Mode of interaction

It's a mistake to think of a request to an LLM in the same way as calling a function. Given the same set of inputs in the same order, a function acts predictably. We can write tests and inject faults and harden a function for a wide variety of inputs.

An LLM is not like that. A better way to think about it is as if the LLM were a user and to treat the data we get from them as such. Like a user, an LLM is nondeterministic, often wrong (partially or wholly) and sometimes plain random. To guard our apps under these conditions, we need to build the same guardrails around LLM input as we do around user input.

If we can do that successfully, then we can bring extraordinary abilities to apps in the form of problem solving and creativity that can rival that of a human.

LLMs are good at some things and bad at others; the key is to bring them into your apps for the good while mitigating the bad. As an example, let's consider the task list in the Crossword Companion:

The task list is the set of clues that need solving. The goal is to use colors and solutions in the task list to show progress during the solving process. The initial implementation provided the model with a tool for managing the task list, asking it to provide updates on progress as it went. Flash could not solve the puzzle this way, but Pro could. Unfortunately, it solved it in big chunks, only remembering to update the task list once or twice with a big delay in between. No amount of prompting could convince it to update the tasks as it went. You'll see the same behavior with modern AI agents managing their own task lists; that's just where we are in the evolution of LLMs at the moment.

So how do we get consistent, deterministic updates of the task list? Take task management out of the LLM's hands and handle it in the code.

To generalize, before applying an LLM solution to a problem you're facing, ask yourself whether an LLM is the best tool for the job. Is human-like problem solving and creativity worth the tradeoff in unpredictability?

The answer to that question comes with experimentation. Here are some examples from the sample:

It's a judgement call for sure, but if you can reasonably write the code to do it, your results will be predictable. However, if writing the code would be unreasonably difficult, then consider an LLM, knowing you'll have to build the guardrails like we did in the sample.

There's more than just one pivot to consider besides code vs. LLM. Models operate in roughly two modes: "ask" and "agent".

A LLM is in "ask" mode when we prompt it without giving it tools to affect change in the world, for example, no tools at all or tools just for looking up data. Both the crossword interference model and clue solver models run in ask mode, using tools only for additional data.

On the other hand, when we give an LLM a set of tools that allow it to operate on our behalf in the world – like reading and writing files, executing bash commands, loading web pages, calling web APIs, and so on – that LLM is in "agent" mode.

The difference between ask and agent mode is not the model you choose or the prompts you give it, but the tools you supply. The combination of the tools and the agentic loop described in the Tool calls section allow an LLM to call any number of those tools as often as it decides. Giving it that power puts the responsibility on you to make sure you treat it as unpredictable; more like a person than a program.

You do that the same way that you validate user input, by building up a suite of tests to see how your app works against LLM responses. Give real LLMs a wide variety of prompts and mock the tools to evaluate how the LLM is using them. Like your first user testing experience, your first LLM testing results might surprise you. Use that data to build the guardrails you need to harden your app.

In the sample, we didn't have to guard against harm, but we did have to guard against imperfect results. It was extensive testing against real-world data that led to the institution of human-in-the-loop guards against attempting to solve an invalid puzzle or conflicting solutions. In this way, Flutter and Firebase AI Logic make the perfect combination to harness the power of an LLM and bring unique capabilities to your apps.