Document management with semantic search across your contracts

In buying groups, document management always comes last. Someone sets up a network share, creates a folder per supplier, and for five years that is enough. Then a member asks whether their agreement carries the same logistics discount as their neighbour's, and searching the contracts means opening forty files by hand.

The problem is not storage, it is retrieval

A central purchasing body produces few documents, but each of them carries real legal and financial weight. Every one holds a rule that will eventually be applied to an amount, and nobody remembers it at the moment it has to be found.

• The framework approval agreement, with its effective date, its term and its tacit renewal conditions.
• The amendments — often three or four per supplier — that change a rebate scale without altering a line of the original agreement.
• Price grids and logistics terms, attached as annexes and therefore rarely indexed with the agreement itself.
• Administrative papers with an expiry date: insurance certificates, compliance certificates, bank details, mandates.
• Purchasing committee minutes, where in practice half of the rules actually applied are decided.

A network share files all of this perfectly well. It answers no questions. "Which agreements include a logistics discount?", "Which ones expire before March?", "Where is the signed copy of the 2023 amendment?" are management questions, not filing questions. That is why reorganising the folder tree never fixes anything for more than six months.

Full text, metadata, semantic: what each kind of search really does

These three search modes are routinely sold under the same word. They index different things, answer different questions, and cost very different amounts to put in place.

What semantic search across contracts does — and does not do

Semantic search compares the meaning of a question with the meaning of indexed passages. In practice the user asks in plain language, the engine returns the five or six closest passages, each with its document, page and originating supplier. That is not an answer: it is a shortcut to where the answer sits.

That boundary is not stylistic caution. A semantic engine fails in two known ways: it misses a passage phrased unusually, and it offers a plausible passage that is beside the point. The first failure is visible and can be fixed by enriching the index. The second is only visible if the interface always shows the source extract and lets you open the original document in one click. That is a design decision, not a late configuration setting.

Scanned documents, access rights and retention

How we deliver it

1. 1

   Inventory of document types

   Two workshops to list the families actually produced, their mandatory metadata and who may consult them. That document, not the tool, is what determines search quality.

2. 2

   Migrating what exists

   The network share is taken as it is, OCR'd, then attached to suppliers, agreements and member companies. Whatever no automatic rule can attach is listed and handled by hand: there always is some, and it is better counted at the start than discovered at go-live.

3. 3

   Indexing and acceptance on your own questions

   Full-text, metadata and vector indexes are built over the same corpus. Acceptance runs on about thirty real questions supplied by your teams, not on demonstration queries chosen by us.

4. 4

   Rights, retention, switchover

   Access opened per company, retention periods configured, a short user briefing, then the network share set to read-only for a quarter before archiving.

MEKANO is a Lyon-based development studio building the management applications of buying groups and member networks. Document management is rarely the first batch of a project: it usually follows supplier approval and revenue declarations, because it only becomes valuable once documents are attached to a supplier, an agreement and a financial year. The group of more than 60 member companies for which we delivered this platform brought it live in exactly that order.

What usually goes wrong

• Migrating the network share without arbitration: fifteen years of duplicates and drafts become fifteen years of noise in the results.
• Assigning metadata entry to nobody in particular. A mandatory field with no owner stays empty.
• Expecting semantic search to compensate for missing structure: it amplifies good document organisation, it does not replace it.
• Opening the system to all members before settling access rights, then restricting afterwards — a document that has been read cannot be unread.
• Digitising without deciding retention periods, and ending up with a repository that cannot be purged without a legal ruling.