": "8.9.0",
+    "eslint-config-prettier": "8.10.0",
     "husky": "^8.0.3",
     "lint-staged": "13.2.3",
-    "prettier": "3.0.0",
+    "prettier": "3.0.1",
     "rimraf": "5.0.1",
     "typescript": "5.1.6",
     "zod": "3.21.4",

package/prompt/AlpacaPromptMapping.cjs

@@ -0,0 +1,33 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.InstructionToAlpacaPromptMapping = void 0;
+const DEFAULT_SYSTEM_PROMPT_INPUT = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.";
+const DEFAULT_SYSTEM_PROMPT_NO_INPUT = "Below is an instruction that describes a task. Write a response that appropriately completes the request.";
+/**
+ * Maps an instruction prompt to the Alpaca prompt format.
+ *
+ * If the instruction has a system prompt, it overrides the default system prompt
+ * (which can impact the results, because the model may be trained on the default system prompt).
+ *
+ * @see https://github.com/tatsu-lab/stanford_alpaca#data-release
+ */
+const InstructionToAlpacaPromptMapping = () => ({
+    stopTokens: [],
+    map: (instruction) => {
+        let text = instruction.system ??
+            (instruction.input != null
+                ? DEFAULT_SYSTEM_PROMPT_INPUT
+                : DEFAULT_SYSTEM_PROMPT_NO_INPUT);
+        text += "\n\n### Instruction:\n";
+        if (instruction.system != null) {
+            text += `${instruction.system}\n`;
+        }
+        text += instruction.instruction;
+        if (instruction.input != null) {
+            text += `\n\n### Input:\n${instruction.input}`;
+        }
+        text += "\n\n### Response:\n";
+        return text;
+    },
+});
+exports.InstructionToAlpacaPromptMapping = InstructionToAlpacaPromptMapping;

package/prompt/AlpacaPromptMapping.d.ts

@@ -0,0 +1,11 @@
+import { InstructionPrompt } from "./InstructionPrompt.js";
+import { PromptMapping } from "./PromptMapping.js";
+/**
+ * Maps an instruction prompt to the Alpaca prompt format.
+ *
+ * If the instruction has a system prompt, it overrides the default system prompt
+ * (which can impact the results, because the model may be trained on the default system prompt).
+ *
+ * @see https://github.com/tatsu-lab/stanford_alpaca#data-release
+ */
+export declare const InstructionToAlpacaPromptMapping: () => PromptMapping<InstructionPrompt, string>;

package/prompt/AlpacaPromptMapping.js

@@ -0,0 +1,29 @@
+const DEFAULT_SYSTEM_PROMPT_INPUT = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.";
+const DEFAULT_SYSTEM_PROMPT_NO_INPUT = "Below is an instruction that describes a task. Write a response that appropriately completes the request.";
+/**
+ * Maps an instruction prompt to the Alpaca prompt format.
+ *
+ * If the instruction has a system prompt, it overrides the default system prompt
+ * (which can impact the results, because the model may be trained on the default system prompt).
+ *
+ * @see https://github.com/tatsu-lab/stanford_alpaca#data-release
+ */
+export const InstructionToAlpacaPromptMapping = () => ({
+    stopTokens: [],
+    map: (instruction) => {
+        let text = instruction.system ??
+            (instruction.input != null
+                ? DEFAULT_SYSTEM_PROMPT_INPUT
+                : DEFAULT_SYSTEM_PROMPT_NO_INPUT);
+        text += "\n\n### Instruction:\n";
+        if (instruction.system != null) {
+            text += `${instruction.system}\n`;
+        }
+        text += instruction.instruction;
+        if (instruction.input != null) {
+            text += `\n\n### Input:\n${instruction.input}`;
+        }
+        text += "\n\n### Response:\n";
+        return text;
+    },
+});

