The Builder

Run Python & R code inside your search sessions, sandboxed and persistent

“The primitive here is a REPL with persistent session state embedded in a retrieval interface — that's actually a non-trivial thing to ship correctly, and sandboxed container isolation per session is the right call, not a toy iframe. The DX bet is that you never leave the search context to crunch numbers, which works until you need pip installs beyond the pre-loaded environment or you want to pull in your own data files without pasting CSVs into a chat box. The moment of truth is asking it to analyze a dataset you found in the same session — if that works end-to-end without copy-paste, that's genuinely useful. It's not replacing a Jupyter notebook for serious work, but it doesn't need to: it earns its keep for quick validation tasks where spinning up a local environment is the thing that was stopping you.”

RAG model with citation-level grounding for regulated enterprise search

“The primitive is clear: a RAG model that returns answers with document-level citations baked into the response structure, not bolted on post-hoc. The DX bet is on the connectors — pre-built integrations to Salesforce, SharePoint, and Confluence mean the 'connect your data' step doesn't require you to write a chunking pipeline at 2am. The moment of truth is whether those connectors handle real enterprise data shapes (nested Confluence spaces, Salesforce custom objects) without breaking — the docs suggest yes but I haven't stress-tested edge schemas. What earns the ship is that citation grounding is a first-class output type, not a hallucinated footer: the API returns source references as structured fields, which means downstream auditing is an engineering problem you can actually solve.”

Extended thinking for grad-level math, science, and coding

“The primitive here is straightforward: a reasoning model that allocates more inference compute to hard problems before returning a result. The DX bet OpenAI made is to hide all of that behind the same ChatGPT interface you already use — no new API surface to learn, no config, just select o3 Pro from the model picker. The moment of truth is dropping a genuinely hard coding problem or a graduate-level proof and watching whether the extended thinking trace actually catches errors that o3 misses — in my experience, it does on non-trivial linear algebra and dynamic programming. The honest caveat: if you're accessing this via API you're paying per-token and the latency is real; this is not a drop-in for production pipelines. Ship for the specific use case of hard reasoning problems where correctness matters more than speed.”