Nick: max num tokens for llm extract (for now) + slice the max

apps/api/src/controllers/scrape.ts

@@ -106,6 +106,9 @@ export async function scrapeController(req: Request, res: Response) {
     const extractorOptions = req.body.extractorOptions ?? {
       mode: "markdown"
     }
+    if (extractorOptions.mode === "llm-extraction") {
+      pageOptions.onlyMainContent = true;
+    }
     const origin = req.body.origin ?? "api";
     const timeout = req.body.timeout ?? 30000; // Default timeout of 30 seconds
 

apps/api/src/lib/LLM-extraction/models.ts

@@ -1,25 +1,38 @@
 import OpenAI from "openai";
 import { Document } from "../../lib/entities";
+import { numTokensFromString } from "./helpers";
 
 export type ScraperCompletionResult = {
   data: any | null;
   url: string;
 };
 
+const maxTokens = 32000;
+const modifier = 4;
 const defaultPrompt =
   "You are a professional web scraper. Extract the contents of the webpage";
 
 function prepareOpenAIDoc(
   document: Document
-): OpenAI.Chat.Completions.ChatCompletionContentPart[] {
+): [OpenAI.Chat.Completions.ChatCompletionContentPart[], number] {
-  // Check if the markdown content exists in the document
+  let markdown = document.markdown;
-  if (!document.markdown) {
+
+// Check if the markdown content exists in the document
+  if (!markdown) {
     throw new Error(
       "Markdown content is missing in the document. This is likely due to an error in the scraping process. Please try again or reach out to help@mendable.ai"
     );
   }
 
-  return [{ type: "text", text: document.markdown }];
+  // count number of tokens
+  const numTokens = numTokensFromString(document.markdown, "gpt-4");
+
+  if (numTokens > maxTokens) {
+    // trim the document to the maximum number of tokens, tokens != characters
+    markdown = markdown.slice(0, (maxTokens * modifier));
+  }
+
+  return [[{ type: "text", text: markdown }], numTokens];
 }
 
 export async function generateOpenAICompletions({
@@ -38,7 +51,7 @@ export async function generateOpenAICompletions({
   temperature?: number;
 }): Promise<Document> {
   const openai = client as OpenAI;
-  const content = prepareOpenAIDoc(document);
+  const [content, numTokens] = prepareOpenAIDoc(document);
 
   const completion = await openai.chat.completions.create({
     model,
@@ -72,6 +85,7 @@ export async function generateOpenAICompletions({
   return {
     ...document,
     llm_extraction: llmExtraction,
+    warning: numTokens > maxTokens ? `Page was trimmed to fit the maximum token limit defined by the LLM model (Max: ${maxTokens} tokens, Attemped: ${numTokens} tokens). If results are not good, email us at help@mendable.ai so we can help you.` : undefined,
   };
 }
 

apps/api/src/lib/entities.ts

@@ -72,6 +72,7 @@ export class Document {
   };
   childrenLinks?: string[];
   provider?: string;
+  warning?: string;
 
   constructor(data: Partial<Document>) {
     if (!data.content) {

apps/api/src/scraper/WebScraper/utils/excludeTags.ts

@@ -34,8 +34,6 @@ export const excludeNonMainTags = [
   "#nav",
   ".breadcrumbs",
   "#breadcrumbs",
-  ".form",
-  "form",
   "#search-form",
   ".search",
   "#search",
@@ -51,10 +49,6 @@ export const excludeNonMainTags = [
   "#tag",
   ".category",
   "#category",
-  ".comment",
+  ".cookie",
-  "#comment",
+  "#cookie"
-  ".reply",
-  "#reply",
-  ".author",
-  "#author",
 ];