package/prompt/InstructionPrompt.d.ts

@@ -1,7 +1,19 @@
 /**
- * A single instruction prompt. It can contain an optional system message to provide context for the language model.
+ * A single instruction prompt. It can contain an optional system message to define the role and behavior of the language model
+ * and an optiona input to provide context for the language model.
  */
 export type InstructionPrompt = {
+    /**
+     * Optional system message to provide context for the language model. Note that for some models,
+     * changing the system message can impact the results, because the model may be trained on the default system message.
+     */
     system?: string;
+    /**
+     * The instruction for the model.
+     */
     instruction: string;
+    /**
+     * Optional additional input or context, e.g. a the content from which information should be extracted.
+     */
+    input?: string;
 };

package/prompt/Llama2PromptMapping.cjs

@@ -18,7 +18,7 @@ const InstructionToLlama2PromptMapping = () => ({
     stopTokens: [END_SEGMENT],
     map: (instruction) => `${BEGIN_SEGMENT}${BEGIN_INSTRUCTION}${instruction.system != null
         ? ` ${BEGIN_SYSTEM}${instruction.system}${END_SYSTEM}`
-        : ""} ${instruction.instruction} ${END_INSTRUCTION}\n`,
+        : ""} ${instruction.instruction}${instruction.input != null ? `\n\n${instruction.input}` : ""} ${END_INSTRUCTION}\n`,
 });
 exports.InstructionToLlama2PromptMapping = InstructionToLlama2PromptMapping;
 const ChatToLlama2PromptMapping = () => ({

package/prompt/Llama2PromptMapping.js

@@ -15,7 +15,7 @@ export const InstructionToLlama2PromptMapping = () => ({
     stopTokens: [END_SEGMENT],
     map: (instruction) => `${BEGIN_SEGMENT}${BEGIN_INSTRUCTION}${instruction.system != null
         ? ` ${BEGIN_SYSTEM}${instruction.system}${END_SYSTEM}`
-        : ""} ${instruction.instruction} ${END_INSTRUCTION}\n`,
+        : ""} ${instruction.instruction}${instruction.input != null ? `\n\n${instruction.input}` : ""} ${END_INSTRUCTION}\n`,
 });
 export const ChatToLlama2PromptMapping = () => ({
     map: (chatPrompt) => {

package/prompt/OpenAIChatPromptMapping.cjs

@@ -15,6 +15,12 @@ const InstructionToOpenAIChatPromptMapping = () => ({
             role: "user",
             content: instruction.instruction,
         });
+        if (instruction.input != null) {
+            messages.push({
+                role: "user",
+                content: instruction.input,
+            });
+        }
         return messages;
     },
     stopTokens: [],

package/prompt/OpenAIChatPromptMapping.js

@@ -12,6 +12,12 @@ export const InstructionToOpenAIChatPromptMapping = () => ({
             role: "user",
             content: instruction.instruction,
         });
+        if (instruction.input != null) {
+            messages.push({
+                role: "user",
+                content: instruction.input,
+            });
+        }
         return messages;
     },
     stopTokens: [],

package/prompt/TextPromptMapping.cjs

@@ -4,9 +4,17 @@ exports.ChatToTextPromptMapping = exports.InstructionToTextPromptMapping = void
 const validateChatPrompt_js_1 = require("./chat/validateChatPrompt.cjs");
 const InstructionToTextPromptMapping = () => ({
     stopTokens: [],
-    map: (instruction) => instruction.system != null
-        ? `${instruction.system}\n\n${instruction.instruction}`
-        : instruction.instruction,
+    map: (instruction) => {
+        let text = "";
+        if (instruction.system != null) {
+            text += `${instruction.system}\n\n`;
+        }
+        text += instruction.instruction;
+        if (instruction.input != null) {
+            text += `\n\n${instruction.input}`;
+        }
+        return text;
+    },
 });
 exports.InstructionToTextPromptMapping = InstructionToTextPromptMapping;
 /**

package/prompt/TextPromptMapping.js

@@ -1,9 +1,17 @@
 import { validateChatPrompt } from "./chat/validateChatPrompt.js";
 export const InstructionToTextPromptMapping = () => ({
     stopTokens: [],
-    map: (instruction) => instruction.system != null
-        ? `${instruction.system}\n\n${instruction.instruction}`
-        : instruction.instruction,
+    map: (instruction) => {
+        let text = "";
+        if (instruction.system != null) {
+            text += `${instruction.system}\n\n`;
+        }
+        text += instruction.instruction;
+        if (instruction.input != null) {
+            text += `\n\n${instruction.input}`;
+        }
+        return text;
+    },
 });
 /**
  * A mapping from a chat prompt to a text prompt.

