n8nauto
Get Started

Autonomous Deep Research Agent with OpenRouter, SerpAPI and Jina AI

This template builds an autonomous research agent inside n8n: you type a question into a chat window, and the workflow returns a structured, multi-source research report. It is aimed at analysts, founders and content teams who would otherwise spend an hour Googling, copy-pasting articles and asking an LLM to summarise them. Under the hood it chains an OpenRouter LLM, a SerpAPI Google search, the Jina AI Reader for full page extraction, and a Wikipedia tool, all orchestrated by two LangChain agents with conversation memory.

How It Works

The flow starts at the Chat Message Trigger, which exposes a chat UI inside n8n. Whatever the user asks is passed straight into Generate Search Queries using LLM, a Basic LLM Chain wired to LLM Response Provider (OpenRouter). The prompt instructs the model to break the research question into a JSON array of focused Google queries.

Parse and Chunk JSON Data is a Code node that parses that LLM output and emits one item per query. Those items go into Split Data for SerpAPI Batching, which loops one query at a time into Perform SerpAPI Search Request, an HTTP Request node hitting https://serpapi.com/search. The raw SERP JSON is normalised by Format SerpAPI Organic Results, which keeps the organic results array with titles, snippets and URLs.

Those URLs feed Split Data for Jina AI Batching. Each URL is sent to Perform Jina AI Analysis Request, which calls the Jina Reader endpoint (https://r.jina.ai/) to download a clean, LLM-friendly markdown version of the page. The same batch also fans out into Extract Relevant Context via LLM, a LangChain Agent that uses LLM Memory Buffer (Input Context) to keep track of what has already been read and pulls only the passages relevant to the original question.

Once all batches finish, Generate Comprehensive Research Report takes over. It is a second LangChain Agent with two connected sub-nodes: LLM Memory Buffer (Report Context) for long-form continuity and Fetch Wikipedia Information as a fallback knowledge tool. It returns a structured report with sections, citations and a summary, which appears back in the chat panel.

Nodes Included

  • Chat Message Trigger — entry point that opens the n8n chat panel and receives the research question.
  • Generate Search Queries using LLM — Basic LLM Chain that turns the question into a JSON array of search queries.
  • LLM Response Provider (OpenRouter) — shared OpenRouter chat model used by all three LangChain agents/chains.
  • Parse and Chunk JSON Data — Code node that parses the query JSON and emits one item per query.
  • Split Data for SerpAPI Batching — Split In Batches loop that paces SerpAPI calls one query at a time.
  • Perform SerpAPI Search Request — HTTP Request that calls SerpAPI's Google search endpoint.
  • Format SerpAPI Organic Results — Code node that flattens the SERP JSON into a clean list of URLs and snippets.
  • Split Data for Jina AI Batching — Split In Batches loop that fans out URLs to Jina AI and to the context extractor.
  • Perform Jina AI Analysis Request — HTTP Request to r.jina.ai that returns a markdown version of each page.
  • Extract Relevant Context via LLM — LangChain Agent that filters the fetched pages down to passages that actually answer the question.
  • LLM Memory Buffer (Input Context) — window memory the extractor uses to avoid duplicating context across batches.
  • Fetch Wikipedia Information — LangChain Wikipedia tool, no credentials required, attached to the report agent.
  • LLM Memory Buffer (Report Context) — window memory keeping the final agent coherent across long outputs.
  • Generate Comprehensive Research Report — LangChain Agent that writes the final structured report sent back to chat.

Setting Up Credentials

This workflow connects to three external services. Wikipedia needs no key; the rest do.

OpenRouter

  1. Sign in at openrouter.ai and open Keys (openrouter.ai/keys).
  2. Click Create Key, give it a name, and copy the value starting with sk-or-....
  3. Add credit on the Credits page; pay-per-use applies even if a specific model is free.
  4. In n8n, add a new credential of type OpenRouter API and paste the key. Select the model you want (for example openai/gpt-4o-mini or anthropic/claude-3.5-sonnet) on the LLM Response Provider (OpenRouter) node.

