Every file read
at the detail it needs
We build features and products on the Gemini API, with each document, photo and video read at the detail it needs, replies your code can read, and tests run before every model change, so you pay for detail where it matters.
Thanks, we’ll be in touch soon.
{{ heroStatus }}
You’re in good company
A few of the companies we’ve worked with.
When the prototype meets real files
Many features on the Gemini API start in Google AI Studio, on unpaid use, tried on a few clean files. A request is what the product sends: its instructions, and files such as photos, PDF pages or video. A model answers it with a reply.
Then real files arrive: long PDFs, phone photos, videos. How much detail the model reads in each is left at its default, or set once for the whole request, whatever each file holds. Every task gets the model’s default thinking, the reasoning it does before it answers, however simple the task. Replies are read by matching their wording, the prototype’s key carries on, and a model name left as it was can be routed to a newer model.
Cost and waiting grow with every file, and replies change though nobody changed the code.
The prototype did its job: it showed the idea works.
A product needs its settings chosen on purpose: the detail for each file, the thinking for each task, a set shape for replies, paid use for real data, and a test set, real cases with the answers you expect, to check every change.
What we build
Files, thinking and replies set for each task, real data on paid use, and a test set behind every change, for one feature or a whole product.
Settings for each task
Instructions, examples, a model and a thinking level chosen for each job: the lightest that passes your tests.
Replies in a shape you set
Each reply set to follow a schema, the fields and kinds of value it must have, so your code reads values, not wording, and checks each one.
Files read at the right detail
Documents, photos and video sent with the detail each one needs, set file by file.
Actions in your systems
Functions the model can only ask for: your code checks each request, decides what runs and runs it.
Tests from your own files
Photos, pages and videos from your own work, each with the answer it should get, run whenever a setting or a model changes.
Cost and waiting kept in check
Detail and thinking set for each task, work that can wait sent in batches, and limits for each project.
Paid use and keys
Real data only on paid use, a key for each project and environment, and a record of who holds which.
Moves before a retirement
Each deprecation read early, the next model tested on your files, then switched on a date you choose.
After launch
Failed requests, blocked replies and the cost of each feature tracked, and the answers people flag reviewed with your team.
- Assistants that take actions in your systems?AI Agents
- Answers for customers and staff, from your own documents?AI Chatbots & Assistants
- Your product inside the assistants people use?MCP & Agent Integration
Detail, set file by file
A request can carry photos, PDF pages and video. How much detail the model reads in each is a setting: on the current Gemini models it can be set once for the whole request, or file by file. Left unset, a default tuned for each kind of file applies.
More detail reads fine print and small parts better, and it also costs more and takes longer. A receipt, a photo and a video in one request rarely need the same detail.
So a team can check the replies on its test set at each setting, and keep the lightest that still reads what the reply needs.
A claim sent for review
An exampleThe instructions
Check the claim against its files
A scanned receiptMedium detail
The reply needs its small print. For document pages, the guide Google publishes finds medium detail usually enough: more detail rarely reads a standard page better.
The small print reads, without paying for detail the page doesn’t need.
A photo of the damageHigh detail
The reply needs to see the damage closely. The guide Google publishes recommends high detail for most images, since it reads small details better.
Small parts show clearly, and this one file costs more and takes longer.
A walk-around videoLow detail
The reply needs what happens from frame to frame, not fine print. The guide Google publishes finds low detail enough to describe most video, and keeps high detail for frames with dense text or small details.
Each frame is read lightly, so a long video costs less and comes back sooner.
The question
What was bought, what’s damaged, and does the video match?
One request, a different detail for each file.
Know which model is answering
Google names each model as stable, preview, latest or experimental. A latest name points to the newest release of a model, which can be a preview, so a product names the stable model its tests passed on.
The deprecations page gives each model its earliest shutdown date and a recommended replacement, and Google confirms the exact date with notice. Some deprecated names aren’t turned off: requests to them are routed to a newer model automatically, so the model answering a feature, and its price, can change with no change to the code.
The docs now mark the generateContent guides as legacy, beside a newer Interactions API in beta, and still recommend generateContent for stable production work. A move between the two is one more change tried on the test set first.
A product on a deprecated model gets a next step: its replacement run on the test set, and the switch made before the date.
Start with one feature and its files
The review covers the models in use and where each stands on the deprecations page; paid or unpaid use, project by project; how requests are built, which files go in and at what detail; thinking levels; how replies are read; the test set, or the cases that could become one; the tools and what each can do; keys, projects and limits; the data settings; errors, blocked replies and rate limits; which API each request goes through; and where the cost of a request can come down.
Then the first changes, in the order we’d make them, starting from a test set.
What you receive is a written review and a proposal for that first step.
Questions worth asking
Does it matter whether we’re on paid use?
Yes. On unpaid use, which covers Google AI Studio and the API’s unpaid quota, Google uses what you send and the replies to improve its products, people may review them, and its terms ask that nothing sensitive, confidential or personal be sent. On paid use, through a project with an active billing account, Google doesn’t use your prompts or replies to improve its products, and keeps them for a limited time only to catch misuse and meet legal duties. We keep real data on paid use only, set project by project, and write it down.
How much should the model think on each task?
As much as the task needs. By default the model sets its own effort for each request; a thinking level changes that. Google suggests the lightest levels for simple tasks such as sorting or looking up a fact, the default for moderate ones, and the most for complex planning. Thinking is billed with the reply, so a lighter level costs less and answers sooner. We set a level for each task by your tests, lighter wherever the task is simple.
What happens when a model we use is deprecated?
Google announces it, lists its earliest shutdown date and a recommended replacement on the deprecations page, and confirms the exact date with notice. On that date the model is turned off, or, for some names, its requests are routed to a newer model instead, as the page says for each. We run the replacement on your test set well before then, so the move happens on a date you choose.
Can we use the same models through our cloud account?
Yes. Google also offers Gemini models through Gemini Enterprise Agent Platform, on its cloud, under that platform’s own terms and controls. That’s our cloud work: software development with Google Cloud services.
What does it cost to work with you?
It depends on what your features read and do, and on the first step you choose. Before any work starts, you get a written proposal with the scope, the price or rate, and how you pay.
Can we hire developers for our own team?
Yes, through a dedicated team: Webair developers who stay with your product over time and work beside your own people.
Gemini, Google AI Studio and Google Cloud are trademarks of Google LLC, and the other Google product names on this page may be too. Webair is independent of Google, and isn’t sponsored by, endorsed by or affiliated with it.
Let’s build it better
Tell us which files your product reads and what it does with each reply, and we’ll suggest where to begin, in a short call.
Thanks, we’ll be in touch soon.
{{ ctaStatus }}