llm-fns 1.0.6 → 1.0.7
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/createJsonSchemaLlmClient.d.ts +29 -0
- package/dist/createJsonSchemaLlmClient.js +190 -0
- package/dist/createZodLlmClient.d.ts +3 -21
- package/dist/createZodLlmClient.js +14 -180
- package/dist/createZodLlmClient.spec.js +13 -0
- package/dist/index.d.ts +1 -0
- package/dist/index.js +1 -0
- package/dist/llmFactory.d.ts +2 -0
- package/dist/llmFactory.js +6 -1
- package/package.json +1 -1
|
@@ -0,0 +1,29 @@
|
|
|
1
|
+
import OpenAI from 'openai';
|
|
2
|
+
import { PromptFunction, LlmPromptOptions, IsPromptCachedFunction } from "./createLlmClient.js";
|
|
3
|
+
export type JsonSchemaLlmClientOptions = Omit<LlmPromptOptions, 'messages' | 'response_format'> & {
|
|
4
|
+
maxRetries?: number;
|
|
5
|
+
/**
|
|
6
|
+
* If true, passes `response_format: { type: 'json_object' }` to the model.
|
|
7
|
+
* If false, only includes the schema in the system prompt.
|
|
8
|
+
* Defaults to true.
|
|
9
|
+
*/
|
|
10
|
+
useResponseFormat?: boolean;
|
|
11
|
+
/**
|
|
12
|
+
* A hook to process the parsed JSON data before it is validated.
|
|
13
|
+
* This can be used to merge partial results or perform other transformations.
|
|
14
|
+
* @param data The parsed JSON data from the LLM response.
|
|
15
|
+
* @returns The processed data to be validated.
|
|
16
|
+
*/
|
|
17
|
+
beforeValidation?: (data: any) => any;
|
|
18
|
+
};
|
|
19
|
+
export interface CreateJsonSchemaLlmClientParams {
|
|
20
|
+
prompt: PromptFunction;
|
|
21
|
+
isPromptCached: IsPromptCachedFunction;
|
|
22
|
+
fallbackPrompt?: PromptFunction;
|
|
23
|
+
disableJsonFixer?: boolean;
|
|
24
|
+
}
|
|
25
|
+
export declare function createJsonSchemaLlmClient(params: CreateJsonSchemaLlmClientParams): {
|
|
26
|
+
promptJson: <T>(messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[], schema: Record<string, any>, validator: (data: any) => T, options?: JsonSchemaLlmClientOptions) => Promise<T>;
|
|
27
|
+
isPromptJsonCached: (messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[], schema: Record<string, any>, options?: JsonSchemaLlmClientOptions) => Promise<boolean>;
|
|
28
|
+
};
|
|
29
|
+
export type JsonSchemaClient = ReturnType<typeof createJsonSchemaLlmClient>;
|
|
@@ -0,0 +1,190 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.createJsonSchemaLlmClient = createJsonSchemaLlmClient;
|
|
4
|
+
const createLlmRetryClient_js_1 = require("./createLlmRetryClient.js");
|
|
5
|
+
function createJsonSchemaLlmClient(params) {
|
|
6
|
+
const { prompt, isPromptCached, fallbackPrompt, disableJsonFixer = false } = params;
|
|
7
|
+
const llmRetryClient = (0, createLlmRetryClient_js_1.createLlmRetryClient)({ prompt, fallbackPrompt });
|
|
8
|
+
async function _tryToFixJson(brokenResponse, schemaJsonString, errorDetails, options) {
|
|
9
|
+
const fixupPrompt = `
|
|
10
|
+
An attempt to generate a JSON object resulted in the following output, which is either not valid JSON or does not conform to the required schema.
|
|
11
|
+
|
|
12
|
+
Your task is to act as a JSON fixer. Analyze the provided "BROKEN RESPONSE" and correct it to match the "REQUIRED JSON SCHEMA".
|
|
13
|
+
|
|
14
|
+
- If the broken response contains all the necessary information to create a valid JSON object according to the schema, please provide the corrected, valid JSON object.
|
|
15
|
+
- If the broken response is missing essential information, or is too garbled to be fixed, please respond with the exact string: "CANNOT_FIX".
|
|
16
|
+
- Your response must be ONLY the corrected JSON object or the string "CANNOT_FIX". Do not include any other text, explanations, or markdown formatting.
