#Method
#Architecture
#Growth
If the cost of serving a customer changes with every page they generate, what goes on the invoice? Not a seat, which consumes nothing, and not an inve...
Jesse James Richard
Read previous|
May 16, 2026Turning uploaded files into retrievable context
|
7 min read
#Method
#Data
#AI & Agents
A large language model is brilliant and knows nothing about your business. Files are how Giant Context closes that gap. A file here is the material the platform reads to work out what a business is, rather than something to store and hand back. Everything the platform knows about a customer came from their files, and a customer with no files is a customer it cannot do much for.
What a file is for here
A company should not have to pour its information into a set of documents and then pour the same information into a website and an outbound mail campaign, and it should not have to chase down every place a fact appears when that fact changes. Tie the corporate documents to the marketing surfaces, and the marketing stops being a separate job.
When a customer uploads their brand guidelines, a quarter of sales call recordings, a folder of product photos and last year's positioning deck, they are teaching the platform who they are. Everything the AI later writes, every page it builds, every campaign it proposes, is reasoned from that pile. Give it nothing and it writes about a company like theirs rather than about theirs. Give it the real materials and it writes from their own documents, in their own voice, about their own products.
So the file system's job is turning every file, whatever its format, into something the platform can reason over.
Everything becomes text, then meaning
A pile of files is a pile of formats, and a language model reasons over text alone. So the first job is collapsing every format into text.
A PDF or a document is read for its text. An image is classified for what it shows and read for any text printed on it. An audio file, a sales call or a voice memo, is transcribed. A video is transcribed too. This work runs in background workers, because transcription and OCR take minutes rather than milliseconds.
The output of all of it is the same shape. Text, broken into chunks, each chunk turned into an embedding, a long list of numbers that places that chunk's meaning in a space where similar meanings sit near each other. The embeddings live in the same Postgres the rest of the platform runs on. A quarterly report and a podcast episode and a product photo, formats that share nothing, end up as neighbors in the same searchable space if they are about the same thing.
Four formats, one pipeline. Text pulled from documents, labels and printed text from images, transcripts from audio and video, all of it landing in one place where the platform can find what is relevant by meaning instead of by filename.
Two settings on every file
Every file carries two settings. One decides who can see the file. The other decides what the platform is allowed to do with what is inside it.
An image is public because it is going to appear on the customer's website, served to strangers. A confidential strategy memo is private. The platform still reasons from it.
Private governs the file rather than the reasoning. The file is never served to anyone, and everything the platform learns from it can still shape a public page. That is what makes the split worth having. A customer can hand over sensitive material without any of it ever leaking onto a public page, and the material a platform reasons from has to be able to stay private, or nobody will give it anything real.
The other setting decides what the platform may do with the contents. A fact may be asserted as the customer's own claim and cited. Context is read to know where a business sits and is never repeated as that business's claim. I wrote about why that split exists when I built the grounding pipeline.
It matters most where a mistake is expensive. A testimonial or a named team member is a fact or it does not go in, because a generated testimonial that reads plausibly and was never said is a legal problem.
The machine reasons, and points at its source
When the platform writes something, a page, a campaign idea, a piece of copy, it does not write from the void. It searches the customer's embedded files for the chunks relevant to what it is doing, and it reasons from those. And it remembers where they came from. Every idea the platform generates carries the id of the source file and the source chunk it was grounded in. The output can point back at the exact paragraph of the exact document that justified it.
When the platform proposes a campaign, a customer can ask why, and the answer is a citation. Their own Q3 report said this. Their own founder interview said that.
Where this is going
The reasoning layer, Mind, reads everything on record about a customer and decides what that customer's marketing needs. Not waiting for a request. Reading the current state and acting on it. And Mind has exactly one source. The files.
The customer syncs their documents and goes back to running their company. Change a product sheet, and what the platform holds about that product changes, and what it decides to say about that product changes with it. Add a case study, and it has new evidence to argue from. Delete something, and it stops drawing on it. The customer's job shrinks to keeping their own truth current, and the platform's job is to read it, work out what follows, and act.
Wiring this up yourself
If you are building anything that reasons over a customer's own material, this transfers.
Ingestion is translation. Every format the customer can give you, document, image, audio, video, has to become text and then embeddings, or the model cannot think with it. Do that work in background workers.
Put the vectors next to your data. Ours live in the same Postgres as everything else, which means a semantic search is a query rather than a distributed transaction across two databases.
And carry the source through to the output. Store the source file and chunk on everything you generate. It turns the AI said so into your own document said so.
Hiring an early engineer, or building something like this
Remote, Pacific time, full-time or contract. Get in touch.