⚡ Custom Deep Research Bot | v2.0
Performs in-depth online research using your custom Google queries to answer your questions. First, create up to 4 specific Google searches to gather information, then ask your actual question to get a focused AI response based on those search results
Created by Morteza H.
Updated on Mar 31, 00:03
Installed 45 times
How to Use
- Get HARPA AI from https://get.harpa.ai
- Open HARPA AI with Alt+A on this page
- Click IMPORT COMMAND
Content
- type: say
message: >-
**Tip:** This command performs a thorough in-depth online search for a given
query or topic and is best used for research. It takes time and is token
hungry (~20k tokens on GPT-4o-mini per search depth).
label: TIP ABOUT COMMAND
- param: depth
message: >-
**Set the search depth.**
Search depth indicates how many search pages will be analyzed. Each depth
level generates an extra sub-query, improving search quality.
options:
- label: 🔍 SEARCH
value: 1
- label: 2 SUB-QUERIES
value: 2
- label: 3 SUB-QUERIES
value: 3
- label: 4 SUB-QUERIES
value: 4
optionsInvalid: false
type: ask
default: ''
label: SET DEPTH LEVEL
- message: >-
| Operator | Description | Example |
|----------|------------|----------|
| site: | Restricts results to a specific domain | site:arxiv.org
"transformer model" |
| -site: | Excludes a specific domain from results | AI research
-site:wikipedia.org |
| "..." | Forces exact phrase match | "machine learning architecture" |
| filetype: | Searches for specific file types | filetype:pdf "reinforcement
learning" |
| intitle: | Finds pages with specific term in title | intitle:AI benchmarks
|
| allintitle: | All terms must appear in title | allintitle:machine learning
python |
| intext: | Finds pages with specific term in body text | intext:neural
network topology |
| allintext: | All terms must appear in body text | allintext:GPT
architecture training |
| define: | Shows definition of a term | define:neural network |
| OR \| | Searches for either term | GPT-4 OR "large language model" |
| AND | Requires both terms to be present | transformer AND "attention
mechanism" |
| inurl: | Finds pages with specific term in URL |
inurl:artificial-intelligence |
| * | Wildcard for any word | "deep * model" |
| -term | Excludes a specific term | AI -chatbot |
| () | Groups search terms for complex queries | (GPT-4 OR LLaMA)
architecture |
| before: after: | Time-based filtering | AI breakthroughs after:2023 |
label: TIP ABOUT SEARCH OPERATOR
type: say
- param: userQuery1
message: >-
 Please enter {{depth}}
search queries that will be used to retrieve Google search results:
options: null
condition: '{{depth}} >= 1'
label: ASK FIRST SEARCH QUERY
type: ask
default: ''
- param: autoQueries.boolean
message: >-
Would you like to have similar sub-queries automatically generated by an AI
based on the first sub-query you specified?
options:
- label: ✅ YES
value: 'yes'
- label: ⛔ NO
value: 'no'
vision:
mode: area
send: true
enabled: false
hint: ''
label: AUTO QUERIES BOOLEAN
condition: '{{depth}} > 1'
type: ask
default: ''
optionsInvalid: false
- type: group
steps:
- param: userQuery2
condition: '{{depth}} >= 2'
message: ''
type: ask
options: null
default: ''
- param: userQuery3
condition: '{{depth}} >= 3'
message: ''
type: ask
options: null
default: ''
- param: userQuery4
condition: '{{depth}} >= 4'
message: ''
type: ask
options: null
default: ''
label: ASK SEARCH QUERIES
condition: '{{autoQueries.boolean}} = no'
- param: searchQuery
message: >-
Please enter your specific search query now. This query will be used as the
basis for the AI-generated response. The more precisely you phrase your
question or request, the better the AI can address it. What would you like
to know?
vision:
enabled: false
mode: area
hint: ''
send: true
label: THE REAL AND SPECIFIC QUERY FROM USER
type: ask
options: null
default: ''
- steps:
- prompt: >-
You are a Research Agent. I need you to break down my search query into
{{depth}} clarifying queries in {{language}}, the answers to which are
necessary to form a detailed response to my query: "{{userQuery1}}"
Follow the instructions:
- The first sub-query should be as close as possible to My Query in
wording. If my query is short and simple - repeat it in the first
sub-query.
- Subsequent sub-queries should clarify the first one by focusing on
different aspects or using alternative wordings.
- Each sub-query should include keywords from [my search query]. For
example, if asking about a product, include its name, brand, or key
features in all sub-queries to find relevant info.
