Where OpenAI wins for us, and where it doesn't
Technology
OpenAI is not our default chat model. It shows up in two places in our work: the embedding layer under retrieval, and generation inside an agent pipeline that was already built around GPT-4o. Both are jobs where the model is a component, not the product.
The assistant on this site embeds our CMS content with an OpenAI model and stores the vectors in Pinecone. For Quicknode we put GPT-4o inside a LangGraph pipeline that writes structured web3 tool profiles into Contentful, taking a directory that started at 700 curated entries toward a potential 5,000. This page is about those jobs, not about OpenAI in general.
Where
The jobs we actually give OpenAI
- Embeddings, where nobody notices the model
The assistant on this site embeds our CMS content with an OpenAI model and keeps the vectors in Pinecone. Nobody reading an answer asks which model produced the embedding behind it. It is the least visible OpenAI dependency we run, and the one that has never been the bottleneck.
- Picking an embedding model is a commitment
Worth knowing before you start. A vector index is built for one model at one dimension count. Change either and every new query vector mismatches the stored ones, retrieval quietly returns worse results rather than failing loudly, and the only real fix is a full re-index.
- Two builds, and we will name them
The OpenAI work we can point at is two systems: the embeddings behind this site's assistant, and GPT-4o generating tool profiles inside Quicknode's Builder's Guide. That is the extent of it. A vendor page that names no builds is describing the vendor, not the work.
OpenAI work
Where we've shipped it
Writing
How we think about this
The model you argue about in a kickoff is the chat model. The one you are actually stuck with is the embedding model, because it is baked into an index rather than passed as a parameter.
So we treat the two decisions differently. Chat models get compared, swapped, and revisited as new versions land. Embedding models get picked once, deliberately, with the cost of a full re-index priced in before anything ships.
That is the OpenAI decision worth spending time on. Two weeks to a build plan. Start a conversation.