package/prompt/VicunaPromptMapping.cjs

@@ -0,0 +1,55 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ChatToVicunaPromptMapping = void 0;
+const validateChatPrompt_js_1 = require("./chat/validateChatPrompt.cjs");
+const DEFAULT_SYSTEM_PROMPT = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.";
+/**
+ * A mapping from a chat prompt to a Vicuna prompt.
+ *
+ * Overridding the system message in the first chat message can affect model respones.
+ *
+ * Vicuna prompt template:
+ * ```
+ * A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
+ *
+ * USER: {prompt}
+ * ASSISTANT:
+ * ```
+ */
+const ChatToVicunaPromptMapping = () => ({
+    map: (chatPrompt) => {
+        (0, validateChatPrompt_js_1.validateChatPrompt)(chatPrompt);
+        let text = "";
+        for (let i = 0; i < chatPrompt.length; i++) {
+            const message = chatPrompt[i];
+            // system message:
+            if (i === 0 &&
+                "system" in message &&
+                typeof message.system === "string") {
+                text += `${message.system}\n\n`;
+                continue;
+            }
+            // first message was not a system message:
+            if (i === 0) {
+                text += `${DEFAULT_SYSTEM_PROMPT}\n\n`;
+            }
+            // user message
+            if ("user" in message) {
+                text += `USER: ${message.user}\n`;
+                continue;
+            }
+            // ai message:
+            if ("ai" in message) {
+                text += `ASSISTANT:\n${message.ai}\n`;
+                continue;
+            }
+            // unsupported message:
+            throw new Error(`Unsupported message: ${JSON.stringify(message)}`);
+        }
+        // AI message prefix:
+        text += `ASSISTANT: `;
+        return text;
+    },
+    stopTokens: [`\nUSER:`],
+});
+exports.ChatToVicunaPromptMapping = ChatToVicunaPromptMapping;

package/prompt/VicunaPromptMapping.d.ts

@@ -0,0 +1,16 @@
+import { PromptMapping } from "./PromptMapping.js";
+import { ChatPrompt } from "./chat/ChatPrompt.js";
+/**
+ * A mapping from a chat prompt to a Vicuna prompt.
+ *
+ * Overridding the system message in the first chat message can affect model respones.
+ *
+ * Vicuna prompt template:
+ * ```
+ * A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
+ *
+ * USER: {prompt}
+ * ASSISTANT:
+ * ```
+ */
+export declare const ChatToVicunaPromptMapping: () => PromptMapping<ChatPrompt, string>;

package/prompt/VicunaPromptMapping.js

@@ -0,0 +1,51 @@
+import { validateChatPrompt } from "./chat/validateChatPrompt.js";
+const DEFAULT_SYSTEM_PROMPT = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.";
+/**
+ * A mapping from a chat prompt to a Vicuna prompt.
+ *
+ * Overridding the system message in the first chat message can affect model respones.
+ *
+ * Vicuna prompt template:
+ * ```
+ * A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
+ *
+ * USER: {prompt}
+ * ASSISTANT:
+ * ```
+ */
+export const ChatToVicunaPromptMapping = () => ({
+    map: (chatPrompt) => {
+        validateChatPrompt(chatPrompt);
+        let text = "";
+        for (let i = 0; i < chatPrompt.length; i++) {
+            const message = chatPrompt[i];
+            // system message:
+            if (i === 0 &&
+                "system" in message &&
+                typeof message.system === "string") {
+                text += `${message.system}\n\n`;
+                continue;
+            }
+            // first message was not a system message:
+            if (i === 0) {
+                text += `${DEFAULT_SYSTEM_PROMPT}\n\n`;
+            }
+            // user message
+            if ("user" in message) {
+                text += `USER: ${message.user}\n`;
+                continue;
+            }
+            // ai message:
+            if ("ai" in message) {
+                text += `ASSISTANT:\n${message.ai}\n`;
+                continue;
+            }
+            // unsupported message:
+            throw new Error(`Unsupported message: ${JSON.stringify(message)}`);
+        }
+        // AI message prefix:
+        text += `ASSISTANT: `;
+        return text;
+    },
+    stopTokens: [`\nUSER:`],
+});