SerpAPI

  1. Create an account at serpapi.com.
  2. Open Dashboard → Your Account and copy the Private API Key.
  3. The Perform SerpAPI Search Request node sends the key as the api_key query parameter, so the simplest path is to open that node, find the api_key entry under Query Parameters, and paste your key directly there. If you prefer to keep it out of the workflow JSON, switch the node to Generic Credential Type → Query Auth and store the key as a Query Auth credential with name api_key.

Jina AI Reader

  1. Go to jina.ai/reader and sign in.
  2. Open API Key Manager and copy your token (starts with jina_). The Reader endpoint also works without a key, but authenticated requests get a much higher rate limit and longer page support.
  3. In n8n, on Perform Jina AI Analysis Request, set Authentication to Generic Credential Type → HTTP Header Auth. Create a credential with Name Authorization and Value Bearer YOUR_JINA_KEY.

Customizing the Workflow

Tighten the search prompt in Generate Search Queries using LLM to control breadth: ask for 3 queries for quick answers, 8–10 for thorough reports. The number of queries is the biggest driver of cost and runtime because every query triggers a SerpAPI call plus several Jina fetches.

Swap the report destination by adding a node after Generate Comprehensive Research Report. Common choices: a Notion page, a Google Doc via the Google Docs node, a Slack channel, or an email via Gmail. For long-running runs, replace the Chat Trigger with a Webhook so you can call the workflow from your own UI or a Telegram bot.

Switch knowledge sources by editing the report agent's tools. The current build uses Wikipedia; you can attach additional LangChain tools such as Tavily, Wolfram Alpha, or a custom HTTP tool that hits an internal knowledge base.

Common Issues

  • Empty or malformed JSON from the query generator. If Parse and Chunk JSON Data throws, the model returned prose instead of JSON. Lower the temperature on the OpenRouter node and add "Respond with valid JSON only, no markdown fences." to the system prompt.
  • Jina AI timeouts or 451 responses. Some sites block the Reader. Wrap the HTTP Request in Continue On Fail and add a Filter node downstream so a few dead URLs do not kill the whole run; the report agent can still work from the surviving pages.
17 nodesmanual trigger211 views0 copiesAI