|
|
17
|
+
|
|
18
|
+
REQUIRED JSON SCHEMA:
|
|
19
|
+
${schemaJsonString}
|
|
20
|
+
|
|
21
|
+
ERROR DETAILS:
|
|
22
|
+
${errorDetails}
|
|
23
|
+
|
|
24
|
+
BROKEN RESPONSE:
|
|
25
|
+
${brokenResponse}
|
|
26
|
+
`;
|
|
27
|
+
const messages = [
|
|
28
|
+
{ role: 'system', content: 'You are an expert at fixing malformed JSON data to match a specific schema.' },
|
|
29
|
+
{ role: 'user', content: fixupPrompt }
|
|
30
|
+
];
|
|
31
|
+
const useResponseFormat = options?.useResponseFormat ?? true;
|
|
32
|
+
const response_format = useResponseFormat
|
|
33
|
+
? { type: 'json_object' }
|
|
34
|
+
: undefined;
|
|
35
|
+
const { maxRetries, useResponseFormat: _useResponseFormat, ...restOptions } = options || {};
|
|
36
|
+
const completion = await prompt({
|
|
37
|
+
messages,
|
|
38
|
+
response_format,
|
|
39
|
+
...restOptions
|
|
40
|
+
});
|
|
41
|
+
const fixedResponse = completion.choices[0]?.message?.content;
|
|
42
|
+
if (fixedResponse && fixedResponse.trim() === 'CANNOT_FIX') {
|
|
43
|
+
return null;
|
|
44
|
+
}
|
|
45
|
+
return fixedResponse || null;
|
|
46
|
+
}
|
|
47
|
+
async function _parseOrFixJson(llmResponseString, schemaJsonString, options) {
|
|
48
|
+
let jsonDataToParse = llmResponseString.trim();
|
|
49
|
+
// Robust handling for responses wrapped in markdown code blocks
|
|
50
|
+
const codeBlockRegex = /```(?:json)?\s*([\s\S]*?)\s*```/;
|
|
51
|
+
const match = codeBlockRegex.exec(jsonDataToParse);
|
|
52
|
+
if (match && match[1]) {
|
|
53
|
+
jsonDataToParse = match[1].trim();
|
|
54
|
+
}
|
|
55
|
+
if (jsonDataToParse === "") {
|
|
56
|
+
throw new Error("LLM returned an empty string.");
|
|
57
|
+
}
|
|
58
|
+
try {
|
|
59
|
+
return JSON.parse(jsonDataToParse);
|
|
60
|
+
}
|
|
61
|
+
catch (parseError) {
|
|
62
|
+
if (disableJsonFixer) {
|
|
63
|
+
throw parseError; // re-throw original error
|
|
64
|
+
}
|
|
65
|
+
// Attempt a one-time fix before failing.
|
|
66
|
+
const errorDetails = `JSON Parse Error: ${parseError.message}`;
|
|
67
|
+
const fixedResponse = await _tryToFixJson(jsonDataToParse, schemaJsonString, errorDetails, options);
|
|
68
|
+
if (fixedResponse) {
|
|
69
|
+
try {
|
|
70
|
+
return JSON.parse(fixedResponse);
|
|
71
|
+
}
|
|
72
|
+
catch (e) {
|
|
73
|
+
// Fix-up failed, throw original error.
|
|
74
|
+
throw parseError;
|
|
75
|
+
}
|
|
76
|
+
}
|
|
77
|
+
throw parseError; // if no fixed response
|
|
78
|
+
}
|
|
79
|
+
}
|
|
80
|
+
async function _validateOrFix(jsonData, validator, schemaJsonString, options) {
|
|
81
|
+
try {
|
|
82
|
+
if (options?.beforeValidation) {
|
|
83
|
+
jsonData = options.beforeValidation(jsonData);
|
|
84
|
+
}
|
|
85
|
+
return validator(jsonData);
|
|
86
|
+
}
|
|
87
|
+
catch (validationError) {
|
|
88
|
+
if (disableJsonFixer) {
|
|
89
|
+
throw validationError;
|
|
90
|
+
}
|
|
91
|
+
// Attempt a one-time fix for schema validation errors.
