Google Doc sections given context, then stored in Pinecone

Google Doc sections given context, then stored in Pinecone

An n8n workflow that prepares one Google Doc for search. It splits the text into sections, has a model write a short line of context for each section, and stores every section with its context in Pinecone as a Gemini vector.

What is inside

• workflow.json: the workflow as n8n exports it, 16 nodes. Thirteen run, and three are sticky notes. The thirteen are a manual trigger, the Drive download, the text extraction, a Code node, a Split Out and a Loop Over Items. The rest are the agent with its OpenRouter chat model, a Set node, and the Pinecone store with its loader, text splitter and embeddings. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The OpenRouter Chat Model node had no model set, so it ran n8n's default, openai/gpt-4.1-mini. OpenAI's deprecation page, read on 12 November 2026, shuts that model down on 14 January 2027 and names gpt-6-luna. The node now sets openai/gpt-6-luna, which OpenRouter's public model list shows. The agent has no tool, so that model fits.
• Embeddings Google Gemini used models/text-embedding-004. Google's page, read on 12 November 2026, shut it down on 14 January 2026 and names gemini-embedding-2. The node now uses models/gemini-embedding-001, a text-only model that is also this node's default. It returns 3072 values by default.
• Split Document Text Into Sections split on one exact line of dashes around [SECTIONEND], and an empty section would have gone to the agent. The code now cuts at [SECTIONEND] with or without dashes around it, trims each section and drops empty ones.
• The credential links are gone. The Google Doc in Get Document From Google Drive is empty, so you pick your own, because the original pointed at a named person's document.
• The sticky notes are written again, and the one that linked to a video is left out.
• The file's name is ours.
• Nothing else. The agent's prompt, the Set node, the splitter and the loop are as the file had them.

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

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google account and a Google Doc.
• An OpenRouter API key.
• 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 context-rag-test 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, an OpenRouter 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 Get Document From Google Drive, pick your Google Doc.
5. In Pinecone Vector Store, pick your index.
6. In your Doc, put a line that reads [SECTIONEND] between the sections.
7. Press Test workflow.

What happens in a run

1. Get Document From Google Drive downloads the Doc and converts it to plain text.
2. Extract Text Data From Google Document puts that text into a field called data.
3. Split Document Text Into Sections cuts data at each [SECTIONEND] marker. It keeps the sections in section and the same sections as one JSON string in document.
4. Prepare Sections For Looping turns the sections into one item each, and Loop Over Items hands them on one at a time.
5. AI Agent - Prepare Context sends the model the whole document and the one section, with the request for a short line of context that helps a search find the section, and nothing else in the answer.
6. Concatenate the context and section text makes a field section_chunk from the model's line, a full stop and the section.
7. Pinecone Vector Store stores section_chunk in insert mode. Embeddings Google Gemini makes the vector, Default Data Loader reads the field and Recursive Character Text Splitter has a chunk size of 100000, so a section stays whole. Then the loop moves to the next section.

What to change

• The marker between sections: the pattern in Split Document Text Into Sections.
• What the model is asked: the text field in AI Agent - Prepare Context.
• The model that writes the context: OpenRouter Chat Model.
• The embedding model: Embeddings Google Gemini. If you change it, make a new index of its size.
• The index: Pinecone Vector Store.

Good to know

• Keep each section and its context within gemini-embedding-001's limit of 2,048 tokens.
• The workflow handles one document per run. Run it twice and everything is stored twice, because the Pinecone node's Clear Namespace option is off.
• Each model call carries the whole document, so a long document costs more with every section.
• The loop's finished output has nothing connected, so the run ends when the last section is stored.
• Data: the document text goes to OpenRouter, the provider of the model you pick and Google, and the vectors go to Pinecone.
• Cost: one model call and one embedding call per section.