package/prompt/index.cjs

@@ -14,12 +14,14 @@ var __exportStar = (this && this.__exportStar) || function(m, exports) {
     for (var p in m) if (p !== "default" && !Object.prototype.hasOwnProperty.call(exports, p)) __createBinding(exports, m, p);
 };
 Object.defineProperty(exports, "__esModule", { value: true });
+__exportStar(require("./AlpacaPromptMapping.cjs"), exports);
 __exportStar(require("./InstructionPrompt.cjs"), exports);
 __exportStar(require("./Llama2PromptMapping.cjs"), exports);
 __exportStar(require("./OpenAIChatPromptMapping.cjs"), exports);
 __exportStar(require("./PromptMapping.cjs"), exports);
 __exportStar(require("./PromptMappingTextGenerationModel.cjs"), exports);
 __exportStar(require("./TextPromptMapping.cjs"), exports);
+__exportStar(require("./VicunaPromptMapping.cjs"), exports);
 __exportStar(require("./chat/ChatPrompt.cjs"), exports);
 __exportStar(require("./chat/trimChatPrompt.cjs"), exports);
 __exportStar(require("./chat/validateChatPrompt.cjs"), exports);

package/prompt/index.d.ts

@@ -1,9 +1,11 @@
+export * from "./AlpacaPromptMapping.js";
 export * from "./InstructionPrompt.js";
 export * from "./Llama2PromptMapping.js";
 export * from "./OpenAIChatPromptMapping.js";
 export * from "./PromptMapping.js";
 export * from "./PromptMappingTextGenerationModel.js";
 export * from "./TextPromptMapping.js";
+export * from "./VicunaPromptMapping.js";
 export * from "./chat/ChatPrompt.js";
 export * from "./chat/trimChatPrompt.js";
 export * from "./chat/validateChatPrompt.js";

package/prompt/index.js

@@ -1,9 +1,11 @@
+export * from "./AlpacaPromptMapping.js";
 export * from "./InstructionPrompt.js";
 export * from "./Llama2PromptMapping.js";
 export * from "./OpenAIChatPromptMapping.js";
 export * from "./PromptMapping.js";
 export * from "./PromptMappingTextGenerationModel.js";
 export * from "./TextPromptMapping.js";
+export * from "./VicunaPromptMapping.js";
 export * from "./chat/ChatPrompt.js";
 export * from "./chat/trimChatPrompt.js";
 export * from "./chat/validateChatPrompt.js";

package/util/api/postToApi.cjs

@@ -18,10 +18,12 @@ const createJsonResponseHandler = (responseSchema) => async ({ response, url, re
 exports.createJsonResponseHandler = createJsonResponseHandler;
 const createTextResponseHandler = () => async ({ response }) => response.text();
 exports.createTextResponseHandler = createTextResponseHandler;
-const postJsonToApi = async ({ url, apiKey, body, failedResponseHandler, successfulResponseHandler, abortSignal, }) => (0, exports.postToApi)({
+const postJsonToApi = async ({ url, headers, body, failedResponseHandler, successfulResponseHandler, abortSignal, }) => (0, exports.postToApi)({
     url,
-    apiKey,
-    contentType: "application/json",
+    headers: {
+        ...headers,
+        "Content-Type": "application/json",
+    },
     body: {
         content: JSON.stringify(body),
         values: body,
@@ -31,15 +33,8 @@ const postJsonToApi = async ({ url, apiKey, body, failedResponseHandler, success
     abortSignal,
 });
 exports.postJsonToApi = postJsonToApi;
-const postToApi = async ({ url, apiKey, contentType, body, successfulResponseHandler, failedResponseHandler, abortSignal, }) => {
+const postToApi = async ({ url, headers = {}, body, successfulResponseHandler, failedResponseHandler, abortSignal, }) => {
     try {
-        const headers = {};
-        if (apiKey !== undefined) {
-            headers["Authorization"] = `Bearer ${apiKey}`;
-        }
-        if (contentType !== null) {
-            headers["Content-Type"] = contentType;
-        }
         const response = await fetch(url, {
             method: "POST",
             headers,