Workflow JSON

{"id": "WLSqXECfQF7rOj2A", "meta": {"instanceId": "cba4a4a2eb5d7683330e2944837278938831ed3c042e20da6f5049c07ad14798"}, "name": "Open Deep Research - AI-Powered Autonomous Research Workflow", "tags": [], "nodes": [{"id": "b7b70ba1-0267-4d2b-91f4-5cc4fd22fd03", "name": "Chat Message Trigger", "type": "@n8n/n8n-nodes-langchain.chatTrigger", "position": [-1940, 160], "webhookId": "cb0b9dbe-1f35-441a-b062-29624b0ebc6a", "parameters": {"options": {}}, "typeVersion": 1.1}, {"id": "55a8a512-f2d4-4aed-93e5-dd9bfa2dcaad", "name": "Generate Search Queries using LLM", "type": "@n8n/n8n-nodes-langchain.chainLlm", "position": [-1760, 160], "parameters": {"text": "=User Query: {{ $('Chat Message Trigger').item.json.chatInput }}", "messages": {"messageValues": [{"message": "=You are an expert research assistant. Given a user's query, generate up to four distinct, precise search queries that would help gather comprehensive information on the topic. Return only a JSON list of strings, for example: ['query1', 'query2', 'query3']."}]}, "promptType": "define"}, "typeVersion": 1.5}, {"id": "5f92361a-b490-479d-8360-c87a100b470e", "name": "LLM Response Provider (OpenRouter)", "type": "@n8n/n8n-nodes-langchain.lmChatOpenRouter", "position": [-1760, 700], "parameters": {"model": "google/gemini-2.0-flash-001", "options": {}}, "credentials": {"openRouterApi": {"id": "", "name": "[Your openRouterApi]"}}, "typeVersion": 1}, {"id": "4ab360eb-858f-48b8-a00d-71867d4f0c93", "name": "Parse and Chunk JSON Data", "type": "n8n-nodes-base.code", "position": [-1420, 160], "parameters": {"jsCode": "// Parse the input JSON string and split it into four chunks\nconst rawText = $json.text;\n\n// Remove Markdown JSON code blocks if present\nconst cleanedText = rawText.replace(/```json|```/g, '').trim();\n\ntry {\n const jsonArray = JSON.parse(cleanedText);\n if (!Array.isArray(jsonArray)) {\n throw new Error('The JSON is not an array.');\n }\n const chunkSize = Math.ceil(jsonArray.length / 4);\n const chunks = [];\n for (let i = 0; i < jsonArray.length; i += chunkSize) {\n chunks.push(jsonArray.slice(i, i + chunkSize));\n }\n return chunks.map(chunk => ({ json: { chunk } }));\n} catch (error) {\n return [{ json: { error: error.message } }];\n}\n"}, "typeVersion": 2}, {"id": "5a3ac393-8355-449f-93cb-b98e8bee9b80", "name": "Perform SerpAPI Search Request", "type": "n8n-nodes-base.httpRequest", "position": [-780, 180], "parameters": {"url": "https://serpapi.com/search", "options": {}, "sendQuery": true, "queryParameters": {"parameters": [{"name": "q", "value": "={{ $('Parse and Chunk JSON Data').item.json.chunk }}"}, {"name": "api_key", "value": "={{ $credentials.SerpAPI.key }}"}, {"name": "engine", "value": "google"}]}}, "typeVersion": 4.2}, {"id": "dad82469-830d-40fb-9f6b-b9fefef41267", "name": "Perform Jina AI Analysis Request", "type": "n8n-nodes-base.httpRequest", "position": [80, 160], "parameters": {"url": "=https://r.jina.ai/{{ $json.url }}", "options": {}, "authentication": "genericCredentialType", "genericAuthType": "httpHeaderAuth"}, "credentials": {"httpHeaderAuth": {"id": "", "name": "[Your httpHeaderAuth]"}}, "typeVersion": 4.2}, {"id": "e21bbdf6-a903-491e-920c-ef7576f9ce80", "name": "Format SerpAPI Organic Results", "type": "n8n-nodes-base.code", "position": [-460, 140], "parameters": {"jsCode": "// Format the organic search results from SerpAPI\nconst results = $input.first().json.organic_results;\nif (results.length === 0) {\n return [{ json: { error: 'No search results found.' } }];\n}\nconst formattedResults = results.map(result => ({\n title: result.title || 'No title available',\n url: result.link || 'No link available',\n source: result.source || result.displayed_link || 'Unknown source'\n}));\nreturn formattedResults.map(result => ({ json: result }));\n"}, "typeVersion": 2}, {"id": "a856c8e8-5c3c-4a2f-9086-66deee1afd06", "name": "Extract Relevant Context via LLM", "type": "@n8n/n8n-nodes-langchain.agent", "position": [-1280, 520], "parameters": {"text": "=User Queries: {{ $('Parse and Chunk JSON Data').all().map(item => item.json.chunk[0]).join(', ') }}\nWebpage Contents: \n\"\"\"\n{{ $json.data }}\n\"\"\"", "options": {"systemMessage": "=You are an expert information extractor. Given the user's query, the search query that led to this page, and the webpage content, extract all relevant pieces of information that are useful to answer the query. Return only the relevant context as plain text without any additional commentary."}, "promptType": "define"}, "typeVersion": 1.7}, {"id": "6d5c6698-0b4f-438c-91b9-3597f5d3e904", "name": "Generate Comprehensive Research Report", "type": "@n8n/n8n-nodes-langchain.agent", "position": [-740, 520], "parameters": {"text": "=Extracted Contexts (Merged):\n\"\"\"\n{{ $json.output }}\n\"\"\"", "options": {"systemMessage": "You are an expert researcher and report writer. Based on the gathered contexts and the original user query, generate a comprehensive, well-structured report. Include all relevant insights and conclusions without unnecessary commentary.