Semantic clause search: finding the right clause, not just a keyword match

A lawyer needs the firm’s standard position on capping liability. They search the clause library for “liability cap” and get nothing useful, because the approved clause opens “the aggregate liability of the Supplier under or in connection with this Agreement shall not exceed…”. The clause is there. It is approved, it is current, and the search could not see it, because the lawyer and the draftsman used different words for the same idea. Semantic clause search exists to close exactly that gap — retrieval by meaning rather than by matching characters. It is a real improvement and it is not, on its own, the answer. This piece sets out what semantic search does well, where it performs worse than plain keyword search, and why the thing that most determines whether a clause library gets used is not the search method at all.

The vocabulary problem keyword search cannot solve

Keyword search matches tokens. It is fast, it is predictable, and when you know the wording it is the best tool available. Its limitation is that it requires the person searching to guess the words the author chose.

In drafting, those words diverge constantly. A lawyer thinks “change of control”; the clause says “if any person acquires the ability to direct the affairs of”. A lawyer thinks “we can walk away”; the clause is headed “termination for convenience”. A lawyer thinks “penalty for late delivery”; the clause says “liquidated damages” and the firm’s house style never uses the word penalty at all.

The practical consequence is not that the search returns nothing. It is that the search returns something, the lawyer concludes the library does not have what they need, and they go back to the document they used last time. That is how a well-built clause library quietly stops being used — not through rejection, but through a series of small unsuccessful searches.

Semantic clause search addresses this directly. Instead of matching words, it compares meaning, so a query phrased the way a lawyer thinks can retrieve a clause phrased the way it was drafted. As Alex Smith of iManage puts it in a piece on search and findability in legal knowledge work, “search is a tool; findability is the outcome” — and the outcome is what the firm is actually paying for.

💡 Insight

It is worth watching how lawyers search rather than asking them. Firms often find that the queries people actually type are shorter and more conversational than anyone expects — two or three words describing a situation rather than a clause heading. A library indexed only against formal headings will miss most of them.

Where semantic clause search performs worse than keyword search

It is worth being honest about this, because the market generally is not. Retrieval by meaning has specific, well-documented failure modes on legal text, and they are not edge cases.

Terms of art collide with everyday language. Legal English reuses ordinary words with specialised meanings, and models trained on general text learn the ordinary sense. “Consideration” means thinking carefully to most language models and something quite different to a contract lawyer. “Service” means customer service rather than formal delivery of process. “Sanction” means a penalty, and also its opposite. An analysis of hybrid search across millions of legal passages sets these collisions out in detail, along with two further problems: identifiers such as a section number carry identity rather than meaning, and negation gets compressed away — “except as provided in subsection (c)” and “notwithstanding any other provision” are exactly the low-frequency, high-consequence phrases that flip a clause’s effect and that similarity scoring tends to flatten.

For a clause library, the last of those matters most. A lawyer searching for a mutual indemnity does not want the one-way version ranked first because the two are textually similar. The difference between them is the entire point.

So the sensible position is not semantic search instead of keyword search. It is both, with the structure the firm has already built sitting on top of them.

What good retrieval actually looks like in a clause library

XpressDox Intelligent Clause Management provides several routes to the same content, and in practice lawyers use different ones for different situations. That variety is the feature.

Keyword search, for when the wording is known. It remains the fastest route and it is exact, which matters when you are looking for a specific defined term.

Short codes, for the clauses an individual uses constantly. A few characters typed into the document beats any search, and experienced users converge on this quickly.

Categories and tags, for browsing by practice area or document type when you are not sure what you are looking for but you know roughly where it lives.

Favourites and recently used clauses, which quietly handle a large share of day-to-day drafting because most lawyers reuse a small working set.

Semantic and natural-language search, for the case this article opened with — where the lawyer knows what the clause needs to achieve but not how the firm words it.

A library that offers only the last of those is harder to use than one that offers all of them, which is the opposite of how retrieval is usually marketed. The best outcome is that a lawyer rarely searches at all, because short codes and favourites cover the routine work and search is reserved for the unfamiliar.

💡 Insight

Firms often find retrieval performance is decided by how the library was structured rather than by which search method is switched on. Consistent categories, meaningful tags and clause titles written for a searcher rather than a drafter improve every route at once. It is worth investing there before evaluating search technology in detail.

The part that decides everything: what is being searched

Here is the argument we would make more strongly than most vendors would find comfortable. Search quality is a secondary variable. The primary one is whether the content being searched has been governed.

Consider the same query run against two collections. The first is the firm’s document history — everything ever produced. Good semantic search will find several plausible clauses, and the lawyer has no way to tell which one the firm stands behind, when it was last reviewed, or whether it was a concession made for one client in one deal. The retrieval worked perfectly and the lawyer is no better off.

The second is a governed library, where every clause carries an owner, an approval state, a version history and any approved variants. The same query returns fewer results, and each one answers the question behind the search: is this what we say? That is why we treat structured legal precedent management as the foundation and retrieval as the interface to it.

