Chat with your company documents from Google Drive

Chat with your company documents from Google Drive

An n8n workflow that answers questions from your company documents. One flow watches a Google Drive folder and stores each new or changed file in a Pinecone index, using Gemini to make the vectors. A second flow takes a question from n8n's chat, and an agent answers from the closest pieces.

What is inside

• workflow.json: the workflow as n8n exports it, 18 nodes. Fifteen run, and three are sticky notes. The fifteen are two Drive triggers and a download step, the Pinecone store with its loader, text splitter and embeddings, and the chat trigger. The rest are the agent with its memory and chat model, and the document tool with its own Pinecone search, embeddings and chat model. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The two Embeddings Google Gemini nodes used models/text-embedding-004. Google's deprecation page, read on 12 November 2026, shut that model down on 14 January 2026 and names gemini-embedding-2. That model returns one combined vector when a request holds several inputs, and Mindket did not want to rely on that without a run. Both nodes now use models/gemini-embedding-001, a text-only model that is also this node's own default. It returns 3072 values by default.
• The two Google Gemini Chat Model nodes used models/gemini-2.0-flash-exp. Google's page does not list that name any more, and the gemini-2.0-flash family shut down on 1 June 2026. Both nodes now use models/gemini-3.8-flash, which Google's models page lists as stable.
• The credential links are gone. Both Drive triggers have an empty folder to pick, because the original pointed at the author's own folder.
• The three sticky notes are written again, and the set-up note now gives the index size.
• The file's name is ours.
• Nothing else. The agent's rules, the document tool, the splitter and the loader are as the file had them.

Mindket read this copy against n8n's source and Google's pages. It was not run on a live n8n.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google account with Drive, and a folder for your documents.
• A Google AI Studio API key for the Gemini credential.
• A Pinecone account and an API key.

Set it up

1. In Pinecone, create an index named company-files with 3072 dimensions and the cosine metric, and copy an API key.
2. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
3. Add a Google Drive credential, a Google Gemini (PaLM) credential with your AI Studio key and a Pinecone credential. Select them in the nodes that ask for them.
4. In Google Drive File Created and Google Drive File Updated, pick the folder that holds your documents.
5. In both Pinecone nodes, pick the company-files index.
6. Publish the workflow, add a document to the folder and give it a minute.
7. Open the chat in n8n and ask a question.

What happens in a run

Filling the index:

1. Google Drive File Created and Google Drive File Updated check the folder every minute. Both pass the file on to Download File From Google Drive, which downloads it under its own name.
2. Pinecone Vector Store stores the file in insert mode. Default Data Loader reads the downloaded file, Recursive Character Text Splitter cuts it into pieces of about 1,000 characters that overlap by 100, and Embeddings Google Gemini turns each piece into a vector.

Answering a question:

1. When chat message received passes your question to AI Agent, which has Window Buffer Memory and Google Gemini Chat Model.
2. The agent's system message makes it a helpful HR assistant for questions about company policies. It must use the company_documents_tool for anything that comes from the documents. When the documents hold no answer, it must say so in a fixed line.
3. Vector Store Tool searches through Pinecone Vector Store (Retrieval) and Embeddings Google Gemini (retrieval), returns the closest pieces and has Google Gemini Chat Model (retrieval) write the answer from them.

What to change

• The assistant's role and rules: the system message in AI Agent.
• How the document tool is described to the agent: Vector Store Tool.
• The size of a piece: Recursive Character Text Splitter.
• The folder: both Drive triggers.
• The models: the two Gemini chat nodes and the two embeddings nodes. If you change an embeddings model, make a new index of its size and fill it again.

Good to know

• An index made for another vector size than the one in step 1 will not take the vectors.
• A changed file is added again. The Pinecone node's Clear Namespace option is off, so the old pieces of that file stay beside the new ones.
• The loader reads PDF, Word, text, CSV, JSON and EPub. A Google Doc is downloaded as a Word file. Google Sheets and Slides arrive as Excel and PowerPoint files, which the loader cannot read.
• The triggers react to files created or changed after you publish the workflow, so add your documents then.
• The chat memory holds the last five interactions inside the n8n process, and a restart clears it.
• Both embeddings nodes must use the same model, or the search finds nothing useful. They do in this copy.
• Data: the text of your files and your questions go to Google and Pinecone. Anyone who can open the chat can read answers drawn from every file in the folder.
• Cost: Google bills the Gemini calls and Pinecone bills the index, each under your plan.