package/util/api/postToApi.d.ts

@@ -7,18 +7,17 @@ export type ResponseHandler<T> = (options: {
 }) => PromiseLike<T>;
 export declare const createJsonResponseHandler: <T>(responseSchema: z.ZodType<T, z.ZodTypeDef, T>) => ResponseHandler<T>;
 export declare const createTextResponseHandler: () => ResponseHandler<string>;
-export declare const postJsonToApi: <T>({ url, apiKey, body, failedResponseHandler, successfulResponseHandler, abortSignal, }: {
+export declare const postJsonToApi: <T>({ url, headers, body, failedResponseHandler, successfulResponseHandler, abortSignal, }: {
     url: string;
-    apiKey?: string | undefined;
+    headers?: Record<string, string> | undefined;
     body: unknown;
     failedResponseHandler: ResponseHandler<ApiCallError>;
     successfulResponseHandler: ResponseHandler<T>;
     abortSignal?: AbortSignal | undefined;
 }) => Promise<T>;
-export declare const postToApi: <T>({ url, apiKey, contentType, body, successfulResponseHandler, failedResponseHandler, abortSignal, }: {
+export declare const postToApi: <T>({ url, headers, body, successfulResponseHandler, failedResponseHandler, abortSignal, }: {
     url: string;
-    apiKey?: string | undefined;
-    contentType: string | null;
+    headers?: Record<string, string> | undefined;
     body: {
         content: string | FormData;
         values: unknown;

package/util/api/postToApi.js

@@ -13,10 +13,12 @@ export const createJsonResponseHandler = (responseSchema) => async ({ response,
     return parsedResult.data;
 };
 export const createTextResponseHandler = () => async ({ response }) => response.text();
-export const postJsonToApi = async ({ url, apiKey, body, failedResponseHandler, successfulResponseHandler, abortSignal, }) => postToApi({
+export const postJsonToApi = async ({ url, headers, body, failedResponseHandler, successfulResponseHandler, abortSignal, }) => postToApi({
     url,
-    apiKey,
-    contentType: "application/json",
+    headers: {
+        ...headers,
+        "Content-Type": "application/json",
+    },
     body: {
         content: JSON.stringify(body),
         values: body,
@@ -25,15 +27,8 @@ export const postJsonToApi = async ({ url, apiKey, body, failedResponseHandler,
     successfulResponseHandler,
     abortSignal,
 });
-export const postToApi = async ({ url, apiKey, contentType, body, successfulResponseHandler, failedResponseHandler, abortSignal, }) => {
+export const postToApi = async ({ url, headers = {}, body, successfulResponseHandler, failedResponseHandler, abortSignal, }) => {
     try {
-        const headers = {};
-        if (apiKey !== undefined) {
-            headers["Authorization"] = `Bearer ${apiKey}`;
-        }
-        if (contentType !== null) {
-            headers["Content-Type"] = contentType;
-        }
         const response = await fetch(url, {
             method: "POST",
             headers,

package/util/api/retryWithExponentialBackoff.cjs

@@ -20,7 +20,7 @@ async function _retryWithExponentialBackoff(f, { maxTries, delay, backoffFactor,
     catch (error) {
         const newErrors = [...errors, error];
         const tryNumber = newErrors.length;
-        if (tryNumber > maxTries) {
+        if (tryNumber >= maxTries) {
             throw new RetryError_js_1.RetryError({
                 message: `Failed after ${tryNumber} tries.`,
                 reason: "maxTriesExceeded",

package/util/api/retryWithExponentialBackoff.js

@@ -16,7 +16,7 @@ async function _retryWithExponentialBackoff(f, { maxTries, delay, backoffFactor,
     catch (error) {
         const newErrors = [...errors, error];
         const tryNumber = newErrors.length;
-        if (tryNumber > maxTries) {
+        if (tryNumber >= maxTries) {
             throw new RetryError({
                 message: `Failed after ${tryNumber} tries.`,
                 reason: "maxTriesExceeded",