chore: first commit

src/agent.ts

@@ -7,6 +7,7 @@ import {SafeSearchType, search} from "duck-duck-scrape";
 import {rewriteQuery} from "./tools/query-rewriter";
 import {dedupQueries} from "./tools/dedup";
 import {evaluateAnswer} from "./tools/evaluator";
+import {buildURLMap} from "./tools/getURLIndex";
 
 // Proxy setup remains the same
 if (process.env.https_proxy) {
@@ -164,32 +165,55 @@ function getSchema(allowReflect: boolean): ResponseSchema {
   };
 }
 
-function getPrompt(question: string, context?: any[], allQuestions?: string[], allowReflect: boolean = false, badContext?: any[]) {
+function getPrompt(question: string, context?: any[], allQuestions?: string[], allowReflect: boolean = false, badContext?: any[], knowledge?: any[], allURLs?: Record<string, string>) {
+  // console.log('Context:', context);
+  // console.log('All URLs:', JSON.stringify(allURLs, null, 2));
+
+  const knowledgeIntro = knowledge?.length ?
+    `
+## Knowledge
+You have successfully gathered some knowledge from the following questions:
+
+${JSON.stringify(knowledge, null, 2)}
+
+`
+    : '';
+
   const badContextIntro = badContext?.length ?
-    `Your last unsuccessful answer contains these previous actions and knowledge:
+    `
-    ${JSON.stringify(badContext, null, 2)}
+## Unsuccessful Answer Analysis
+Your last unsuccessful answer are:
+
+${JSON.stringify(badContext, null, 2)}
     
-    Learn to avoid these mistakes and think of a new approach, from a different angle, e.g. search for different keywords, read different URLs, or ask different questions.
+Learn to avoid these mistakes and think of a new approach, from a different angle, e.g. search for different keywords, read different URLs, or ask different questions.
     `
     : '';
 
   const contextIntro = context?.length ?
-    `Your current context contains these previous actions and knowledge:
-    ${JSON.stringify(context, null, 2)}
     `
+## Context
+You have conducted the following actions:
+
+${JSON.stringify(context, null, 2)}
+
+`
     : '';
 
   let actionsDescription = `
-Using your training data and prior lessons learned, answer the following question with absolute certainty:
+## Actions
-
-${question}
 
 When uncertain or needing additional information, select one of these actions:
 
+${allURLs ? `
 **read**:
-- Access external URLs to gather more information
+- Access any URLs from below to gather external knowledge
+
+${JSON.stringify(allURLs, null, 2)}
+
 - When you have enough search result in the context and want to deep dive into specific URLs
-- It allows you access the full content behind specific URLs
+- It allows you access the full content behind any URLs
+` : ''}
 
 **search**:
 - Query external sources using a public search engine
@@ -202,7 +226,8 @@ When uncertain or needing additional information, select one of these actions:
 ${allowReflect ? `- If doubts remain, use "reflect" instead` : ''}`;
 
   if (allowReflect) {
-    actionsDescription += `\n\n**reflect**:
+    actionsDescription += `
+**reflect**:
 - Perform critical analysis through hypothetical scenarios or systematic breakdowns
 - Identify knowledge gaps and formulate essential clarifying questions
 - Questions must be:
@@ -210,14 +235,19 @@ ${allowReflect ? `- If doubts remain, use "reflect" instead` : ''}`;
   - Focused on single concepts
   - Under 20 words
   - Non-compound/non-complex
-`;
+`.trim();
   }
 
   return `
-You are an advanced AI research analyst specializing in multi-step reasoning.
+You are an advanced AI research analyst specializing in multi-step reasoning. Using your training data and prior lessons learned, answer the following question with absolute certainty:
+
+## Question
+${question}
 
 ${contextIntro.trim()}
 
+${knowledgeIntro.trim()}
+
 ${badContextIntro.trim()}
 
 ${actionsDescription.trim()}
@@ -240,7 +270,7 @@ async function getResponse(question: string, tokenBudget: number = 30000000) {
   let allKeywords = [];
   let allKnowledge = [];  // knowledge are intermedidate questions that are answered
   let badContext = [];
-
+  let allURLs: Record<string, string> = {};
   while (totalTokens < tokenBudget) {
     // add 1s delay to avoid rate limiting
     await sleep(1000);
@@ -249,7 +279,16 @@ async function getResponse(question: string, tokenBudget: number = 30000000) {
     console.log('Gaps:', gaps)
     const allowReflect = gaps.length <= 1;
     const currentQuestion = gaps.length > 0 ? gaps.shift()! : question;
-    const prompt = getPrompt(currentQuestion, context, allQuestions, allowReflect, badContext);
+    // update all urls with buildURLMap
+    allURLs = {...allURLs, ...buildURLMap(context)};
+    const prompt = getPrompt(
+      currentQuestion,
+      context,
+      allQuestions,
+      allowReflect,
+      badContext,
+      allKnowledge,
+      allURLs);
     console.log('Prompt len:', prompt.length)
 
     const model = genAI.getGenerativeModel({
@@ -283,7 +322,11 @@ async function getResponse(question: string, tokenBudget: number = 30000000) {
         if (evaluation.is_valid_answer) {
           return action;
         } else {
-          badContext.push({...context, "Why this is a bad answer?": evaluation.reasoning});
+          badContext.push({
+            question: currentQuestion,
+            answer: action.answer,
+            "Why this is a bad answer?": evaluation.reasoning
+          });
           context = [];
         }
       } else if (evaluation.is_valid_answer) {
@@ -296,9 +339,18 @@ async function getResponse(question: string, tokenBudget: number = 30000000) {
       if (allQuestions.length) {
         newGapQuestions = await dedupQueries(newGapQuestions, allQuestions)
       }
-      gaps.push(...newGapQuestions);
+      if (newGapQuestions.length > 0) {
-      allQuestions.push(...newGapQuestions);
+        gaps.push(...newGapQuestions);
-      gaps.push(question);  // always keep the original question in the gaps
+        allQuestions.push(...newGapQuestions);
+        gaps.push(question);  // always keep the original question in the gaps
+      } else {
+        console.log('No new questions to ask');
+        context.push({
+          step,
+          ...action,
+          result: 'I have tried all possible questions and found no useful information. I must think out of the box or different angle!!!'
+        });
+      }
     }
 
