How Jev Powers Chem Autocomplete

Chem Autocomplete suggests the next structure while you draw. RDKit builds and checks every candidate from your drawing, and Jev, a decision model from TypeSafe, picks which of those checked cards to show and in what order, in one request of about 0.2 seconds.

What Autocomplete does

Chem Autocomplete watches the structure on the Generator canvas and offers what you are likely to draw next. Draw an amino ester and it offers to protect the amine. Draw a molecule with a stereocenter and it shows the stereoisomers. Grow a chain on a known compound and it tells you which PubChem compound your drawing is turning into, with a link to the record.

Each suggestion is a card with a thumbnail. Add puts the structure on the canvas next to your drawing, and Replace swaps it in for the molecule it was built from. Replace now reads the live canvas at the moment you click, so it replaces the right molecule on the canvas as it is now.

Autocomplete suggests steps and structures. It does not predict reactions.

Chem Autocomplete drawing L-serine methyl ester, then protecting the amine with Fmoc and the alcohol with TBDMS.

RDKit writes the chemistry

No language model writes a structure in Autocomplete. Every candidate is built by RDKit from the structure you drew, using a small set of generators:

  • Protect and deprotect: add a protecting group to an amine, alcohol, phenol, carboxylic acid, carbonyl or thiol, or take one off.
  • Stereo: enumerate the stereoisomers of the drawn structure.
  • Analogs: close structural variations of what you drew.
  • Properties: calculated molecular weight, logP, polar surface area, and hydrogen-bond donor and acceptor counts.
  • PubChem literature: the known compound your drawing matches, or the one it is one step away from.

Because RDKit produces the structures, every card is a valid molecule derived from your drawing. Nothing is generated as free text.

Jev makes the decisions

The generators usually produce more candidates than fit on screen, and most of them are not what you want right now. Choosing among them is the job of Jev, a decision model from TypeSafe (typesafe/jev-1.13), which we call through OpenRouter's Decisions API.

Jev does not write text or chemistry. It answers typed questions: pick one option from a list, score something on a fixed scale, or answer yes or no, and it returns probabilities for each answer. We send it a compact description of the canvas and ask what you are most likely doing (protecting a group, exploring analogs, heading toward a known compound) and which of the checked cards are worth showing, in what order. A choice question always includes a “none” option, so Jev can decline to show anything instead of forcing a weak card.

All of that happens in one request of about 0.2 seconds. A full Autocomplete cycle, from your edit to new cards, measures about 170–300 ms.

PubChem lookups are slower, so they never hold up a cycle. They run in the background, and when a match comes back the card appears on the next check a few seconds later.

Safety and privacy

Before any request goes to Jev, a local RDKit safety check screens the drawn structure, and the generated candidates are screened again before ranking. The check runs on our server with no network call. If it blocks a structure, no Jev request is made.

PubChem receives only the structure. In Private Structure Mode, the PubChem lookup is skipped entirely unless you have allowed external lookups.

What it costs you

Nothing. Autocomplete is free for every registered account and does not use AI Actions. Without an account you get 8 Autocomplete cycles a day; signing in with a free account lifts that limit and saves your canvas.

Try it

The Chem Autocomplete page shows two worked examples captured from the product. To try it on your own drawing, open the Generator with Autocomplete on. It runs on a laptop or desktop.