|
|
92
|
+
const errorDetails = `Schema Validation Error: ${validationError.message}`;
|
|
93
|
+
const fixedResponse = await _tryToFixJson(JSON.stringify(jsonData, null, 2), schemaJsonString, errorDetails, options);
|
|
94
|
+
if (fixedResponse) {
|
|
95
|
+
try {
|
|
96
|
+
let fixedJsonData = JSON.parse(fixedResponse);
|
|
97
|
+
if (options?.beforeValidation) {
|
|
98
|
+
fixedJsonData = options.beforeValidation(fixedJsonData);
|
|
99
|
+
}
|
|
100
|
+
return validator(fixedJsonData);
|
|
101
|
+
}
|
|
102
|
+
catch (e) {
|
|
103
|
+
// Fix-up failed, throw original validation error
|
|
104
|
+
throw validationError;
|
|
105
|
+
}
|
|
106
|
+
}
|
|
107
|
+
throw validationError; // if no fixed response
|
|
108
|
+
}
|
|
109
|
+
}
|
|
110
|
+
function _getJsonPromptConfig(messages, schema, options) {
|
|
111
|
+
const schemaJsonString = JSON.stringify(schema);
|
|
112
|
+
const commonPromptFooter = `
|
|
113
|
+
Your response MUST be a single JSON entity (object or array) that strictly adheres to the following JSON schema.
|
|
114
|
+
Do NOT include any other text, explanations, or markdown formatting (like \`\`\`json) before or after the JSON entity.
|
|
115
|
+
|
|
116
|
+
JSON schema:
|
|
117
|
+
${schemaJsonString}`;
|
|
118
|
+
// Clone messages to avoid mutating the input
|
|
119
|
+
const finalMessages = [...messages];
|
|
120
|
+
// Find the first system message to append instructions to
|
|
121
|
+
const systemMessageIndex = finalMessages.findIndex(m => m.role === 'system');
|
|
122
|
+
if (systemMessageIndex !== -1) {
|
|
123
|
+
// Append to existing system message
|
|
124
|
+
const existingContent = finalMessages[systemMessageIndex].content;
|
|
125
|
+
finalMessages[systemMessageIndex] = {
|
|
126
|
+
...finalMessages[systemMessageIndex],
|
|
127
|
+
content: `${existingContent}\n${commonPromptFooter}`
|
|
128
|
+
};
|
|
129
|
+
}
|
|
130
|
+
else {
|
|
131
|
+
// Prepend new system message
|
|
132
|
+
finalMessages.unshift({
|
|
133
|
+
role: 'system',
|
|
134
|
+
content: commonPromptFooter
|
|
135
|
+
});
|
|
136
|
+
}
|
|
137
|
+
const useResponseFormat = options?.useResponseFormat ?? true;
|
|
138
|
+
const response_format = useResponseFormat
|
|
139
|
+
? { type: 'json_object' }
|
|
140
|
+
: undefined;
|
|
141
|
+
return { finalMessages, schemaJsonString, response_format };
|
|
142
|
+
}
|
|
143
|
+
async function promptJson(messages, schema, validator, options) {
|
|
144
|
+
const { finalMessages, schemaJsonString, response_format } = _getJsonPromptConfig(messages, schema, options);
|
|
145
|
+
const processResponse = async (llmResponseString) => {
|
|
146
|
+
let jsonData;
|
|
147
|
+
try {
|
|
148
|
+
jsonData = await _parseOrFixJson(llmResponseString, schemaJsonString, options);
|
|
149
|
+
}
|
|
150
|
+
catch (parseError) {
|
|
151
|
+
const errorMessage = `Your previous response resulted in an error.
|
|
152
|
+
Error Type: JSON_PARSE_ERROR
|
|
153
|
+
Error Details: ${parseError.message}
|
|
154
|
+
The response provided was not valid JSON. Please correct it.`;
|
|
155
|
+
throw new createLlmRetryClient_js_1.LlmRetryError(errorMessage, 'JSON_PARSE_ERROR', undefined, llmResponseString);
|
|
156
|
+
}
|
|
157
|
+
try {
|
|
158
|
+
const validatedData = await _validateOrFix(jsonData, validator, schemaJsonString, options);
|
|
159
|
+
return validatedData;
|
|
160
|
+
}
|
|
161
|
+
catch (validationError) {
|
|
162
|
+
// We assume the validator throws an error with a meaningful message
|
|
163
|
+
const rawResponseForError = JSON.stringify(jsonData, null, 2);
|
|
164
|
+
const errorDetails = validationError.message;
|
|
165
|
+
const errorMessage = `Your previous response resulted in an error.
