Chat with the GitHub API spec stored in Pinecone with OpenAI

What it is

This n8n workflow is a chatbot that answers questions about the GitHub REST API from its official specification. One part downloads the specification, cuts it into pieces and saves each piece in a Pinecone index as an OpenAI embedding. The other part is a chat window, where an AI Agent finds the best matching pieces and answers from them. Fourteen nodes do the work, beside three sticky notes.

Who it suits

Developers and technical writers who want a working example of retrieval augmented generation in n8n, and anyone who wants to ask about GitHub’s API without reading its reference. Skip it if you need answers from your own documents as they stand. You would change what it downloads first.

You need a little n8n: import a file, add an OpenAI and a Pinecone credential, create an index and run the indexing once.

Where to use it

Any n8n that can reach OpenAI, Pinecone and GitHub’s file host will run it. The chat opens inside n8n, so the answers stay in your editor.

• n8n Cloud: import the file into a new workflow there
• Your own n8n server: the same file and steps
• Your own documents: change the address in HTTP Request to another JSON or text file
• A second index: point both Pinecone nodes at it to keep topics apart

Setup

Download saves a zip with workflow.json and Mindket’s setup notes in your downloads folder. Two accounts are involved.

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json.
2. In Pinecone, create an index with 1536 dimensions, the size of OpenAI’s text-embedding-3-small embeddings.
3. Add an OpenAI credential and select it in Generate Embeddings, Generate User Query Embedding, OpenAI Chat Model and OpenAI Chat Model1.
4. Add a Pinecone credential, select it in Pinecone Vector Store and Pinecone Vector Store (Querying), and choose your index in both.
5. Click Test workflow on the manual trigger and wait for it to finish. The first run takes a while, since the file is about 14 MB.
6. Open the chat and ask a question.

How to use it

Once the index is full, open the chat panel and ask about the GitHub API, for example how to list a repository’s issues. The agent keeps the last few messages of a chat as memory. Run the indexing again only when you change the file or the index.

Example

A developer asks the chat how to list the issues of a repository. The agent calls the vector store tool, which turns the question into an embedding and asks Pinecone for the four closest pieces.

The tool’s own model reads those pieces and answers, and the agent passes the answer on. The reply names the path for listing issues and says that open issues are returned by default.

For newcomers and experts

If you are new. n8n draws a workflow as boxes joined by lines, and each box, a node, does one job. This one has two starts. The manual start fills the index, and the chat start answers questions from it. Run the first once, then open each box of the second to see where an answer comes from.

For experts. Both Pinecone nodes need the same index and the same embedding model, or the vectors will not match. Mindket’s copy sets text-embedding-3-small on both embedding nodes outright, where the original left it to the node’s default.

Good to know

The indexing run is large for a first try, and a model change was needed.

• Models: OpenAI’s deprecations page, read on 12 November 2026, lists gpt-5-mini, the node’s old default, as shutting down on 11 December 2026 with gpt-5.6-terra as the replacement. The original named no model, so Mindket’s copy sets gpt-5.6-terra for the agent and gpt-6-luna for the tool’s own model, which only reads the pieces
• Why not `gpt-6-luna` for the agent: its model page says Chat Completions supports function calling only with reasoning effort off, and an agent needs function calling
• Size: the file splits into roughly 14,000 pieces, each embedded and stored, so the first run takes a while and costs a little on both services
• Run it once: a second indexing run stores every piece again, unless Clear Namespace is on in Pinecone Vector Store
• Index size: an index that does not have 1536 dimensions refuses the vectors, so check it before the first run
• Memory: the agent’s window memory lives in n8n’s own memory, is cleared after an idle hour and does not suit queue mode
• Empty pickers: the author’s own Pinecone index is gone, so choose yours
• Billing: each question costs one agent call and one call inside the tool, plus an embedding, and each is billed by its provider
• Your data: your questions go to OpenAI and Pinecone, and the file you index goes to both too
• Credentials: the JSON holds none, so add the OpenAI and Pinecone credentials yourself