\n\nFormat the report in Markdown with clear headings. For example:\n\n# Research Report: [User Query]\n\n## Key Findings\n- Point 1\n- Point 2\n\n## Detailed Analysis\n### Aspect 1\nSummary of findings.\n_Source:_ [Source Name](URL)\n\n### Aspect 2\nSummary of findings.\n_Source:_ [Another Source](URL)\n\nNow, generate the complete report."}, "promptType": "define"}, "typeVersion": 1.7}, {"id": "05fea6a1-791e-4980-8f2a-2960455066d7", "name": "Split Data for SerpAPI Batching", "type": "n8n-nodes-base.splitInBatches", "position": [-1100, 160], "parameters": {"options": {}}, "typeVersion": 3}, {"id": "df00e7e8-99b8-484a-8047-869474fefee9", "name": "Split Data for Jina AI Batching", "type": "n8n-nodes-base.splitInBatches", "position": [-220, 140], "parameters": {"options": {}}, "typeVersion": 3}, {"id": "2edc683b-65f7-40c3-a22d-7fbf5b67de0a", "name": "LLM Memory Buffer (Input Context)", "type": "@n8n/n8n-nodes-langchain.memoryBufferWindow", "position": [-1160, 740], "parameters": {"sessionKey": "my_test_session", "sessionIdType": "customKey", "contextWindowLength": 20}, "typeVersion": 1.3}, {"id": "23017ae7-72a7-45c7-8edf-d0ba72220675", "name": "LLM Memory Buffer (Report Context)", "type": "@n8n/n8n-nodes-langchain.memoryBufferWindow", "position": [-620, 760], "parameters": {"sessionKey": "my_test_session", "sessionIdType": "customKey", "contextWindowLength": 20}, "typeVersion": 1.3}, {"id": "6bc9533b-e265-47b3-b93a-3a4f86ba0541", "name": "Fetch Wikipedia Information", "type": "@n8n/n8n-nodes-langchain.toolWikipedia", "position": [-580, 920], "parameters": {}, "typeVersion": 1}, {"id": "b25c148e-047d-40a7-8818-94c3504828dd", "name": "Sticky Note: SerpAPI Setup", "type": "n8n-nodes-base.stickyNote", "position": [-940, -20], "parameters": {"color": 7, "width": 420, "height": 140, "content": "## SerpAPI Setup Instructions\n1. Obtain your API key from https://serpapi.com/manage-api-key.\n2. Save your API key securely in n8n credentials (do not use plain text)."}, "typeVersion": 1}, {"id": "e69c9a85-31e4-42b9-a09a-683ec5bb97d1", "name": "Sticky Note: Jina AI Setup", "type": "n8n-nodes-base.stickyNote", "position": [-60, -40], "parameters": {"color": 7, "width": 420, "height": 140, "content": "## Jina AI Setup Instructions\n1. Obtain your API key from https://jina.ai/api-dashboard/key-manager.\n2. Configure your Jina AI credential in n8n to ensure secure API access."}, "typeVersion": 1}, {"id": "dbd204e0-da8e-41d8-814b-f409a23e9573", "name": "Sticky Note: OpenRouter API Setup", "type": "n8n-nodes-base.stickyNote", "position": [-1680, 460], "parameters": {"color": 7, "width": 300, "height": 180, "content": "## OpenRouter API Setup Instructions\n1. Obtain your API key from https://openrouter.ai/settings/keys.\n2. Set up your OpenRouter credential in n8n for secure integration."}, "typeVersion": 1}], "active": false, "pinData": {}, "settings": {"executionOrder": "v1"}, "versionId": "aa857bb3-84c1-4fe6-9464-90fc09163960", "connections": {"Chat Message Trigger": {"main": [[{"node": "Generate Search Queries using LLM", "type": "main", "index": 0}]]}, "Parse and Chunk JSON Data": {"main": [[{"node": "Split Data for SerpAPI Batching", "type": "main", "index": 0}]]}, "Fetch Wikipedia Information": {"ai_tool": [[{"node": "Generate Comprehensive Research Report", "type": "ai_tool", "index": 0}]]}, "Format SerpAPI Organic Results": {"main": [[{"node": "Split Data for Jina AI Batching", "type": "main", "index": 0}]]}, "Perform SerpAPI Search Request": {"main": [[{"node": "Split Data for SerpAPI Batching", "type": "main", "index": 0}]]}, "Split Data for Jina AI Batching": {"main": [[{"node": "Extract Relevant Context via LLM", "type": "main", "index": 0}], [{"node": "Perform Jina AI Analysis Request", "type": "main", "index": 0}]]}, "Split Data for SerpAPI Batching": {"main": [[{"node": "Format SerpAPI Organic Results", "type": "main", "index": 0}], [{"node": "Perform SerpAPI Search Request", "type": "main", "index": 0}]]}, "Extract Relevant Context via LLM": {"main": [[{"node": "Generate Comprehensive Research Report", "type": "main", "index": 0}]]}, "Perform Jina AI Analysis Request": {"main": [[{"node": "Split Data for Jina AI Batching", "type": "main", "index": 0}]]}, "Generate Search Queries using LLM": {"main": [[{"node": "Parse and Chunk JSON Data", "type": "main", "index": 0}]]}, "LLM Memory Buffer (Input Context)": {"ai_memory": [[{"node": "Extract Relevant Context via LLM", "type": "ai_memory", "index": 0}]]}, "LLM Memory Buffer (Report Context)": {"ai_memory": [[{"node": "Generate Comprehensive Research Report", "type": "ai_memory", "index": 0}]]}, "LLM Response Provider (OpenRouter)": {"ai_languageModel": [[{"node": "Generate Search Queries using LLM", "type": "ai_languageModel", "index": 0}, {"node": "Extract Relevant Context via LLM", "type": "ai_languageModel", "index": 0}, {"node": "Generate Comprehensive Research Report", "type": "ai_languageModel", "index": 0}]]}}}