|
|
166
|
+
Error Type: SCHEMA_VALIDATION_ERROR
|
|
167
|
+
Error Details: ${errorDetails}
|
|
168
|
+
The response was valid JSON but did not conform to the required schema. Please review the errors and the schema to provide a corrected response.`;
|
|
169
|
+
throw new createLlmRetryClient_js_1.LlmRetryError(errorMessage, 'CUSTOM_ERROR', validationError, rawResponseForError);
|
|
170
|
+
}
|
|
171
|
+
};
|
|
172
|
+
const retryOptions = {
|
|
173
|
+
...options,
|
|
174
|
+
messages: finalMessages,
|
|
175
|
+
response_format,
|
|
176
|
+
validate: processResponse
|
|
177
|
+
};
|
|
178
|
+
return llmRetryClient.promptTextRetry(retryOptions);
|
|
179
|
+
}
|
|
180
|
+
async function isPromptJsonCached(messages, schema, options) {
|
|
181
|
+
const { finalMessages, response_format } = _getJsonPromptConfig(messages, schema, options);
|
|
182
|
+
const { maxRetries, useResponseFormat: _u, beforeValidation, ...restOptions } = options || {};
|
|
183
|
+
return isPromptCached({
|
|
184
|
+
messages: finalMessages,
|
|
185
|
+
response_format,
|
|
186
|
+
...restOptions
|
|
187
|
+
});
|
|
188
|
+
}
|
|
189
|
+
return { promptJson, isPromptJsonCached };
|
|
190
|
+
}
|
|
@@ -1,28 +1,10 @@
|
|
|
1
1
|
import OpenAI from 'openai';
|
|
2
2
|
import * as z from "zod";
|
|
3
|
-
import { PromptFunction, LlmPromptOptions, IsPromptCachedFunction } from "./createLlmClient.js";
|
|
4
3
|
import { ZodTypeAny } from "zod";
|
|
5
|
-
|
|
6
|
-
|
|
7
|
-
/**
|
|
8
|
-
* If true, passes `response_format: { type: 'json_object' }` to the model.
|
|
9
|
-
* If false, only includes the schema in the system prompt.
|
|
10
|
-
* Defaults to true.
|
|
11
|
-
*/
|
|
12
|
-
useResponseFormat?: boolean;
|
|
13
|
-
/**
|
|
14
|
-
* A hook to process the parsed JSON data before it is validated against the Zod schema.
|
|
15
|
-
* This can be used to merge partial results or perform other transformations.
|
|
16
|
-
* @param data The parsed JSON data from the LLM response.
|
|
17
|
-
* @returns The processed data to be validated.
|
|
18
|
-
*/
|
|
19
|
-
beforeValidation?: (data: any) => any;
|
|
20
|
-
};
|
|
4
|
+
import { JsonSchemaClient, JsonSchemaLlmClientOptions } from "./createJsonSchemaLlmClient.js";
|
|
5
|
+
export type ZodLlmClientOptions = JsonSchemaLlmClientOptions;
|
|
21
6
|
export interface CreateZodLlmClientParams {
|
|
22
|
-
|
|
23
|
-
isPromptCached: IsPromptCachedFunction;
|
|
24
|
-
fallbackPrompt?: PromptFunction;
|
|
25
|
-
disableJsonFixer?: boolean;
|
|
7
|
+
jsonSchemaClient: JsonSchemaClient;
|
|
26
8
|
}
|
|
27
9
|
export interface NormalizedZodArgs<T extends ZodTypeAny> {
|
|
28
10
|
messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[];
|
|
@@ -36,7 +36,6 @@ Object.defineProperty(exports, "__esModule", { value: true });
|
|
|
36
36
|
exports.normalizeZodArgs = normalizeZodArgs;
|
|
37
37
|
exports.createZodLlmClient = createZodLlmClient;
|
|
38
38
|
const z = __importStar(require("zod"));
|
|
39
|
-
const createLlmRetryClient_js_1 = require("./createLlmRetryClient.js");
|
|
40
39
|
const zod_1 = require("zod");
|
|
41
40
|
function isZodSchema(obj) {
|
|
42
41
|
return (typeof obj === 'object' &&
|
|
@@ -89,197 +88,32 @@ function normalizeZodArgs(arg1, arg2, arg3, arg4) {
|
|
|
89
88
|
throw new Error("Invalid arguments passed to promptZod");
|
|
90
89
|
}
|
|
91
90
|
function createZodLlmClient(params) {
|
|
92
|
-
const {
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
const fixupPrompt = `
|
|
96
|
-
An attempt to generate a JSON object resulted in the following output, which is either not valid JSON or does not conform to the required schema.
|
|
97
|
-
|
|
98
|
-
Your task is to act as a JSON fixer. Analyze the provided "BROKEN RESPONSE" and correct it to match the "REQUIRED JSON SCHEMA".
