Private Preview - features may change without notice
GramSpec
Architecture Spec API Reference Pricing About
Start free

Build on it · 19

Your own app on the model

Their team puts “ask your data” inside their own product without building the hard part. We hand over the instructions and the model, and when the model is edited the app can answer more the next time it asks, with no release of theirs.

The GRAM API. In one flow the customer's app fetches the prompt and the model, keeps them until the model changes, asks its own language model, and runs the SQL on its own database. In the other the app sends a question and GramSpec has a language model write the SQL for the app to run.GRAMSPECYOUR INFRASTRUCTUREThe modelGRAM APIthe prompt and themodel, or SQL for aquestionA language modelwe callYour userYour appkeeps the prompt until themodel changesYour language modelany provider, your accountYour databaseyour credentials12312 The GRAM API. In one flow the customer's app fetches the prompt and the model, keeps them until the model changes, asks its own language model, and runs the SQL on its own database. In the other the app sends a question and GramSpec has a language model write the SQL for the app to run.GRAMSPECYOUR INFRASTRUCTUREThe modelGRAM APIthe prompt and themodel, or SQL for aquestionA language modelwe callYour userYour appkeeps the prompt untilthe model changesYour languagemodelYour databaseyour credentials12312
Blue circles: their own language model writes the SQL. Green squares: we write it. Either way their app runs it.

Their own language model

  1. Their app fetches the prompt and the model. One call returns the whole instruction for a language model: the rules, the notes for their database's dialect, and the model itself. It comes with a fingerprint. The app keeps it and asks again with the fingerprint, and is told “nothing changed” until the model is edited.
  2. Their app asks their own language model. Any provider, on their own account. The question and the answer never pass through us.
  3. Their app runs the SQL on their own database. With their own credentials. We see neither the rows nor the login.

We write the SQL

  1. Their app sends a question. In plain words, naming one connected model.
  2. We have a language model write the answer. What goes back is the SQL, a sentence of explanation and a chart to draw. We do not run it: their app does, as in step 3 above.

A whole investigator, on their side. A third call hands over everything needed to run a Reason-style investigation in their own app: the instructions, the model, the definitions of the tools, and a ready page about each table. Their app runs it, with their language model.

The demo that closes. Edit the model in Grammar. Refresh the app built on the API. Ask a harder question than it could answer a minute ago. Nothing was released: the fingerprint changed, so the app fetched the new model. The same API also serves the model as a file, and the list of its versions.