The assistant that never answers from memory

The chat assistant in I Like Movies is a language model with tools, the same kind of thing you have used elsewhere. The difference is a rule I built the whole feature around: it is not allowed to answer a question about films from its own memory. Every title it puts in front of you has been looked up, checked against what you have already watched, and filtered down to what you can actually stream tonight, before you ever see it.

That rule exists because of a specific failure. Ask a general chatbot for a film and it will happily give you one, in a confident voice, with a plausible year and a plausible plot. Sometimes the film is real and you cannot watch it anywhere you subscribe. Sometimes the film does not exist at all. Both answers look identical coming out of a model, because a model that is recalling is indistinguishable from a model that is inventing. The only fix is to stop it recalling and make it check. This is how the app does that.

The rule it cannot talk its way around

The assistant's instructions are blunt about this. When you ask it to list, count, or rank a person's films, it is told to fetch that filmography from a tool first and work only from what comes back, and never to claim a complete count like "all twelve of her films" without having done so. The same holds for your lists: it is told never to say a list does not exist without calling the tool that reads your lists, and never to confirm that it added or removed something until the tool that does the adding has actually returned.

This matters more than it sounds. A model asked for "every film Christopher Nolan directed" will produce a list that feels complete and is quietly wrong, usually by missing whatever is newest, because the newest titles came out after its training data was frozen. The tool that answers this reads the person's credits live and includes unreleased and upcoming titles on purpose. So the count is the real count, and a film announced last month is in the list rather than being denied out of existence.

Every title becomes a real record

When the assistant mentions a title, that title gets resolved to a specific entry in the film database before anything happens with it. The lookup tries the cheap path first: your own library and watchlist, already loaded for this conversation, since a title you own is one it already knows the identity of. If it is not there, it searches the film database directly.

Two things about that search are deliberate. First, it never guesses whether you meant the film or the television series of the same name. The year you give, or the model supplies, is what separates them, because the same name at the same numeric id can be a film in one place and a show in another, and guessing hands you the wrong thing. Second, if the search comes back with nothing usable, the assistant does not fill the gap with something invented. It tells you it could not find the title and stops. "I couldn't find that, nothing was changed" is a real answer the code can return; a made-up substitute is not.

It already knows what you have seen

Before the assistant reads a word of your message, its context has been seeded with your own history: the films and shows you have rated, the ones you own, and every title you have marked as watched. Each of those carries a seen marker, and the rule attached to it is that a seen title is a signal about your taste and never a thing to recommend back to you.

The awkward case is someone with years of history, more ratings than fit sensibly into one prompt. Rather than let the older ones fall out of view, where the assistant would no longer know you had seen them, there is a plain name-only roster that carries the overflow, so the never-recommend set stays complete no matter how much you have rated. And the parts of the system that pull candidate suggestions are gated against the full set of what you have watched, independently of anything the prompt trims. The result you feel is simple: it does not suggest a film you have already seen. The work behind that being reliable is most of the effort.

Only what you can actually put on tonight

A recommendation you cannot watch is not much better than one that does not exist. So the last check is streaming. Every title on your watchlist knows which services carry it in your country. The app takes that set and intersects it with the services you have told it you actually subscribe to. Only the titles that survive both are eligible when you ask what to watch tonight.

In pseudocode the check is nothing clever, which is the point:

for each title on your watchlist, keep it only if at least one service that carries it is also a service you have. Otherwise drop it.

That is why a "what should I watch tonight" answer cites its source on every line, either "in your library" for something you own or something in the shape of "on Netflix, in your watchlist" for something streamable. If a watchlist title is not on any service you have, it is not offered for tonight, because you cannot watch it tonight. And if you have not told the app which services you have, that whole pool is empty by design, and it falls back to recommending from the titles you own rather than guessing at a subscription you might not hold.

The honest version of an assistant feature

A few smaller pieces hold the same line. The prompt carries today's date, so "this year" means the actual current year and not the year the model was trained in. It is told plainly that titles released after its training cutoff still exist, so it never tells you a recent film is not real. And if your library and your streamable watchlist are both empty, it is told to say there is nothing to watch right now and offer to suggest things to add, rather than improvising picks from its own memory and passing them off as available.

None of this makes the model less useful. It is still the part that understands "something short and funny for a Tuesday" or the film with the submarine and the cook. It does the language. The facts, which title, have you seen it, can you stream it, are answered by ordinary code that has a right answer and looks it up. Splitting the work that way is the whole trick. The model is good at understanding what you meant and bad at remembering the world exactly; so it is trusted with the first and never the second.

If you want the longer story of how the app itself got built, that is written up separately, and the free watch orders and app comparisons are kept current by hand for the same reason this assistant checks its work: a wrong answer stated confidently is worse than no answer at all.

I Like Movies is at ilikemovies.app, and on Google Play.