|
|
99
|
-
|
|
100
|
-
- If the broken response contains all the necessary information to create a valid JSON object according to the schema, please provide the corrected, valid JSON object.
|
|
101
|
-
- If the broken response is missing essential information, or is too garbled to be fixed, please respond with the exact string: "CANNOT_FIX".
|
|
102
|
-
- Your response must be ONLY the corrected JSON object or the string "CANNOT_FIX". Do not include any other text, explanations, or markdown formatting.
|
|
103
|
-
|
|
104
|
-
REQUIRED JSON SCHEMA:
|
|
105
|
-
${schemaJsonString}
|
|
106
|
-
|
|
107
|
-
ERROR DETAILS:
|
|
108
|
-
${errorDetails}
|
|
109
|
-
|
|
110
|
-
BROKEN RESPONSE:
|
|
111
|
-
${brokenResponse}
|
|
112
|
-
`;
|
|
113
|
-
const messages = [
|
|
114
|
-
{ role: 'system', content: 'You are an expert at fixing malformed JSON data to match a specific schema.' },
|
|
115
|
-
{ role: 'user', content: fixupPrompt }
|
|
116
|
-
];
|
|
117
|
-
const useResponseFormat = options?.useResponseFormat ?? true;
|
|
118
|
-
const response_format = useResponseFormat
|
|
119
|
-
? { type: 'json_object' }
|
|
120
|
-
: undefined;
|
|
121
|
-
const { maxRetries, useResponseFormat: _useResponseFormat, ...restOptions } = options || {};
|
|
122
|
-
const completion = await prompt({
|
|
123
|
-
messages,
|
|
124
|
-
response_format,
|
|
125
|
-
...restOptions
|
|
126
|
-
});
|
|
127
|
-
const fixedResponse = completion.choices[0]?.message?.content;
|
|
128
|
-
if (fixedResponse && fixedResponse.trim() === 'CANNOT_FIX') {
|
|
129
|
-
return null;
|
|
130
|
-
}
|
|
131
|
-
return fixedResponse || null;
|
|
132
|
-
}
|
|
133
|
-
async function _parseOrFixJson(llmResponseString, schemaJsonString, options) {
|
|
134
|
-
let jsonDataToParse = llmResponseString.trim();
|
|
135
|
-
// Robust handling for responses wrapped in markdown code blocks
|
|
136
|
-
const codeBlockRegex = /```(?:json)?\s*([\s\S]*?)\s*```/;
|
|
137
|
-
const match = codeBlockRegex.exec(jsonDataToParse);
|
|
138
|
-
if (match && match[1]) {
|
|
139
|
-
jsonDataToParse = match[1].trim();
|
|
140
|
-
}
|
|
141
|
-
if (jsonDataToParse === "") {
|
|
142
|
-
throw new Error("LLM returned an empty string.");
|
|
143
|
-
}
|
|
144
|
-
try {
|
|
145
|
-
return JSON.parse(jsonDataToParse);
|
|
146
|
-
}
|
|
147
|
-
catch (parseError) {
|
|
148
|
-
if (disableJsonFixer) {
|
|
149
|
-
throw parseError; // re-throw original error
|
|
150
|
-
}
|
|
151
|
-
// Attempt a one-time fix before failing.
|
|
152
|
-
const errorDetails = `JSON Parse Error: ${parseError.message}`;
|
|
153
|
-
const fixedResponse = await _tryToFixJson(jsonDataToParse, schemaJsonString, errorDetails, options);
|
|
154
|
-
if (fixedResponse) {
|
|
155
|
-
try {
|
|
156
|
-
return JSON.parse(fixedResponse);
|
|
157
|
-
}
|
|
158
|
-
catch (e) {
|
|
159
|
-
// Fix-up failed, throw original error.
|
|
160
|
-
throw parseError;
|
|
161
|
-
}
|
|
162
|
-
}
|
|
163
|
-
throw parseError; // if no fixed response
|
|
164
|
-
}
|
|
165
|
-
}
|
|
166
|
-
async function _validateOrFixSchema(jsonData, dataExtractionSchema, schemaJsonString, options) {
|
|
167
|
-
try {
|
|
168
|
-
if (options?.beforeValidation) {
|
|
169
|
-
jsonData = options.beforeValidation(jsonData);
|
|
170
|
-
}
|
|
171
|
-
return dataExtractionSchema.parse(jsonData);
|
|
172
|
-
}
|
|
173
|
-
catch (validationError) {
|
|
174
|
-
if (!(validationError instanceof zod_1.ZodError) || disableJsonFixer) {
|
|
175
|
-
throw validationError;
|
|
176
|
-
}
|
|
177
|
-
// Attempt a one-time fix for schema validation errors.