- Use Google search operators like "OR", "AND", "site:", "-",
"filetype:", "intitle:", "intext:", "inurl:", and parentheses to refine
the sub-queries and get the best search results. For example:
- Use "OR" to include synonyms or related terms
- Use quotation marks for exact phrases
- Use parentheses to group search terms for complex queries
- Use "site:" or "-site:" to include or exclude specific domains (sparingly and only when relevant)
- Be cautious with the "-" and "-site:" operator to avoid excluding with the minus operator potentially relevant results
- If My Query is saying to exclude sites or information, then USE the
minus operator like "-site:" or "-term" IN EVERY query. If they say
something like "No site X and/or site Y", then use "-site:" for both
sites in every query.
- Configure your Google search operators to maximize relevant results.
Use common operators and avoid overly restrictive combinations that
could yield zero matches.
- Sub-queries should be brief. Imagine typing a brief Google query for a
precise answer.
- Do not repeat my command and instructions.
- Response in JSON array with {{depth}} objects (queries).
- Write nothing other than the JSON.
Example input: "I want to buy a cordless vacuum cleaner under $500,
bagless, only from brand XY, and only handheld models that are long and
don't require bending down."
Example JSON response:
[
{
"query": cordless vacuum cleaner (under OR less than OR up to) $500 (from OR by) brand XY
},
{
"query": handheld vacuum brand XY (long OR no bending)
},
{
"query": vacuum cleaner bagless (under OR up to) $500 (brand OR manufacturer) XY
},
{
"query": XY handheld vacuum (bagless OR without bags) (long OR ergonomic OR no bending) (price OR cost) under 500
}
]
JSON response:
param: subqueries
condition: '{{autoQueries.boolean}} = yes'
type: gpt
label: CREATE SUBQUERIES
silent: true
dumb: false
isolated: true
- prompt: >-
You are a Research Agent. I need you to convert my search queries below
into JSON array in language: {{language}}:
1. "{{userQuery1}}"
2. "{{userQuery2}}"
3. "{{userQuery3}}"
4. "{{userQuery4}}"
Follow the instructions:
- Put my queries in a JSON array.
- Do not repeat my command and instructions.
- Response in JSON array with {{depth}} objects (queries).
- Write nothing other than the JSON.
- DO NOT change anything in the queries provided above.
- ONLY use the queries shown above, not the ones from the example.
- DO not create more than {{depth}} queries in JSON Array.
Example JSON response:
[
{
"query": "Steam engine history and industrial impact"
},
{
"query": "First practical steam engine inventor and date"
}
]
JSON response:
label: CONVERT USER QUERIES TO JSON
param: subqueries
condition:
- '{{autoQueries.boolean}} = no'
- '{{autoQueries.boolean}} ='
type: gpt
silent: true
dumb: false
isolated: true
- type: calc
func: extract-json
to: subqueries
param: subqueries
index: ''
- message: >-
🧩 To give a comprehensive response, your question was split into
**{{subqueries.length}}** sub-queries.
type: say
- steps:
- value: '{{item.query}}'
func: set
param: query
format: text
type: calc
- message: '⏳ Scanning information for sub-query: "**{{query}}**".'
type: say
- param: information
value: '{{serp {{query}}}}'
type: calc
func: set
format: ''
- func: serp.extract-links
to: links
type: calc
from: information
- param: total
value: '{{links.length}}'
format: number
type: calc
func: set
- steps:
- value: '{{page {{item.url}}}}'
param: content
type: calc
func: set
format: ''
- type: js
code: |-
const content = args['content'];
return {
chars: content.length,
estimatedTokens: Math.ceil(content.length / 4),
estimatedWords: Math.ceil(content.length / 4 * 0.75)
};
param: count
timeout: 15000
onFailure: SAY STATUS
label: COUNT
args: content
silent: true
- param: pageUrl
value: '{{item.url}}'
type: calc
func: set
format: text
- code: |-
const regex = /^(?:https?:\/\/)?(?:www\.)?([^\/]+)/;
const testString = pageUrl;
const matches = testString.match(regex);
if (matches) {
const hostname = matches[1];
return hostname;
}
param: hostname
label: HOSTNAME
type: js
args: pageUrl
timeout: 15000
silent: true
- func: increment
param: index
type: calc
delta: 1
- args: index, total
code: |-
function calculatePercentage(index, total) {
index = Number(index);
total = Number(total);
if (isNaN(index) || isNaN(total)) {
return "Error";
}
return (index / total * 100).toFixed(1) + "%";
}
let result = calculatePercentage(index, total);
return result;
param: percentage
label: PERCENTAGE INDEX TOTAL
type: js
timeout: 15000
silent: true
- steps:
- steps:
- prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize
useful information to answer my query: "{{searchQuery}}"
Instructions:
- Imagine you are a Google Search engine, collecting
relevant information from various websites to form an
answer to my query in {{language}}.
