A private assistant for your company documents
BENCH 03 / DOCUMENT ANSWERS
Ask a question. Get an answer with its source, or an honest no.
Six documents from a fictional forty-person logistics company: a staff handbook, two procedures, a pricing policy, a client onboarding checklist and a fleet maintenance log. Pick a question. The tool shows which passages it used and how closely they matched. If the documents do not say, it says so.
Open this benchQuestions
Sample data
Answer
Corporate clients can request a refund within 14 days of delivery.
will not: answer without a source
Authored sample · 5 passages scored in code · line at 0.60
Answer
Friday overtime is approved by the operations manager.
will not: answer without a source
Authored sample · 5 passages scored in code · line at 0.60
Answer
Every 10,000 km or every three months, whichever comes first.
will not: answer without a source
Authored sample · 5 passages scored in code · line at 0.60
No answer in these documents
These documents do not answer that. Closest passage: Staff handbook, section 4 (match 0.24, below the 0.60 line).
- ClosestStaff handbook · Section 4
will not: answer without a source
Authored sample · 3 passages scored in code · line at 0.60
Sources disagree
Two documents disagree. Client onboarding checklist v4, page 2 asks for a signed engagement letter, a trade licence copy and an invoicing contact. The 2023 procedure also asked for a deposit; v4 removed it.
will not: answer without a source
Authored sample · 5 passages scored in code · line at 0.60
Passages considered
Match
Pricing policy · Section 3
Corporate clients can request a refund within 14 days of delivery. The refund window starts on the delivery date on the signed proof of delivery. Refund requests from corporate clients after the window go to the finance manager.
Pricing policy · Section 4
Individual customers who pay by card can request a refund within 7 days of delivery. The refund goes back to the same card.
Pricing policy · Section 1
This policy covers rates, surcharges and refunds for every account. Finance owns it and reviews it each January.
Procedure: new client setup (2023) · Section 2
Before we start work for a new client, we need a signed rate agreement, a copy of the client's trade licence and a deposit equal to one month of expected charges. Do not start collections for a new client until the deposit has cleared.
Client onboarding checklist v4 · Page 2
Before we start, we need three things from a new client: a signed engagement letter, a copy of their trade licence and a named invoicing contact. Do not start the first collection for a new client until all three are on file.
Procedure: overtime and weekend work · Section 3
Friday is a half working day. Friday overtime, meaning any work after 12:00 on a Friday, is approved by the operations manager. A shift supervisor cannot approve Friday overtime.
Procedure: overtime and weekend work · Section 2
Overtime from Monday to Thursday is approved by the shift supervisor before the shift ends.
Procedure: overtime and weekend work · Section 4
Work on a Saturday or Sunday is approved by the operations manager at least 48 hours ahead and is recorded on the rota.
Staff handbook · Section 1
Office hours are Monday to Thursday, 7:30 to 16:30, and Friday, 7:30 to 12:00. Warehouse shifts follow the rota published every Thursday. During Ramadan, office hours are 9:00 to 14:00.
Staff handbook · Section 2
Staff with more than one year of service have 30 calendar days of annual leave a year. Request leave in the HR portal at least two weeks ahead. Your line manager approves annual leave.
Fleet maintenance log · Section 1
Every truck is serviced every 10,000 km or every 3 months, whichever comes first. A truck that misses its service is taken off the rota until it has been serviced.
Fleet maintenance log · Section 4
T-07, 7.5 tonne box truck. 9,840 km since last service. Full service, front brake pads replaced. Back on the rota the same day.
Fleet maintenance log · Section 5
T-03, 12 tonne curtain-sider. Three months since last service at 6,200 km. Full service, tail lift hydraulic hose replaced.
Fleet maintenance log · Section 6
T-11, refrigerated van. Unit not holding 4°C. Compressor serviced by the refrigeration contractor. Off the rota for two days.
Staff handbook · Section 2
Staff with more than one year of service have 30 calendar days of annual leave a year. Request leave in the HR portal at least two weeks ahead. Your line manager approves annual leave.
Staff handbook · Section 4
Keep shared desks clear at the end of each day. Visitors sign in at reception and wear a visitor badge while in the office. If you bring a visitor to the office, stay with them until they leave. Visitors do not enter the warehouse floor without safety boots and a hi-vis vest. Eat in the pantry, not at your desk.
Staff handbook · Section 1
Office hours are Monday to Thursday, 7:30 to 16:30, and Friday, 7:30 to 12:00. Warehouse shifts follow the rota published every Thursday. During Ramadan, office hours are 9:00 to 14:00.
Staff handbook · Section 5
Office staff wear business casual. Warehouse staff and drivers wear the issued uniform and safety boots at all times on site.
Procedure: new client setup (2023) · Section 2
Before we start work for a new client, we need a signed rate agreement, a copy of the client's trade licence and a deposit equal to one month of expected charges. Do not start collections for a new client until the deposit has cleared.
Client onboarding checklist v4 · Page 2
Before we start, we need three things from a new client: a signed engagement letter, a copy of their trade licence and a named invoicing contact. Do not start the first collection for a new client until all three are on file.
Procedure: new client setup (2023) · Section 1
This procedure sets out how a new client account is opened. Owner: finance. Last reviewed March 2023.
Pricing policy · Section 3
Corporate clients can request a refund within 14 days of delivery. The refund window starts on the delivery date on the signed proof of delivery. Refund requests from corporate clients after the window go to the finance manager.
Procedure: overtime and weekend work · Section 6
Drivers have at least 12 hours off between the end of one shift and the start of the next.
What the system does
It answers a question from your own policies, procedures and records, and shows the passages it used with the document and section each came from. If the documents do not answer the question, it says so and shows the closest passage it found.
Finding the passages comes first and is plain code: the documents are split into sections, each section is scored against the question, and only sections above a fixed threshold count as evidence. A model then writes the answer from those sections alone. On the bench, the scores are computed by that code on the page.
When two documents disagree, as an old and a new procedure often do, it shows both and says which is newer instead of picking one.
Where a person approves
The assistant answers your staff; it does not act. It sends nothing to clients, changes no document and decides no case.
A named person owns the documents. Questions it could not answer, and documents that disagree, go to that person as a weekly list, so the gap is fixed in the document rather than worked around in the assistant.
What it connects to
It reads from and writes to the systems the business already uses: it reads documents from wherever they are kept, and writes the weekly list to wherever your team tracks work. Which systems those are is the first thing we establish on the first call.
If one of them has no way to connect, we say so on that call.
It keeps the permissions those systems already have: a person only gets answers from documents they could open themselves.
What it does badly
- It cannot answer from documents it was not given. If the answer lives in someone's head or an email thread, it will refuse rather than guess.
- On the bench it finds passages by matching words, so a question worded differently from the document can miss the passage that answers it. A client build also matches on meaning and still misses sometimes, which is why the closest passage is always shown.
- It cannot tell which of two documents is right, only which is newer. An out-of-date document nobody removed is still a source.
- It reads tables, scanned pages and diagrams less reliably than plain text. We check those during the pilot and leave out what it cannot read.
What phase one includes
- One set of documents for one team, chosen with you on the first call.
- 30 to 50 real questions from that team with the correct answer and source for each, used to score the assistant before anyone relies on it.
- A pilot on a copy of the documents, then live use by that team.
- The weekly list of unanswered questions and conflicting documents.
- A log of every question, the passages used and the answer given.
Bring us one task.
Tell us which documents your team searches most often, and we will reply with what we would build first.
Describe a task