Documents and AI
Very long documents
Two hundred pages of which four values matter. How to approach it.
A tender pack, a policy or a framework contract can run to two hundred pages. Of all of them, what you need usually fits on half a page: four dates, two amounts and three conditions. The problem is not reading it, it is not having to read it again every time someone asks.
What changes versus a short document
| Aspect | Short document | Long document |
|---|---|---|
| The reading | Straightforward | The same, just slower |
| What gets extracted | Almost everything relevant | Only the fields you choose |
| The risk | One value misread | A correct value from the wrong place |
| The review | Read it in full | Check field by field |
Important
The third row is what is specific to long documents. Two hundred pages hold many dates and many amounts; the typical error is not misreading a number, it is taking the correct number from the wrong clause.
How to approach it
- 1
Decide beforehand which four or five values matter
If the list runs to twenty, reviewing will take as long as reading.
- 2
Check each value against its place in the document
A plausible value is not enough: it must come from the right clause.
- 3
Keep it linked to the file, not loose
The document lives where it is used; the data, where it is consulted.
- 4
And note the doubt if there is one
An ambiguous condition is not resolved by extracting it: flag it for someone to read.
Watch out
Do not use an automatic summary to take a contractual decision. It helps you orient and know which clause to open; what binds you is the text, not the summary.
When the document is also scanned
A two-hundred-page tender scanned at poor quality is the worst case: a lot of text and little reliability. If it is an important document that will be consulted for years, it pays to obtain the electronic original rather than to fix the scan's reading.
Worth knowing
Upload size limits are the same as for any document. If a file will not fit, it is almost always because it was scanned at a far higher resolution than needed, not because it has too many pages.
›Does it take much longer?
Longer than a short one, yes. Not a problem if you upload it when it arrives rather than when you need it.
›Can I search inside it afterwards?
Yes, and that is half the value: finding the clause without going through it all.
›What if only three pages interest me?
Upload the whole document and extract only from those: the rest stays there just in case.
A real case
The situation
A company wins a 180-page framework contract and every query means rereading it.
What you do
- Extracts six key values and checks each against its clause
- Keeps the document linked to the client file
What you get
Queries are answered on the spot and the contract is only opened when the text must be quoted.
The situation
A two-hundred-page document takes a while to process.
What you do
- Leaves the process running and moves on
What you get
Processing time blocks nobody.
The situation
Only one figure is needed from a very long document.
What you do
- Extracts only the fields that will be used
What you get
Review narrows to the useful.
The situation
The document exceeds the size limit.
What you do
- Splits it into meaningful parts
What you get
Each part goes in and is locatable.
The situation
Searching inside the document by hand is impossible.
What you do
- Searches by content within the document
What you get
You reach the exact point.
The situation
The whole thing is uploaded when only two annexes mattered.
What you do
- Uploads what will be consulted
What you get
The archive does not grow without reason.
This article answers
- extract data from a very long pdf
- document with hundreds of pages
- reading a full tender document
- pulling information from a long contract