|
|
178
|
-
const errorDetails = `Schema Validation Error: ${JSON.stringify(validationError.format(), null, 2)}`;
|
|
179
|
-
const fixedResponse = await _tryToFixJson(JSON.stringify(jsonData, null, 2), schemaJsonString, errorDetails, options);
|
|
180
|
-
if (fixedResponse) {
|
|
181
|
-
try {
|
|
182
|
-
let fixedJsonData = JSON.parse(fixedResponse);
|
|
183
|
-
if (options?.beforeValidation) {
|
|
184
|
-
fixedJsonData = options.beforeValidation(fixedJsonData);
|
|
185
|
-
}
|
|
186
|
-
return dataExtractionSchema.parse(fixedJsonData);
|
|
187
|
-
}
|
|
188
|
-
catch (e) {
|
|
189
|
-
// Fix-up failed, throw original validation error
|
|
190
|
-
throw validationError;
|
|
191
|
-
}
|
|
192
|
-
}
|
|
193
|
-
throw validationError; // if no fixed response
|
|
194
|
-
}
|
|
195
|
-
}
|
|
196
|
-
function _getZodPromptConfig(messages, dataExtractionSchema, options) {
|
|
91
|
+
const { jsonSchemaClient } = params;
|
|
92
|
+
async function promptZod(arg1, arg2, arg3, arg4) {
|
|
93
|
+
const { messages, dataExtractionSchema, options } = normalizeZodArgs(arg1, arg2, arg3, arg4);
|
|
197
94
|
const schema = z.toJSONSchema(dataExtractionSchema, {
|
|
198
95
|
unrepresentable: 'any'
|
|
199
96
|
});
|
|
200
|
-
const
|
|
201
|
-
const commonPromptFooter = `
|
|
202
|
-
Your response MUST be a single JSON entity (object or array) that strictly adheres to the following JSON schema.
|
|
203
|
-
Do NOT include any other text, explanations, or markdown formatting (like \`\`\`json) before or after the JSON entity.
|
|
204
|
-
|
|
205
|
-
JSON schema:
|
|
206
|
-
${schemaJsonString}`;
|
|
207
|
-
// Clone messages to avoid mutating the input
|
|
208
|
-
const finalMessages = [...messages];
|
|
209
|
-
// Find the first system message to append instructions to
|
|
210
|
-
const systemMessageIndex = finalMessages.findIndex(m => m.role === 'system');
|
|
211
|
-
if (systemMessageIndex !== -1) {
|
|
212
|
-
// Append to existing system message
|
|
213
|
-
const existingContent = finalMessages[systemMessageIndex].content;
|
|
214
|
-
finalMessages[systemMessageIndex] = {
|
|
215
|
-
...finalMessages[systemMessageIndex],
|
|
216
|
-
content: `${existingContent}\n${commonPromptFooter}`
|
|
217
|
-
};
|
|
218
|
-
}
|
|
219
|
-
else {
|
|
220
|
-
// Prepend new system message
|
|
221
|
-
finalMessages.unshift({
|
|
222
|
-
role: 'system',
|
|
223
|
-
content: commonPromptFooter
|
|
224
|
-
});
|
|
225
|
-
}
|
|
226
|
-
const useResponseFormat = options?.useResponseFormat ?? true;
|
|
227
|
-
const response_format = useResponseFormat
|
|
228
|
-
? { type: 'json_object' }
|
|
229
|
-
: undefined;
|
|
230
|
-
return { finalMessages, schemaJsonString, response_format };
|
|
231
|
-
}
|
|
232
|
-
async function promptZod(arg1, arg2, arg3, arg4) {
|
|
233
|
-
const { messages, dataExtractionSchema, options } = normalizeZodArgs(arg1, arg2, arg3, arg4);
|
|
234
|
-
const { finalMessages, schemaJsonString, response_format } = _getZodPromptConfig(messages, dataExtractionSchema, options);
|
|
235
|
-
const processResponse = async (llmResponseString) => {
|
|
236
|
-
let jsonData;
|
|
237
|
-
try {
|
|
238
|
-
jsonData = await _parseOrFixJson(llmResponseString, schemaJsonString, options);
|
|
239
|
-
}
|
|
240
|
-
catch (parseError) {
|
|
241
|
-
const errorMessage = `Your previous response resulted in an error.
