Your company’s own AI, on your own network.

How a file gets read depends on what it is.

Word documents

Read paragraph by paragraph and table by table, in the order they appear. An extractor that only reads paragraphs silently drops every table — and tables are often where the real content lives: line items, schedules, comparison grids.

PDFs

Read page by page. If one page’s text cannot be pulled directly — because that page is really a scanned image inside an otherwise normal PDF — that page alone runs through OCR, so a mixed document does not lose the pages that happen to be pictures.

Spreadsheets

Read cell by cell against their actual row and column headers, rather than assumed to be clean tables. Most real business spreadsheets are not.

Photos and scans

A small vision model writes a plain-language description of what is in the image. The photo’s own EXIF data — when it was taken, and where, if the camera recorded GPS — is folded into what gets indexed, so a photo can be found by date or place even if nobody wrote a caption for it.

A document is too long to hand a model whole, so it gets split into smaller pieces. But a piece that starts mid-section with no idea what it belongs to is a piece waiting to be misread. Every piece carries a breadcrumb back to the section and sub-section heading it came from — so a bullet point still knows which employer, which project, which year it is talking about, even once it has been separated from the heading above it.

This matters more than it sounds. Real documents extract inconsistently between formats, and getting the structure — headings, dates, section boundaries — to survive that extraction reliably took real tuning against real documents, not just the tidy synthetic ones.

A second pass reads the document’s actual content — with the filename treated as a hint, never a verdict — and pulls out a governed set of attributes: the kind of document it is, the customer or vendor it names, the people involved, any document number, dates, and dollar amounts.

“Governed” means the list of valid document types is a fixed, versioned vocabulary rather than whatever word a model feels like using that day, so the same kind of document gets labeled the same way over time.

Every piece goes into a search database that understands both meaning — so “how much did we charge” can find text that says “total due” — and keyword matching, so an exact SKU or invoice number is never missed for lack of an obvious meaning.

Spreadsheets get a second, different home: their rows also go into a real SQL database, because “how many invoices” is a database question, not a search question, and only a database gives an exact answer.

• Backup copies and files left over from a failed ingest attempt are recognized and excluded, so they never compete with the real document for a search result.
• Large documents are embedded in batches rather than one giant request, so an oversized file cannot time out and take the whole pipeline down with it. A batch that fails is retried with a growing wait instead of hammering the same failure in a tight loop.
• A document is marked indexed only after every step for it has actually succeeded. A document that failed partway through can never be cited in an answer as though it were whole.

Some questions want one specific document. Some want a complete set — every invoice for this customer — where a partial answer is worse than a slow one, because it will be silently wrong. Some want a fact synthesized across many documents, like every employer a person has worked for.

That last kind is a different problem entirely: the fact might be scattered across a resume, several cover letters, and a reference letter, and the job is to compile one accurate answer rather than hand back whichever fragment happened to rank highest.

A follow-up like “what about the other one” means nothing on its own. It gets resolved against what was actually discussed, so a subject-less follow-up does not go searching for whatever happens to be semantically nearest to those four words.

If someone says “only look in the resume folder,” that restriction is honored and stays in effect for the rest of the conversation until it is changed. It does not quietly get dropped on the next question.

Every search runs both a meaning-based search and a keyword search, and the two result sets are combined. A document using different words than the question, and a document carrying the exact code or number the question named, both get a fair shot at being found.

The system pulls likely filters out of the question itself — a customer name, a document type, a date range — but a filter only ever boosts a document’s ranking. It never removes a document from consideration.

That distinction matters because filter extraction is not perfect. An apostrophe, a job title that reads like a person’s name, any number of small misreads — and a hard filter would let one of those make the right document invisible.

The one filter that is absolute is ownership. A user only ever sees results scoped to their own files and whatever has been explicitly shared with them, checked on every single search the system runs, with no exceptions.

The system pulls back more candidates than it needs and re-scores them with a second, more careful model built specifically to judge relevance rather than similarity, before deciding what goes into the answer.

If nothing clears the bar, it says so — “I could not find that” — instead of answering from whatever happened to be closest. A system that always answers with something, even when nothing is actually relevant, is worse than one that sometimes admits it does not know.

For “list every invoice” questions, the system does not stop at a fixed number of results. It returns the complete set, with a true total, even when that set runs into the hundreds.

How many, what is the total, which one is biggest — these get routed to the real SQL database built from your spreadsheets, instead of being answered by having a language model eyeball a handful of retrieved text fragments and guess.

Counting a fragment of a spreadsheet and calling it the answer produces a confident number that is often wrong. A database query does not have that failure mode.

Every answer cites what it is based on. And the system is taught, explicitly, what a document actually proves: a cover letter names the person it was addressed to, which is not evidence that person worked there. A quote or a purchase order names a counterparty, which is not by itself evidence of an ongoing relationship.

Getting that distinction right is the difference between an assistant that sounds confident and one that is actually correct.