- Focus on providing detailed factual information,
statistics, and specific examples when available.
- Gather all information that will be useful in
answering my question, including context, numbers, dates
and concrete details.
- Extract as much useful information as possible from
the page content.
- Avoid general phrases.
- When referring to someone's opinion about something,
mention the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant
information found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field:
"info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{content}}.
JSON response:
param: data
type: gpt
isolated: true
silent: true
- index: first
type: calc
func: extract-json
to: data
param: data
- param: data.url
format: auto
type: calc
func: set
value: '{{item.url}}'
- param: data.title
value: '{{item.title}}'
type: calc
func: set
format: auto
- list: array
func: list-add
index: last
type: calc
item: data
- type: jump
to: SAY STATUS
condition: '{{content}} =~ ^[\s\S]{1000,}$'
label: FETCHED
type: group
label: IS IT REDDIT?
condition: >-
{{hostname}} =~
^(?!.*(reddit\.com|redd\.it)).*(?:https?:\/\/|www\.)?[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}.*$
type: group
- steps:
- args: item.url
code: >-
// Get URL from args parameter
const url = args['item.url'];
if (!url) {
return false;
}
const fullUrl = url.startsWith('http') ? url : 'https://' +
url;
window.location.href = fullUrl;
return true;
param: navigate.boolean
onFailure: ABORT NAVIGATE
label: NAVIGATE
type: js
timeout: 15000
silent: true
- type: wait
for: idle
timeout: 6000
- type: wait
for: 2s
silent: true
label: ABORT NAVIGATE
- steps:
- param: setThread
value: '{{thread}}'
label: SET CONTENT
type: calc
func: set
format: ''
- code: |-
// For setThread
const text = args['setThread'];
const tokenCount = Math.ceil(text.length / 4);
return {
chars: text.length,
estimatedTokens: tokenCount,
estimatedWords: Math.ceil(tokenCount * 0.75)
};
onFailure: SOCIAL-MEDIA PROMPT
type: js
args: setThread
param: count
timeout: 15000
silent: true
label: COUNT
- prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize
useful information to answer my query: "{{searchQuery}}"
Instructions:
- Imagine you are a Google Search engine, collecting
relevant information from various websites to form an
answer to my query in {{language}}.
- Focus on providing detailed factual information,
statistics, and specific examples when available.
- Gather all information that will be useful in
answering my question, including context, numbers, dates
and concrete details.
- Extract as much useful information as possible from
the page content.
- Avoid general phrases.
- When referring to someone's opinion about something,
mention the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant
information found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field:
"info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{setThread}}.
JSON response:
type: gpt
isolated: true
param: data
silent: true
label: SOCIAL-MEDIA PROMPT
label: SOCIAL-MEDIA
condition: >-
{{hostname}} =~
^(?:reddit|facebook|twitter|telegram|discord|whatsapp)\.com$
type: group
- steps:
- param: setPageThread
value: '{{page}} {{thread}}'
type: calc
func: set
format: ''
label: SET CONTENT
- code: |-
// For setPageThread
const text = args['setPageThread'];
const tokenCount = Math.ceil(text.length / 4);
return {
chars: text.length,
estimatedTokens: tokenCount,
estimatedWords: Math.ceil(tokenCount * 0.75)
};
onFailure: NOT-SOCIAL-MEDIA PROMPT
type: js
args: setPageThread
param: count
timeout: 15000
silent: true
label: COUNT
- prompt: >-
Act as a professional researcher.
Analyze the given [web page content] and summarize
useful information to answer my query: "{{searchQuery}}"
Instructions:
- Imagine you are a Google Search engine, collecting
relevant information from various websites to form an
answer to my query in {{language}}.
- Focus on providing detailed factual information,
statistics, and specific examples when available.
- Gather all information that will be useful in
answering my question, including context, numbers, dates
and concrete details.
- Extract as much useful information as possible from
the page content.
- Avoid general phrases.
- When referring to someone's opinion about something,
mention the author. Quote if appropriate.
- If you haven't found anything useful, don't make up
information or guess - simply say "No relevant
information found" in {{language}}.