|
|
242
|
-
Error Type: JSON_PARSE_ERROR
|
|
243
|
-
Error Details: ${parseError.message}
|
|
244
|
-
The response provided was not valid JSON. Please correct it.`;
|
|
245
|
-
throw new createLlmRetryClient_js_1.LlmRetryError(errorMessage, 'JSON_PARSE_ERROR', undefined, llmResponseString);
|
|
246
|
-
}
|
|
97
|
+
const validator = (data) => {
|
|
247
98
|
try {
|
|
248
|
-
|
|
249
|
-
return validatedData;
|
|
99
|
+
return dataExtractionSchema.parse(data);
|
|
250
100
|
}
|
|
251
|
-
catch (
|
|
252
|
-
if (
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
const errorMessage = `Your previous response resulted in an error.
|
|
256
|
-
Error Type: SCHEMA_VALIDATION_ERROR
|
|
257
|
-
Error Details: ${errorDetails}
|
|
258
|
-
The response was valid JSON but did not conform to the required schema. Please review the errors and the schema to provide a corrected response.`;
|
|
259
|
-
throw new createLlmRetryClient_js_1.LlmRetryError(errorMessage, 'CUSTOM_ERROR', validationError.format(), rawResponseForError);
|
|
101
|
+
catch (error) {
|
|
102
|
+
if (error instanceof zod_1.ZodError) {
|
|
103
|
+
// Format the error nicely for the LLM
|
|
104
|
+
throw new Error(JSON.stringify(error.format(), null, 2));
|
|
260
105
|
}
|
|
261
|
-
|
|
262
|
-
throw validationError;
|
|
106
|
+
throw error;
|
|
263
107
|
}
|
|
264
108
|
};
|
|
265
|
-
|
|
266
|
-
...options,
|
|
267
|
-
messages: finalMessages,
|
|
268
|
-
response_format,
|
|
269
|
-
validate: processResponse
|
|
270
|
-
};
|
|
271
|
-
// Use promptTextRetry because we expect a string response to parse as JSON
|
|
272
|
-
return llmRetryClient.promptTextRetry(retryOptions);
|
|
109
|
+
return jsonSchemaClient.promptJson(messages, schema, validator, options);
|
|
273
110
|
}
|
|
274
111
|
async function isPromptZodCached(arg1, arg2, arg3, arg4) {
|
|
275
112
|
const { messages, dataExtractionSchema, options } = normalizeZodArgs(arg1, arg2, arg3, arg4);
|
|
276
|
-
const
|
|
277
|
-
|
|
278
|
-
return isPromptCached({
|
|
279
|
-
messages: finalMessages,
|
|
280
|
-
response_format,
|
|
281
|
-
...restOptions
|
|
113
|
+
const schema = z.toJSONSchema(dataExtractionSchema, {
|
|
114
|
+
unrepresentable: 'any'
|
|
282
115
|
});
|
|
116
|
+
return jsonSchemaClient.isPromptJsonCached(messages, schema, options);
|
|
283
117
|
}
|
|
284
118
|
return { promptZod, isPromptZodCached };
|
|
285
119
|
}
|
|
@@ -79,6 +79,19 @@ const createZodLlmClient_js_1 = require("./createZodLlmClient.js");
|
|
|
79
79
|
(0, vitest_1.expect)(result.dataExtractionSchema).toBe(TestSchema);
|
|
80
80
|
(0, vitest_1.expect)(result.options).toBe(options);
|
|
81
81
|
});
|
|
82
|
+
(0, vitest_1.it)('should normalize Messages Array with few-shot examples + Schema (Case 0)', () => {
|
|
83
|
+
const messages = [
|
|
84
|
+
{ role: 'system', content: 'Extract sentiment.' },
|
|
85
|
+
{ role: 'user', content: 'I love this!' },
|
|
86
|
+
{ role: 'assistant', content: JSON.stringify({ sentiment: 'positive' }) },
|
|
87
|
+
{ role: 'user', content: 'I hate this.' }
|
|
88
|
+
];
|
|
89
|
+
const Schema = zod_1.z.object({ sentiment: zod_1.z.enum(['positive', 'negative']) });
|
|
90
|
+
const result = (0, createZodLlmClient_js_1.normalizeZodArgs)(messages, Schema);
|
|
91
|
+
(0, vitest_1.expect)(result.messages).toBe(messages);
|
|
92
|
+
(0, vitest_1.expect)(result.dataExtractionSchema).toBe(Schema);
|
|
93
|
+
(0, vitest_1.expect)(result.options).toBeUndefined();
