fix: handle typescript json format when mistralai returns this

package.json

@@ -116,6 +116,7 @@
     "i18next": "^22.4.5",
     "i18next-browser-languagedetector": "^7.0.1",
     "jsdom": "^21.1.1",
+    "jsonic": "^1.0.1",
     "langchain": "^0.1.7",
     "linkify-it": "^5.0.0",
     "locks": "^0.2.2",
@@ -202,6 +203,7 @@
     "@types/http-errors": "^2.0.3",
     "@types/jest": "^29.4.0",
     "@types/jsdom": "^21.1.1",
+    "@types/jsonic": "^0.3.3",
     "@types/linkify-it": "^3.0.5",
     "@types/locks": "^0.2.1",
     "@types/lodash.debounce": "^4.0.7",

src/features/llm-langchain.ts

@@ -9,6 +9,7 @@ import { CronJob } from 'cron';
 import { minutesToMilliseconds } from 'date-fns/minutesToMilliseconds';
 import { subHours } from 'date-fns/subHours';
 import fs from 'fs/promises';
+import jsonic from 'jsonic';
 import { LLMChain } from 'langchain/chains';
 import { createStuffDocumentsChain } from 'langchain/chains/combine_documents';
 import { createRetrievalChain } from 'langchain/chains/retrieval';
@@ -17,7 +18,12 @@ import mistralTokenizer from 'mistral-tokenizer-js';
 import path from 'path';
 import { z } from 'zod';
 
-import { ChunkEventEmitter, LlmManager, extractFirstJsonCodeContentFromMarkdown } from '@etabli/src/features/llm';
+import {
+  ChunkEventEmitter,
+  LlmManager,
+  extractFirstJsonCodeContentFromMarkdown,
+  extractFirstTypescriptCodeContentFromMarkdown,
+} from '@etabli/src/features/llm';
 import { gptInstances, gptSeed } from '@etabli/src/gpt';
 import { DocumentInitiativeTemplateSchema, ResultSchema, ResultSchemaType } from '@etabli/src/gpt/template';
 import { tokensReachTheLimitError } from '@etabli/src/models/entities/errors';
@@ -470,6 +476,20 @@ CONTEXT:
 
         throw new Error(`the json code block is not present in the answer or the answer has been truncated while saying it's complete`);
       }
+    } else if (answer.text.includes('```ts')) {
+      const typescriptCode = extractFirstTypescriptCodeContentFromMarkdown(answer.text);
+
+      if (!typescriptCode) {
+        console.log(answer.text);
+
+        throw new Error(`the typescript code block is not present in the answer or the answer has been truncated while saying it's complete`);
+      }
+
+      // That's the pattern MistralAI seems to always provide when returning TypeScript format
+      const jsonStringNotStrict = typescriptCode.replace('type ResultSchemaType =', '').trim();
+
+      // A JSON object in TypeScript cannot be parsed due to missing quotes on properties, ending comma... so using a helper for this
+      jsonString = JSON.stringify(jsonic(jsonStringNotStrict));
     }
 
     if (!jsonString) {

src/features/llm.ts

@@ -34,6 +34,13 @@ export function extractFirstJsonCodeContentFromMarkdown(markdown: string): strin
   return !!regexResult ? regexResult[1] : null;
 }
 
+export function extractFirstTypescriptCodeContentFromMarkdown(markdown: string): string | null {
+  const regex = /```ts\n([\s\S]+?)\n```/;
+  const regexResult = regex.exec(markdown);
+
+  return !!regexResult ? regexResult[1] : null;
+}
+
 export async function initLlmSystem() {
   // Note: Prisma does not implement yet locking table though it should help not messing with requesting the LLM system while replacing sensitive components of it
   // This race condition should remain rare and having error thrown should be fine since replayed on the next iteration