Natural language search, relevancy ranking, semantic search, vector search, and hybrid search have all reached today's legal research platforms. Lawyers meet these terms in product descriptions and demonstrations, often with little explanation of what each one means. This chapter explains each technique and how it works. Each is a retrieval tool. An AI agent uses the same ones when it performs the legal research itself. The chapter closes with the line between a better search box and agentic legal research.
From Keywords to Questions
What is natural language search in legal research?
Natural language search in legal research accepts a question in plain English rather than a query built from terms and connectors. The platform extracts the significant concepts from the question, weighs them, and returns a ranked list of the cases it predicts are most relevant. The researcher asks in the language of the problem, with no query syntax to build.
The difference is visible in the query itself. A terms-and-connectors search for a premises liability question might read "slip /s fall /s notice", with every term guessed and every connector chosen by the researcher. The natural language version simply asks when a store owner is liable for a customer's fall.
The major legal research platforms have offered natural language modes since the early 1990s, so the plain-English question is a convenience, not a new form of retrieval. The platform assigns each extracted concept a weight and scores every document on how prominently those concepts appear. For most of the technique's history the matching underneath remained term-based. The vocabulary problem Chapter 3 documented persisted behind a friendlier prompt. A question asked in the client's language still missed the opinion phrased in the judge's words.
What is relevancy ranking in legal research?
In legal research, relevancy ranking orders search results by a score of predicted usefulness rather than by date, court, or citation count. The score blends many signals, including where and how often the query's concepts appear, how later courts have treated the case, and how the platform's other users have interacted with it. Ranking determines what the researcher actually reads, because few researchers read past the first pages of results.
The scores operate only on what the search retrieved, so ranking reaches nothing the query missed. The scoring rules are proprietary and differ by vendor. Susan Nevelow Mart ran the same searches on six legal research platforms in The Algorithm as a Human Artifact and found that each platform returned a markedly different set of top results. A ranking is an editorial judgment rendered by rules nobody outside the vendor can inspect. Chapter 3 reports her measurements and examines how interfaces and ranking algorithms bias legal research.
Searching by Meaning
What is semantic search in legal research?
Semantic search in legal research is retrieval that matches on meaning rather than on shared words. Instead of asking whether a document contains the query's terms, it asks whether the document says something close to what the query means. That difference decides whether a controlling authority phrased in different vocabulary is found or missed.
A negligence case about a serving cart on an airplane and one about a falling retail display may share no keywords at all while raising the same doctrine. Keyword search sees two unrelated fact patterns. Semantic search sees the same legal question. The same matching works across time, connecting a present-day question to an older controlling case phrased in the vocabulary of its era. Commercial legal research platforms now offer it, so a lawyer can use it today from an ordinary search interface.
The technique differs from the two above at the level of matching. Natural language search changes how the researcher phrases the query. Relevancy ranking changes the order of the results. Semantic search changes what the system matches, replacing shared words with shared meaning as the basis of retrieval. The next two questions explain the machinery underneath, the embedding and the vector search that runs on it.
What is an embedding in legal AI?
An embedding in legal AI is a numerical representation of a passage's meaning, a fingerprint that can be compared mathematically. Two passages with similar embeddings say similar things even when they share no vocabulary. Embeddings are what make semantic search work.
The numbers come from a language model trained on very large volumes of text. The model converts a passage into a long list of numbers. Passages that say related things receive lists that sit near one another. Distance between the lists then measures similarity in meaning. The list itself is called a vector, which gives vector search its name.
Embedding opinions passage by passage lets a system match a query to the specific paragraph stating a rule of law, rather than to a whole opinion that addresses many issues. Embeddings also power the connection-finding described in Chapter 5, where similar embeddings surface doctrinally related statements across millions of opinions.
What is vector search in legal research?
Vector search in legal research is retrieval that runs on embeddings, returning the passages closest in meaning to the query. The system embeds the query, compares it against the embedded corpus, and returns the nearest passages. A result needs no term in common with the query, because the comparison measures meaning.
Any passage of legal text can be the query. A lawyer can take a holding, a key sentence, or a full paragraph from a case and search with the text itself. Vector search returns the passages across the corpus that say like things in unlike words. Searching with a favorable holding, for example, finds other opinions articulating the same rule of law, whatever vocabulary each court chose.
Semantic search and vector search are often used interchangeably. In careful usage, semantic search names the capability, matching by meaning, while vector search names the mechanism that delivers it, retrieval by distance between embeddings.
How is semantic search different from keyword search in legal research?
In legal research, semantic search differs from keyword search by retrieving cases on their legal meaning instead of the words they share with the query. Keyword retrieval requires the researcher to guess the vocabulary a judge used. A wrong guess makes a controlling authority invisible. Semantic retrieval matches the question's meaning against every passage in the corpus, so a governing precedent no longer hides behind a judge's word choice.
Each method keeps an advantage the other lacks. Keyword search is the sharper tool when the researcher knows the exact language, a defined term, a statutory section number, a phrase quoted from a controlling opinion. Semantic matching can blur that precision, returning passages that mean roughly the right thing when the task calls for a verbatim match. The two methods miss different authorities, which is why the commercial platforms rarely offer semantic retrieval by itself.
In the recall and precision terms Chapter 3 defined, semantic retrieval raises recall, reaching relevant authorities a keyword query would miss without requiring the researcher to guess vocabulary. The gain is real. The richer result list still goes to a human reader on budgeted hours. Better retrieval changes what the list contains, not who reads it.
| Keyword search | Semantic search | |
|---|---|---|
| Basis of matching | Words shared between the query and the opinion | Meaning, measured through embeddings |
| The researcher supplies | The vocabulary the judge used | The legal question, or a passage of legal text |
| Strongest when | The exact language is known, a term of art, a section number, a quoted phrase | The controlling authority shares no terms with the query |
| Failure mode | Misses authority worded differently from the query | Blurs precision when the exact words matter |
| Effect on the trade-off | Precision first, with recall sacrificed | Raises recall without vocabulary guessing |
What is hybrid search in legal research?
Hybrid search in legal research runs keyword and semantic retrieval together and merges the results into one ranking. Keyword matching is precise when the researcher knows the term of art, while semantic matching reaches the authorities phrased in vocabulary the query never anticipated, so the combination outperforms either method alone.
Platforms that adopt semantic retrieval usually deploy it this way, as one signal blended into the ranking rather than a replacement for the keyword index. The blend is itself a ranking decision, governed by the same proprietary rules that order every result list. The vendor decides how much weight meaning carries against matched words. The result is a better list. It is still a list.
The Line Search Does Not Cross
Is semantic search the same as agentic legal research?
No. Natural language search, relevancy ranking, semantic search, and hybrid search are search tools. The lawyer operates them from a search box the way lawyers have since legal research first moved online. She writes the query, receives a ranked list, reads what the budget allows, and refines the query again. Agentic legal research changes the whole process, not just the search.
In agentic legal research, every technique in this chapter becomes a tool in the agent's hands. AI agents perform the legal research process end to end under a lawyer's supervision. Retrieval tools like these, prepared for the agent's use, are how the agents uncover the law. The agent runs broad searches, reads every authority returned, follows the recorded connections among cases, and writes a legal research memorandum. Advanced retrieval is necessary to that system but does not define it. What defines agentic legal research is the agent performing the process at a reading depth the billable hour never permitted.
Chapter 5 explains the parts of the system no search box offers, the law curated for agent consumption and the recorded connections the agent traverses.
