Ask every document. Leave none of them.
A search system over the files, mail and scans a logistics or construction company actually runs on, hosted on their network, so a bill of lading never sits in someone else's cloud.
The job is already written down. It is just not in one place.
A single movement produces an agent email, a packing list, a scan of a damaged crate, a spreadsheet of rates, and a line in the system of record. Those five objects do not share a database. They live in a shared drive, a mailbox, a phone, and a folder named FINAL-v3. When someone needs the agreement (what was quoted, what was signed, what the agent in Jeddah actually said) they spend a morning hunting, or they ask the person who remembers.
The wrong file is a wrong number.
In project cargo the document is the job. Quote the packing list from last year and the rate is wrong. Miss the photo of the broken stillage and the claim is already late. Tell a customer the wrong transit because the current instruction is in an email nobody indexed. This is not a knowledge-base problem. It is an operations problem with a commercial bill attached.
An index of the company's own corpus, running on the company's own machines.
We build a search system that connects to the sources the operator already has: mail, file servers, TMS and WMS exports, scanned PDFs, photographs, and keeps a live index of them on their network. People query in language: the BL, the job number, the agent, the date range. The answer is the passage, the file, the page, and the timestamp. It is not a chatbot bolted onto a public model. It is their archive, addressable.
How it works.
Connect the sources that already exist.
Mail, file shares, the TMS export, the scan folder from the yard. We do not ask the company to move its documents into a new repository. The system reads where the work already writes.
Keep enough structure to answer with a source.
A packing list is not a blob. A rate sheet has lanes, dates and currencies. A scan has a page. The index stores the passage and the object it came from, so a result can be opened, checked, and used in the job.
Query in the language the desk already uses.
Job numbers, BLs, agent names, site names, the way a coordinator actually asks. Results respect the permissions of the person asking: the commercial file is not a public document inside the company either.
Leave the documents where they are.
The index lives on their network. Queries do not go to a third-party API. Files are not used to train a model they do not own. If the internet drops, the archive is still there.
What is in the system.
Connectors for the places work already lives
Mail, file servers, exports from the system of record, the scan inbox from the floor. New sources attach without a migration project.
Pages, photos and handwriting
A large share of operational truth is a photograph or a scanned packing list. Those have to be searchable, not stored as images nobody will open.
Permissions that match the desk
Commercial files, customer contracts, and the warehouse scan are not the same audience. The index inherits who can see what, or it is not going in.
An audit of what was asked
Who searched, for which job, what they opened. Needed the first time a customer, an insurer, or a regulator asks how a number was found.
What this replaces.
Search inside one mailbox
The agreement is rarely in the mailbox of the person looking. It is in the mailbox of the person who left, or in a shared inbox with four years of noise.
The shared drive named after a person
Folder trees that only make sense to the person who made them. When they go, so does the map.
Asking the one person who remembers
That person is on leave, on a job, or gone. The system has to remember instead.
The file does not leave the network. That is the product.
Freight files are not blog posts. They carry customer names, rates, packing lists, and in some trades material a regulator would not want sitting in a public cloud. Operators in this work do not have the luxury of 'the model is somewhere in Virginia.' The index, the queries, and the documents stay on machines they control. If a jurisdiction, a customer, or an insurer asks where the data lives, the answer is a room, not a region.
On-prem means on-prem
Their servers, their network, their backup. Not a VPC in a vendor account that happens to be in the same country.
No training on their corpus
The documents are not a dataset for a third party. They are the operator's working record.
A query is an internal event
It does not traverse a public API. It does not become someone else's log line.
If the archive cannot be asked, it is not an archive.
We put this in while the mail and the scans keep arriving. The work does not pause for a clean-up. If your operation is still finding the file by hand, that is the conversation.
Book a scoping call