- Do not echo my prompt in your response.
- Respond with a JSON object containing a single field:
"info".
- Write nothing other than the JSON.
Example JSON response:
{
"info": "Write here textual information that would help answer my query. Write as much as you think is necessary, focusing on providing useful information for further processing."
}
[Web page content]: {{setPageThread}}.
JSON response:
type: gpt
isolated: true
param: data
silent: true
label: NOT-SOCIAL-MEDIA PROMPT
condition: >-
{{hostname}} =~
^(?!(?:reddit|facebook|twitter|telegram|discord|whatsapp)\.com)[a-zA-Z0-9-]+\.[a-zA-Z]+$
label: NOT-SOCIAL-MEDIA
type: group
- type: calc
func: extract-json
to: data
param: data
index: first
- type: calc
func: set
param: data.url
format: auto
value: '{{item.url}}'
- type: calc
func: set
param: data.title
format: auto
value: '{{item.title}}'
- type: calc
func: list-add
index: last
list: array
item: data
label: NOT FETCHED
type: group
- message: >-
🔍 Analyzed **{{index}} / {{total}}**,
[{{hostname}}]({{pageUrl}})
- Chars: {{count.chars}}
- Estimated Tokens: {{count.estimatedTokens}}
- Estimated Words: {{count.estimatedWords}}
label: ⛔️SAY STATUS
condition: '{{index}} = {{total1}}'
type: say
- message: |-
🔍 Analyzed **{{percentage}}**, [{{hostname}}]({{pageUrl}})
- Chars: {{count.chars}}
- Estimated Tokens: {{count.estimatedTokens}}
- Estimated Words: {{count.estimatedWords}}
condition: '{{index}} != {{total}}'
type: say
label: SAY STATUS
- code: |-
const content = args['array'];
const stringContent = JSON.stringify(content);
return {
chars: stringContent.length,
estimatedTokens: Math.ceil(stringContent.length / 4),
estimatedWords: Math.ceil(stringContent.length / 4 * 0.75)
};
param: arrayLength
label: ARRAY LENGTH
type: js
args: array
timeout: 15000
onFailure: SAY STATUS
silent: true
- message: |-
✅ **{{percentage}}** pages scanned for "**{{query}}**".
Last checked page: [{{hostname}}]({{pageUrl}})
- Chars: {{count.chars}}
- Estimated Tokens: {{count.estimatedTokens}}
- Estimated Words: {{count.estimatedWords}}
- Array Token length: **{{arrayLength.estimatedTokens}}**
condition: '{{index}} = {{total}}'
type: say
label: SAY STATUS
type: loop
list: links
type: loop
list: subqueries
- message: >-
🤖 Analyzed **{{array.length}}** pages from **{{subqueries.length}}**
Google searches.
Now preparing a comprehensive answer to your question.
type: say
label: DEPTH >= 1
type: group
condition: '{{depth}} >= 1'
- prompt: >-
Please ignore all previous instructions. I want you to only respond in
{{language}}.
You are a Research AI Agent.
I would like you create a comprehensive answer to [MY QUESTION], adhering to
a specific format I provide:
- Utilize [Information from the Web], provided below.
- Extract as much useful information as possible from the web search results
- Integrate insights from multiple web search sources, and ensure inclusion
of relevant links in a markdown fully rendered format: [➊](URL), [➋](URL)
etc, to each bullet point.
- Every URL should correspond to only one symbol: ➊ ➋ ➌ ➍ ➎ ➏ ➐ ➑ ➒ ➓.
- If multiple sources repeat the same information, describe it and cite all
sources at once.
- If you know the answer, add your own knowledge to make it more complete.
- Do not hallucinate facts or information.
- Do not use any other tools.
- Avoid general phrases and be more specific and detailed.
Follow the [RESPONSE FORMAT]:
## Key takeaway:
Provide a single, most important takeaway from the web search results in
{{language}}.
## Detailed answer:
Analyze and present detailed information that helps answer my question.
There should be no limit in words or bullet points to the report. Ensure
that all ideas, facts, and relevant information are concisely reported, and
the answer is comprehensive. Incorporate the maximum number of source links
within the text.
