Three hundred pages, and everything in them that says “shall”.
A solicitation package is a legal document set with the requirements distributed through it on purpose. Document intelligence reads the whole package — including the scanned attachment and the amendment that arrived on Friday — and produces a requirement list, an evaluation criteria list and a compliance matrix, each item citing the page it came from.
- Produces
- Requirements, criteria and a compliance matrix
- Every item carries
- The document and page it came from
- Handles
- Scanned pages, amendments and cross-references
The requirements are not in one place, and they are not all in the scope.
The obligations in a solicitation are spread across the statement of work, the instructions to offerors, the evaluation section, the special contract requirements, the attachments and every amendment. They are written as “shall”, “must” and “will”, they cross-reference each other, and a handful of them are in a table inside a scanned PDF.
Reading them out by hand is a two-day job for somebody senior, and it produces a spreadsheet that is out of date the moment an amendment lands. Most teams do it once, on the first version, and patch it from memory afterwards.
The failure is rarely dramatic. It is a submission that missed one mandatory form, an evaluation criterion nobody weighted the response against, or a page limit that changed in Amendment 3 and left a technical volume fifteen pages over.
- A compliance matrix nobody trustsBuilt by hand from the first version and patched informally, it stops being the authority the week after it is written.
- Amendments applied unevenlyA changed due date is noticed; a changed page limit or a new requirement in the same amendment is not.
- Two days of senior time per packageThe person who can read a solicitation properly is usually the person who should be writing the technical approach.
From a folder of PDFs to a matrix that stays current.
The pipeline runs in the background — a hundred-megabyte package is not something to read inside a web request — and every output is traceable back to a page.
Ingest the package
Every document is stored as published, typed, and queued for extraction; scanned pages go through OCR and are marked as scanned.
A package with each file recorded, and pages that a citation can point at.
Find the obligations
Sentences carrying an obligation are identified with their reference numbers and the section they sit in.
A requirement list distinguishing mandatory items from permissive ones.
Read the evaluation section
Criteria, their weights and the relationship between technical and price are extracted from the evaluation instructions.
The scoring model the response has to be written against, with each criterion cited.
Resolve references and conflicts
Cross-references between sections are followed, and clauses that contradict each other are flagged rather than silently reconciled.
A list of conflicts for a person to resolve, usually by asking before the question deadline.
Build the matrix, and keep it
Requirements are mapped to response sections; an amendment updates the matrix and marks what it touched.
A compliance matrix derived from the documents, current as of the last amendment received.
A page being read, the requirements that came out, and an amendment landing.
The facsimile on the left is decorative — the extracted list beside it is the content. Underneath are the two things a screenshot never shows: what an amendment does to work already drafted, and what happens when two clauses in one package disagree.
Laboratory instrument maintenance — solicitation package
Scanned pages go through OCR first, and a scanned page is marked as one on every requirement taken from it — an extraction you cannot tell was read by a machine is an extraction you cannot check.
- obligation: shall
The contractor shall maintain calibration records for every instrument covered, traceable to a national standard.
- obligation: must
Preventive maintenance must be performed within the interval stated by the manufacturer, and never later than annually.
- obligation: shall
Technicians entering laboratory space shall hold current safety training for the hazard classes present.
- obligation: may
Offerors may propose an alternative response interval, which will be evaluated on risk rather than on price alone.
Amendment 0003 landed — what it changed
- Response dateEvery schedule derived from the old date moved with it, including the internal reviews.
- Technical volume page limitSections already drafted against the old limit are flagged over-length, with the overage per section.
- New requirement C.5.9Added to the compliance matrix as unanswered, and to the outline as a section with no owner.
Two clauses disagree, and neither was silently chosen
Totals are from this fixture package. Nothing here claims an extraction accuracy figure — every requirement carries its citation so it can be checked against the page it came from.
Nothing above is a live query or another customer’s pipeline. Every record in it was written for this page, and no buying organization named in it is real.
What reading a package properly requires.
Most of these exist because of a specific way proposals are disqualified.
Citations on everything
Document and page on every requirement and criterion. An extraction you cannot check against the source is an assertion with a reference number.
Obligation strength preserved
Shall, must, will and may are not the same word. Mandatory items are counted separately, because those are the ones that make a submission non-responsive.
OCR, marked as OCR
Scanned pages are read, and every requirement taken from one says so, because a reader who cannot tell is a reader who cannot check.
Amendment differencing
What changed, what it affected and what work is now over-length or unanswered — applied to the matrix rather than announced in an email.
Conflicts surfaced, not resolved
When two clauses disagree, both are shown with their citations and flagged for a person. Silently choosing one is how a page limit is breached.
The matrix as a derivation
The compliance matrix is generated from extracted requirements and the sections that answer them, so it cannot drift from the response the way a maintained spreadsheet does.
Where the package is the problem.
Three teams whose difficulty is document volume rather than writing.
A federal bidder
- Packages run to hundreds of pages across Sections C, L and M plus attachments, and amendments arrive weekly.
- Full extraction with the evaluation criteria weighted, and amendment differencing applied to the matrix.
- The compliance matrix is the authority again, rather than a spreadsheet somebody maintains alongside it.
An architecture and engineering firm
- Requirements live in scanned specification sections and drawing packages nobody can search.
- OCR with per-page citation, so a specification requirement can be traced back to the sheet it came from.
- The specification is searchable evidence instead of a stack of PDFs opened one at a time.
A software company answering questionnaires
- Security questionnaires arrive as spreadsheets with hundreds of items, each needing a traceable answer.
- Requirement extraction over the workbook, mapped to the content library entries that answer each item.
- Answers are reused with their sources attached, and gaps are visible before the deadline rather than after.
Document intelligence is on every plan, metered by pages.
Every plan reads packages, extracts requirements and builds compliance matrices. What scales is volume: document pages a month, OCR pages a month, and how much storage the packages and exports occupy.
- Every plan includes extraction, citations, evaluation criteria and compliance matrices.
- Document page and OCR page allowances rise with each plan; OCR is metered separately because scanned pages cost more to read.
- Document storage in gigabytes rises with each plan, and packages count towards it.
- When an allowance runs out, the limit is refused with the reason, rather than the work quietly degrading.
Plan names and allowances are mirrored from the billing catalog. Amounts are on the pricing page, which reads them from Stripe rather than from a number typed into a marketing page.
What people ask about document intelligence.
What formats can it read?
The formats solicitations actually arrive in: PDF, including scanned PDFs through OCR, Word documents and spreadsheet workbooks such as pricing schedules and questionnaires. Every extracted item records which file and which page it came from.
How accurate is the extraction?
No accuracy percentage is published here, because a figure with no study behind it is marketing. What the product does instead is make every item checkable: each requirement carries its citation, mandatory items are counted separately, and conflicts are flagged rather than resolved. A reviewer can verify a page in seconds.
What happens when an amendment arrives?
It is ingested against the same package and differenced: changed dates, changed limits and new requirements are identified, the compliance matrix is updated, and anything already drafted that the change affects — a section now over-length, a requirement now unanswered — is flagged.
Are our solicitation documents used to train models?
No. Packages, extracted requirements and everything derived from them serve your organization and are not training data for a shared model.
Does it handle questionnaires as well as solicitations?
Yes. A security questionnaire or a vendor workbook is a requirement list in a different container, and it is extracted the same way — items, obligations and the sections or library entries that answer them.
Give it a package you have already read.
Upload a solicitation your team has been through by hand and compare the requirement list, page by page, against the matrix somebody built.