Quarterly earnings PDFs turned into an analyst report

Quarterly earnings PDFs turned into an analyst report

An n8n workflow in two halves. The first loads quarterly earnings PDFs from Google Drive into a Pinecone index, with Gemini
embeddings. The second asks an OpenAI agent for a report on the last three quarters. The agent searches that index, and the
finished report is added to a Google Doc.

What is inside

• workflow.json: the workflow as n8n exports it, nineteen nodes. Sixteen of them run and three are sticky notes. It holds no
  credential.
• README.md: these notes.

What we changed from the original copy

• Nothing started the half that fills the index, so a test run only reached the agent. We added a Manual Trigger,
  Load earnings PDFs, in front of List Of Files To Load (Google Sheets).
• Both embedding nodes called models/text-embedding-004, which Google shut down on 14 January 2026. They now call
  models/gemini-embedding-2, the replacement on Google's deprecations page, read on 12 November 2026.
• The Gemini chat model inside the vector store tool was models/gemini-2.0-flash-exp. The same page lists the 2.0 Flash models as
  shut down on 1 June 2026 and names gemini-3.6-flash as their replacement, so the node calls models/gemini-3.6-flash.
• The sheet, the Google Doc and every credential link belonged to the file's author. The two IDs are placeholders now, and the
  credential links are gone.
• The sticky note of setup steps is replaced by a short pointer to this file.
• Nothing else. The file has no Code node.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google account for three OAuth2 credentials: Google Sheets, Google Drive and Google Docs.
• A Gemini API key from Google AI Studio, for a Google Gemini(PaLM) Api credential.
• A Pinecone account and API key.
• An OpenAI API key.
• The earnings PDFs, saved in Google Drive.
• A Google Sheet with one row for each PDF, and an empty Google Doc for the reports.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Pinecone, create an index named company-earnings with 3072 dimensions. Google's embeddings page
   gives 3072 as the default size of gemini-embedding-2. An index made for the older model's 768 will refuse the vectors.
3. Fill the sheet's first row with two headings, File URL and 10Q. Under them, add one row for each PDF: its Google Drive link,
   and the name to save the file under.
4. Add the credentials and select each one: Google Sheets in List Of Files To Load (Google Sheets), Google Drive in
   Download File From Google Drive, Google Docs in Save Report to Google Docs, Google Gemini(PaLM) in
   Embeddings Google Gemini, Embeddings Google Gemini (retrieval) and Google Gemini Chat Model1, Pinecone in both Pinecone
   nodes, and OpenAI in OpenAI Chat Model.
5. In List Of Files To Load (Google Sheets), replace YOUR_SHEET_ID with the ID from the sheet's address, the part between /d/
   and /edit, and YOUR_SHEET_TAB with the tab's name.
6. In Save Report to Google Docs, replace YOUR_GOOGLE_DOC_ID with the ID of your document, found the same way.
7. Press the play button on Load earnings PDFs and wait until the loop ends. Check in Pinecone that the index now holds vectors.
8. Open AI Agent and rewrite its prompt and system message for your company. Both are written for Google.
9. Run the other trigger, When clicking ‘Test workflow’, and open the Google Doc.

How a run goes

Filling the index:

1. Load earnings PDFs starts List Of Files To Load (Google Sheets), which reads every row of the tab.
2. Loop Over Items passes on one row at a time, since its batch size is left at 1.
3. Download File From Google Drive opens the file at File URL and names it after the 10Q column.
4. Default Data Loader reads the file as a PDF, one document for each page. Recursive Character Text Splitter cuts those into
   chunks of 1000 characters with no overlap, and Embeddings Google Gemini turns each chunk into a vector.
5. Pinecone Vector Store, set to insert, writes the vectors into the company-earnings index.
6. The node hands control back to Loop Over Items for the next row.

Asking for the report:

1. When clicking ‘Test workflow’ starts AI Agent (version 1.7). Its prompt is typed in the node: a markdown report on Google's
   last three quarters, with the differences, the trends and any outliers.
2. The agent's model is OpenAI Chat Model. Its one tool, Vector Store Tool, is named company_financial_earnings_data_tool.
3. When the agent calls the tool, Pinecone Vector Store (Retrieval) finds the four closest chunks, since Limit is left at 4.
   Google Gemini Chat Model1 answers from them and hands that answer back to the agent. The agent may call the tool many times.
4. Save Report to Google Docs adds the agent's output to the end of your document.

Change the look

• Company and quarters: the sheet's rows, and the prompt and system message in AI Agent.
• The report's sections: the list in the system message.
• Chunk size: Chunk Size in Recursive Character Text Splitter. Load the PDFs again after a change.
• Passages for each search: Limit in Vector Store Tool.
• Models: OpenAI Chat Model and Google Gemini Chat Model1. Keep both embedding nodes on the same model.
• Where the report goes: replace the last node with Gmail, Slack or another node that takes text.

Good to know

• Each press on Load earnings PDFs adds every chunk again, and nothing checks for repeats. Empty the index before you reload.
• The index size must match the embedding model. If you change the model in both embedding nodes, make a new index to suit it.
• The system message says the agent has a Google Docs tool. It has none: Save Report to Google Docs writes the text after the agent
  finishes. The sentence does no harm, and you can delete it.
• The report passes through two models. Gemini writes each tool answer from four chunks, and the agent builds the report from those
  answers. Check the figures that matter against the PDFs before anyone relies on them.
• OpenAI Chat Model sets no model, so n8n's default for that node version applies, gpt-5-mini. OpenAI's list of retiring models, read the
  same day, gives 11 December 2026 as its shutdown and gpt-5.6-terra as the replacement. Pick a model in the node.
• Google's page also lists gemini-2.5-pro and gemini-2.5-flash for 1 December 2026. Neither is used here.
• Billing: Gemini charges for each chunk embedded and again for each search, OpenAI for the agent's turns, and Pinecone for the index.
• Where the data goes: the PDFs' text goes to Google for embedding, the chunks are stored in Pinecone, and the question and the tool's
  answers go to OpenAI. The finished report lands in your Google Doc.