How to Import This Workflow

  1. 1Copy the workflow JSON above using the Copy Workflow JSON button.
  2. 2Open your n8n instance and go to Workflows.
  3. 3Click Import from JSON and paste the copied workflow.

Don't have an n8n instance? Start your free trial at n8nautomation.cloud

Related Templates

Text to Speech (OpenAI)

Converts text into natural-sounding speech using OpenAI's Text-to-Speech API. It sends your input text to OpenAI and receives an audio file in return. This is useful for creating audio versions of articles, generating voiceovers for videos, or providing accessibility features for web content. Quickly transform written content into engaging audio.

8 nodes

Automate Customer Support Issue Resolution using AI Text Classifier

Automate the resolution of customer support issues by classifying their state and applying AI-driven actions. This workflow connects Jira for issue management, OpenAI for AI classification and response generation, and Slack for notifications. Support teams can use this to automatically close resolved tickets, remind customers about open issues, or escalate complex cases.

36 nodes

AI-Powered Candidate Shortlisting Automation for ERPNext

Automate AI-powered candidate shortlisting for ERPNext job applications. This workflow connects ERPNext, Google Gemini, WhatsApp, and Outlook to process resumes, evaluate candidates, and communicate outcomes. Recruiters and HR departments can use this to efficiently screen applicants, automatically reject unqualified candidates, and send acceptance notifications. It significantly reduces manual review time and streamlines the hiring process.

39 nodes

Ready to automate with n8n?

Get affordable managed n8n hosting with 24/7 support.