|
|
94
|
+
});
|
|
82
95
|
(0, vitest_1.it)('should throw error for invalid arguments', () => {
|
|
83
96
|
(0, vitest_1.expect)(() => (0, createZodLlmClient_js_1.normalizeZodArgs)({})).toThrow('Invalid arguments');
|
|
84
97
|
});
|
package/dist/index.d.ts
CHANGED
package/dist/index.js
CHANGED
|
@@ -17,5 +17,6 @@ Object.defineProperty(exports, "__esModule", { value: true });
|
|
|
17
17
|
__exportStar(require("./createLlmClient.js"), exports);
|
|
18
18
|
__exportStar(require("./createLlmRetryClient.js"), exports);
|
|
19
19
|
__exportStar(require("./createZodLlmClient.js"), exports);
|
|
20
|
+
__exportStar(require("./createJsonSchemaLlmClient.js"), exports);
|
|
20
21
|
__exportStar(require("./llmFactory.js"), exports);
|
|
21
22
|
__exportStar(require("./retryUtils.js"), exports);
|
package/dist/llmFactory.d.ts
CHANGED
|
@@ -14,6 +14,8 @@ export declare function createLlm(params: CreateLlmFactoryParams): {
|
|
|
14
14
|
<T extends import("zod").ZodType>(prompt: string, schema: T, options?: import("./createZodLlmClient.js").ZodLlmClientOptions): Promise<boolean>;
|
|
15
15
|
<T extends import("zod").ZodType>(mainInstruction: string, userMessagePayload: string | import("openai/resources/index.js").ChatCompletionContentPart[], dataExtractionSchema: T, options?: import("./createZodLlmClient.js").ZodLlmClientOptions): Promise<boolean>;
|
|
16
16
|
};
|
|
17
|
+
promptJson: <T>(messages: import("openai/resources/index.js").ChatCompletionMessageParam[], schema: Record<string, any>, validator: (data: any) => T, options?: import("./createJsonSchemaLlmClient.js").JsonSchemaLlmClientOptions) => Promise<T>;
|
|
18
|
+
isPromptJsonCached: (messages: import("openai/resources/index.js").ChatCompletionMessageParam[], schema: Record<string, any>, options?: import("./createJsonSchemaLlmClient.js").JsonSchemaLlmClientOptions) => Promise<boolean>;
|
|
17
19
|
promptRetry: {
|
|
18
20
|
<T = import("openai/resources/index.js").ChatCompletion>(content: string, options?: Omit<import("./createLlmRetryClient.js").LlmRetryOptions<T>, "messages">): Promise<T>;
|
|
19
21
|
<T = import("openai/resources/index.js").ChatCompletion>(options: import("./createLlmRetryClient.js").LlmRetryOptions<T>): Promise<T>;
|
package/dist/llmFactory.js
CHANGED
|
@@ -4,18 +4,23 @@ exports.createLlm = createLlm;
|
|
|
4
4
|
const createLlmClient_js_1 = require("./createLlmClient.js");
|
|
5
5
|
const createLlmRetryClient_js_1 = require("./createLlmRetryClient.js");
|
|
6
6
|
const createZodLlmClient_js_1 = require("./createZodLlmClient.js");
|
|
7
|
+
const createJsonSchemaLlmClient_js_1 = require("./createJsonSchemaLlmClient.js");
|
|
7
8
|
function createLlm(params) {
|
|
8
9
|
const baseClient = (0, createLlmClient_js_1.createLlmClient)(params);
|
|
9
10
|
const retryClient = (0, createLlmRetryClient_js_1.createLlmRetryClient)({
|
|
10
11
|
prompt: baseClient.prompt
|
|
11
12
|
});
|
|
12
|
-
const
|
|
13
|
+
const jsonSchemaClient = (0, createJsonSchemaLlmClient_js_1.createJsonSchemaLlmClient)({
|
|
13
14
|
prompt: baseClient.prompt,
|
|
14
15
|
isPromptCached: baseClient.isPromptCached
|
|
15
16
|
});
|
|
17
|
+
const zodClient = (0, createZodLlmClient_js_1.createZodLlmClient)({
|
|
18
|
+
jsonSchemaClient
|
|
19
|
+
});
|
|
16
20
|
return {
|
|
17
21
|
...baseClient,
|
|
18
22
|
...retryClient,
|
|
23
|
+
...jsonSchemaClient,
|
|
19
24
|
...zodClient
|
|
20
25
|
};
|
|
21
26
|
}
|