[MY QUESTION]: {{searchQuery}}
[Information from the Web]:
{{array}}
[COMPREHENSIVE RESPONSE WITH SOURCE LINKS ([➊](URL))]:
label: FINAL GPT
type: gpt
isolated: true
param: gpt
dumb: false
- type: js
args: array
code: |-
function formatMarkdownLinks(args) {
const text = args.array || "";
if (!text) return "No text provided to extract links from.";
// Regex pattern für Links mit ➊, ➋, ➌ etc.
const linkPattern = /\[(➊|➋|➌|➍|➎|➏|➐|➑|➒|➓|⓫|⓬|⓭|⓮|⓯|⓰|⓱|⓲|⓳|⓴|㉑|㉒|㉓|㉔|㉕|㉖|㉗|㉘|㉙|㉚|㉛|㉜|㉝|㉞|㉟|㊱|㊲|㊳|㊴|㊵)\]\(([^)]+)\)/g;
// Erweiterte Liste von Kreiszahlen
const circleNumbers = "➊➋➌➍➎➏➐➑➒➓⓫⓬⓭⓮⓯⓰⓱⓲⓳⓴㉑㉒㉓㉔㉕㉖㉗㉘㉙㉚㉛㉜㉝㉞㉟㊱㊲㊳㊴㊵".split("");
// Map zur Speicherung eindeutiger Links nach Symbol und URL
const uniqueLinks = new Map();
// Alle Übereinstimmungen finden
let match;
let currentIndex = 0;
while ((match = linkPattern.exec(text)) !== null) {
try {
const symbol = match[1];
const url = match[2];
// URL-Objekt erstellen zur Validierung und Analyse
const urlObj = new URL(url);
// Domain und Pfad extrahieren (ohne www.)
const fullPath = urlObj.hostname.replace('www.', '') + urlObj.pathname;
// Kürzen, wenn länger als 42 Zeichen
const shortPath = fullPath.length > 42
? fullPath.slice(0, 42) + '...'
: fullPath.replace(/\/$/, '');
// Erstelle einen Key, der Symbol und URL kombiniert
const key = `${symbol}:${url}`;
// Jede Übereinstimmung wird mit einem eindeutigen Symbol und URL-Kombination hinzugefügt
if (!uniqueLinks.has(key)) {
// Verwende das vorhandene Symbol aus dem Text
uniqueLinks.set(key, `[${symbol}](${url}) [${shortPath}](${url})`);
currentIndex++;
}
} catch (e) {
// Ungültige URLs überspringen
}
}
// Wenn keine formatierten Links gefunden wurden, nach einfachen URLs suchen
if (uniqueLinks.size === 0) {
const urlPattern = /(https?:\/\/[^\s\)\]\'",:;]+)/g;
let urlMatch;
currentIndex = 0;
while ((urlMatch = urlPattern.exec(text)) !== null) {
try {
const url = urlMatch[1].replace(/[.,;:?!]+$/, ''); // URL bereinigen
// URL-Objekt erstellen zur Validierung und Analyse
const urlObj = new URL(url);
// Domain und Pfad extrahieren (ohne www.)
const fullPath = urlObj.hostname.replace('www.', '') + urlObj.pathname;
// Kürzen, wenn länger als 42 Zeichen
const shortPath = fullPath.length > 42
? fullPath.slice(0, 42) + '...'
: fullPath.replace(/\/$/, '');
// Stoppe, wenn wir mehr URLs haben als Kreiszahlen
if (currentIndex >= circleNumbers.length) break;
// Jede URL wird mit einem eigenen Symbol hinzugefügt, auch wenn sie mehrfach vorkommt
const key = `${currentIndex}:${url}`;
uniqueLinks.set(key, `[${circleNumbers[currentIndex]}](${url}) [${shortPath}](${url})`);
currentIndex++;
} catch (e) {
// Ungültige URLs überspringen
}
}
}
// Formatierte Ausgabe zurückgeben oder eine Nachricht, wenn keine Links gefunden wurden
return uniqueLinks.size > 0
? Array.from(uniqueLinks.values()).join('\n')
: "No numbered reference links found in the text.";
}
return formatMarkdownLinks(args);
param: sources
timeout: 15000
onFailure: DISPLAY SOURCES
label: EXTRACT SOURCES
- type: say
message: |-
 Sources
{{sources}}
label: DISPLAY SOURCES
Notice: Please read before using
This automation command is created by a community member. HARPA AI team does not audit community commands.
Please review the command carefully and only install if you trust the creator.
Contact us
All rights reserved © HARPA AI TECHNOLOGIES LLC, 2021 — 2025
Designed and engineered in Finland 🇫🇮
CAN WE STORE COOKIES?
Our website uses cookies for the purposes of accessibility and security. They also allow us to gather statistics in order to improve the website for you. More info: Privacy Policy