     // Rest of the action handling remains the same
@@ -310,27 +362,37 @@ async function getResponse(question: string, tokenBudget: number = 30000000) {
         if (allKeywords.length) {
           keywordsQueries = await dedupQueries(keywordsQueries, allKeywords)
         }
-        const searchResults = [];
+        if (keywordsQueries.length > 0) {
-        for (const query of keywordsQueries) {
+          const searchResults = [];
-          const results = await search(query, {
+          for (const query of keywordsQueries) {
-            safeSearch: SafeSearchType.STRICT
+            console.log('Searching:', query);
-          });
+            const results = await search(query, {
-          const minResults = results.results.map(r => ({
+              safeSearch: SafeSearchType.STRICT
-            title: r.title,
+            });
-            url: r.url,
+            const minResults = results.results.map(r => ({
-            description: r.description,
+              title: r.title,
-          }));
+              url: r.url,
-          searchResults.push({query, results: minResults});
+              description: r.description,
-          allKeywords.push(query);
+            }));
-          await sleep(5000);
+            searchResults.push({query, results: minResults});
-        }
+            allKeywords.push(query);
+            await sleep(5000);
+          }
 
-        context.push({
+          context.push({
-          step,
+            step,
-          question: currentQuestion,
+            question: currentQuestion,
-          ...action,
+            ...action,
-          result: searchResults
+            result: searchResults
-        });
+          });
+        } else {
+          console.log('No new queries to search');
+          context.push({
+            step,
+            ...action,
+            result: 'I have tried all possible queries and found no new information. I must think out of the box or different angle!!!'
+          });
+        }
       } else if (action.action === 'read' && action.URLTargets?.length) {
         const urlResults = await Promise.all(
           action.URLTargets.map(async (url: string) => {

src/tools/dedup.ts

@@ -101,7 +101,7 @@ export async function dedupQueries(newQueries: string[], existingQueries: string
     const result = await model.generateContent(prompt);
     const response = await result.response;
     const json = JSON.parse(response.text()) as DedupResponse;
-    console.log('Analysis:', json);
+    console.log('Dedup:', json);
     return json.unique_queries;
   } catch (error) {
     console.error('Error in deduplication analysis:', error);

src/tools/evaluator.ts

@@ -45,7 +45,7 @@ const genAI = new GoogleGenerativeAI(apiKey);
 const model = genAI.getGenerativeModel({
   model: modelName,
   generationConfig: {
-    temperature: 0.1,
+    temperature: 0,
     responseMimeType: "application/json",
     responseSchema: responseSchema
   }
@@ -59,7 +59,7 @@ Core Evaluation Criteria:
 2. Clarity: Answer should be clear and unambiguous
 3. Informativeness: Answer must provide substantial, useful information
 4. Specificity: Generic or vague responses are not acceptable
-5. Definitiveness: "I don't know" or highly uncertain responses are not valid
+5. Definitiveness: "I don't know", "lack of information" or highly uncertain responses are not valid
 6. Relevance: Answer must be directly related to the question topic
 7. Accuracy: Information provided should be factually sound (if verifiable)
 
@@ -79,6 +79,13 @@ Evaluation: {
   "reasoning": "The answer is comprehensive, specific, and covers all key system requirements across different operating systems. It provides concrete numbers and necessary additional components."
 }
 
+Question: "what is the twitter account of jina ai's founder?"
+Answer: "The provided text does not contain the Twitter account of Jina AI's founder."
+Evaluation: {
+  "is_valid_answer": false,
+  "reasoning": "The answer is not definitive and fails to provide the requested information. Don't know, can't derive, lack of information is unacceptable,"
+}
+
 Now evaluate this pair:
 Question: ${JSON.stringify(question)}
 Answer: ${JSON.stringify(answer)}`;

src/tools/getURLIndex.ts

@@ -0,0 +1,39 @@
+interface SearchResult {
+  title: string;
+  url: string;
+  description: string;
+}
+
+interface QueryResult {
+  query: string;
+  results: SearchResult[];
+}
+
+interface StepData {
+  step: number;
+  question: string;
+  action: string;
+  reasoning: string;
+  searchQuery?: string;
+  result?: QueryResult[];
+}
+
+export function buildURLMap(data: StepData[]): Record<string, string> {
+  const urlMap: Record<string, string> = {};
+
+  data.forEach(step => {
+    if (step.result && Array.isArray(step.result)) {
+      step.result.forEach(queryResult => {
+        if (queryResult.results && Array.isArray(queryResult.results)) {
+          queryResult.results.forEach(result => {
+            if (!urlMap[result.url]) {
+              urlMap[result.url] = `${result.title}`;
+            }
+          });
+        }
+      });
+    }
+  });
+
+  return urlMap;
+}

src/tools/query-rewriter.ts

@@ -111,6 +111,7 @@ export async function rewriteQuery(query: string): Promise<string[]> {
     const result = await model.generateContent(prompt);
     const response = await result.response;
     const json = JSON.parse(response.text()) as KeywordsResponse;
+    console.log('Rewriter:', json)
     return json.keywords;
   } catch (error) {
     console.error('Error in query rewriting:', error);