The same logic applies to search across a document management system. Firms running iManage, NetDocuments or SharePoint alongside the clause library — see the technology integrations page — are searching two different kinds of collection, and it is worth being clear internally about which one answers which question. The DMS tells you what the firm has produced. The clause library tells you what the firm has approved.

Where AI sits in this, and where it does not

Semantic and natural-language retrieval are AI capabilities, and we would rather describe their role precisely than lead with the acronym.

What they do: reduce the distance between how a lawyer thinks about a problem and how a clause was drafted. What they do not do: decide whether a clause is correct, current or appropriate for the matter. That judgement stays with the lawyer, and the governance layer is what gives them the information to exercise it — who owns this, what state is it in, when did it last change.

This is what we mean by governed AI workflows. The predictability does not come from constraining the model. It comes from the fact that everything the model can reach has already been approved by a named person. Point the same capability at an ungoverned collection and it will be just as clever and considerably less useful.

Common mistakes when evaluating clause search

Five come up repeatedly in evaluations.

1. Testing with clause headings. Evaluators type the formal name of the clause they are looking for, which is the one query almost guaranteed to work. Test with the words a lawyer would use under time pressure instead.
2. Testing on a demonstration library. A curated set of two hundred well-titled clauses will make any search look excellent. Ask to test against something closer to your own content, including the awkwardly titled clauses.
3. Judging semantic search only on what it finds. Also check what it ranks highly and should not — the one-way indemnity when you asked for mutual, the clause carrying a carve-out you did not ask for. Precision matters more than recall in drafting.
4. Treating search as a substitute for structure. Better retrieval over a disorganised library returns better-ranked guesses. Categories, tags, ownership and approval improve every search route simultaneously.
5. Ignoring the routes that are not search. Short codes, favourites and recently used clauses handle much of the daily volume. A system evaluated only on its search box will be assessed on the part lawyers use least often.

💡 Insight

It is worth building a short evaluation set before any demonstration — ten real queries in the words your lawyers would use, with the clause you expect back written down in advance. Firms often find this single exercise separates products more clearly than a feature comparison, and it costs an hour.

Our view: retrieval is a governance question wearing a technical costume

After fifteen years working with firms in more than 25 countries, our position is that semantic clause search is worth having and is rarely the constraint. The firms struggling to get value from a clause library are almost never struggling because the search is insufficiently clever. They are struggling because the library contains content nobody has approved, or because it sits outside the drafting environment, or because no one owns it.

That is why we build retrieval into the wider platform rather than treating it as a headline capability. The library lives inside Microsoft Word, because a search that requires leaving the document competes with copy and paste and copy and paste is very hard to beat. And every result carries its governance with it, because a clause without an owner and an approval state is a suggestion rather than an answer.

If that sounds like an argument against our own capability, it is not. It is an argument about sequence. Get the content governed, then let better retrieval make it easy to reach. Doing it the other way round produces a fast route to material the firm has not decided about.

Where to start

Take ten queries in your lawyers’ own words — not clause headings — and run them against whatever you use today. Write down what you expected and what came back. That gap is the practical case for better retrieval, and it is more persuasive internally than any demonstration, because it is your firm’s vocabulary against your firm’s content.

If it would help to run that exercise against a governed library and see the difference, that is a conversation our team has regularly. You can book a demo or write to us at support@xpressdox.com.

Frequently asked questions

What is semantic clause search?

Semantic clause search retrieves clauses by meaning rather than by matching words. A lawyer can search for the effect they need — capping liability, walking away from a contract — and find the approved clause even though it is drafted in different language. It closes the gap between how lawyers describe a problem and how the clause was written.

Is semantic search better than keyword search?

Not universally, and the useful systems run both. Keyword search is faster and exact when you know the wording, and it handles defined terms, section numbers and terms of art more reliably. Semantic search handles the case where you know the outcome but not the phrasing. Legal language makes each one weak where the other is strong, which is why hybrid retrieval outperforms either alone.

Will semantic search return the wrong clause?

It can, and this is the thing to test for. Clauses that are textually similar but legally opposite — a mutual versus a one-way indemnity, a provision with and without a carve-out — are exactly where similarity scoring struggles, because negation and exceptions are short phrases that carry large consequences. Evaluate on precision, and keep approval states visible so the lawyer can see what they are being offered.

Does better search fix a disorganised clause library?

No. Better retrieval over ungoverned content returns better-ranked guesses, because nothing in the result tells the lawyer which version the firm stands behind. Ownership, approval states, categories and version history are what make a result usable; search is the interface to them rather than a substitute for them.

How should we evaluate clause search during a demonstration?

Write ten real queries in advance, in the words your lawyers would actually use rather than formal clause headings, and note the clause you expect back for each. Then run them against content resembling your own, not a curated demonstration set. Check what is ranked highly and should not be, as well as what is found.