From 948f7866df3c3c8d9f8be0a055de25795a5dee5f Mon Sep 17 00:00:00 2001
From: Rafael Miller <150964962+rafaelsideguide@users.noreply.github.com>
Date: Thu, 30 Jan 2025 16:11:55 -0300
Subject: [PATCH] Added instructions for empty string to extract prompts
 (#1114)

* added instructions to extract prompts

* getting back the null for not strings
---
 apps/api/src/lib/extract/build-prompts.ts            | 2 +-
 apps/api/src/lib/extract/completions/singleAnswer.ts | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/apps/api/src/lib/extract/build-prompts.ts b/apps/api/src/lib/extract/build-prompts.ts
index 05f5bf6f..6d89eb2f 100644
--- a/apps/api/src/lib/extract/build-prompts.ts
+++ b/apps/api/src/lib/extract/build-prompts.ts
@@ -97,7 +97,7 @@ export function buildBatchExtractSystemPrompt(
 ): string {
   return (
     (systemPrompt ? `${systemPrompt}\n` : "") +
-    `Always prioritize using the provided content to answer the question. Do not make up an answer. Do not hallucinate. Be concise and follow the schema always if provided. If the document provided is not relevant to the prompt nor to the final user schema ${JSON.stringify(multiEntitySchema)}, return null. Here are the urls the user provided of which he wants to extract information from: ` +
+    `Always prioritize using the provided content to answer the question. Do not make up an answer. Do not hallucinate. In case you can't find the information and the string is required, instead of 'N/A' or 'Not speficied', return an empty string: '', if it's not a string and you can't find the information, return null. Be concise and follow the schema always if provided. If the document provided is not relevant to the prompt nor to the final user schema ${JSON.stringify(multiEntitySchema)}, return null. Here are the urls the user provided of which he wants to extract information from: ` +
     links.join(", ")
   );
 }
diff --git a/apps/api/src/lib/extract/completions/singleAnswer.ts b/apps/api/src/lib/extract/completions/singleAnswer.ts
index 0f32d8c6..df66263a 100644
--- a/apps/api/src/lib/extract/completions/singleAnswer.ts
+++ b/apps/api/src/lib/extract/completions/singleAnswer.ts
@@ -26,7 +26,7 @@ export async function singleAnswerCompletion({
       mode: "llm",
       systemPrompt:
         (systemPrompt ? `${systemPrompt}\n` : "") +
-        "Always prioritize using the provided content to answer the question. Do not make up an answer. Do not hallucinate. Return 'null' the property that you don't find the information. Be concise and follow the schema always if provided. Here are the urls the user provided of which he wants to extract information from: " +
+        "Always prioritize using the provided content to answer the question. Do not make up an answer. Do not hallucinate. In case you can't find the information and the string is required, instead of 'N/A' or 'Not speficied', return an empty string: '', if it's not a string and you can't find the information, return null. Be concise and follow the schema always if provided. Here are the urls the user provided of which he wants to extract information from: " +
         links.join(", "),
       prompt: "Today is: " + new Date().toISOString() + "\n" + prompt,
       schema: rSchema,