Semantic vs Keyword Case-Law Search in India: What's the Difference?

Explainer · Case Search Technology

Updated for 2026 · Claw Editorial

Keyword, or Boolean, search matches the exact words you type. Semantic, or AI, search matches the meaning behind your question, so it can find a relevant judgment even when that judgment uses completely different words. Neither one is simply "better". Keyword search is precise but brittle: it misses synonyms, paraphrase, and misspelled names. Semantic search is more forgiving but can surface a plausible-looking result that is not actually on point, if it is not grounded in real judgment text. This explainer covers how each works, exactly where each one fails, why party-name matching is a special case on its own, and how to combine both approaches for case-law research you can actually rely on.

01How keyword (Boolean) search works, and where it fails

Keyword search, sometimes called Boolean search, looks for an exact match between the words you type and the words in the document. It is the oldest and most literal way to search a body of text, and most Indian legal databases were originally built on it.

How it works

A keyword search engine indexes every word in every judgment, then looks up your query terms in that index. Boolean operators (AND, OR, NOT) and proximity operators (finding two words within a set distance of each other) let you narrow or widen the search. Type "breach AND contract AND damages" and the engine returns only judgments containing all three words, somewhere in the text. This is fast, predictable, and easy to audit: you can see exactly why a result matched, because the matching word is right there on the page.

Where it fails

The weakness is exactly the same as the strength: keyword search only knows words, not meaning. Three failure modes show up constantly in Indian legal research.

• Synonyms and drafting variation. A judgment might discuss "termination of employment" while your query says "wrongful dismissal". Both describe the same legal question, but a keyword search treats them as unrelated unless you manually think of and type every variant.
• Paraphrase. The legal point you are researching, such as whether a surety is discharged when a creditor gives the principal debtor more time to pay, can be expressed in a judgment using entirely different sentence structure and vocabulary than the one in your head. Keyword search cannot bridge that gap.
• Misspelled or variant party names. If a party’s name is spelled differently in the record than in your query, a pure keyword search returns nothing, even though the case exists. This is common enough in Indian records that it deserves its own section below.

In practice, keyword search rewards a researcher who already knows the right vocabulary and can guess every phrasing a judgment might use. That is a reasonable ask for a senior practitioner working in a familiar area. It is a much harder ask for anyone researching outside their usual specialisation, or a junior lawyer still learning the terminology of a new practice area.

02How semantic (AI) search works, and where it fails

Semantic search takes a different approach: instead of matching words, it tries to match meaning. This is the technology behind most modern AI-based legal research tools.

How it works

A semantic search system converts text, both your query and every judgment in the database, into a numerical representation called an embedding. An embedding places a piece of text as a point in a high-dimensional space, positioned so that texts with similar meaning sit close together, even if they share very few actual words. When you search, the system compares your query’s embedding to the embeddings of every judgment and returns the ones that sit closest in meaning, not the ones that share the most literal words. This is why a semantic search for "wrongful dismissal" can surface a judgment that only ever uses the phrase "termination of employment", and why a paraphrased legal question can still find the judgment that actually answers it.

Where it fails

Semantic search trades precision for recall, and that trade has a real cost if it is not managed carefully.

• Plausible but irrelevant results. Because the system is matching meaning rather than exact fact, it can return a judgment that reads as relevant, uses similar language and a similar fact pattern, but is not actually on the point you are researching. The result looks right until you read it closely.
• Ungrounded generation is a bigger risk than ungrounded retrieval. Semantic search that only retrieves real judgment text and shows you the source is one thing. A general-purpose AI model that generates a summary or a citation from its own memory, without pointing back to a real judgment, is a different and riskier thing. That is how a citation can be invented outright rather than merely mismatched. This is a separate and serious question worth understanding on its own; see our explainer on whether legal AI hallucinates citations in India.
• It can be a black box. A keyword match is self-explaining: the word is right there in the text. A semantic match is not always obvious at a glance, which is why a trustworthy tool shows you the source judgment alongside the result, so you can verify the match yourself rather than taking the system’s word for it.

Keyword search misses the case that used different words. Semantic search can surface the case that only looks like it answers your question. A tool that is honest about both failure modes is more trustworthy than one that claims to have solved them.

03The special case: matching misspelled and transliterated names

Party-name search deserves its own treatment, separate from ordinary legal-principle search, because Indian court records make it a genuinely hard problem on its own.

Indian names are transliterated from Hindi, Urdu, Tamil, Telugu, Kannada, Malayalam, Marathi, Punjabi, Gujarati, and other regional languages into Roman script, and there is no single standard way to do that. The same person or company can appear under several different spellings across the same court system: "Krishnamurthy" and "Krishnamoorthy", "Mohammed" and "Mohd", "Subramaniam" and "Subramanian". Add in decades of records typed by different clerks, plus OCR errors from scanning older paper filings, and a single party name can accumulate a long tail of spelling variants across an archive.

Neither plain keyword search nor plain semantic search solves this cleanly on its own. A keyword search for one spelling will simply miss every record filed under another. Semantic search, built to match legal meaning rather than proper nouns, is not designed for this job either: two judgments about completely different parties can be about the "same" legal meaning while being about entirely different people.

The tool built for this job is phonetic, or "sounds like", matching, sometimes combined with edit-distance ("fuzzy") matching for typos and OCR errors. Soundex is the classic algorithm: it encodes a name by its consonant sounds and drops vowels, so names that sound alike get grouped together even when spelled differently. Soundex was built for English-language names, so on its own it under-performs on Indian names, which involve aspirated consonants, "v"/"w" interchange, and transliteration patterns that standard Western phonetic algorithms were never designed to handle. Legal search tools that work well on Indian names generally need additional, India-specific tuning on top of a base phonetic algorithm. For the full detail on how this works, see our explainer on phonetic name matching.

There is also a reverse risk worth naming honestly: the name-collision problem. Widening a search to catch every spelling variant of a common name, such as "Ramesh Kumar" or "ABC Enterprises", can also pull in results belonging to entirely different, unrelated parties who simply share that name. A tool tuned only to avoid missing a variant, with no way to help a reviewer tell same-named parties apart using case context, address, or other detail, trades one failure mode for another. Solving this well means casting a wide net for spelling and transliteration variants, then giving the reviewer enough detail in each result to confirm identity, not just a name match.

If you are searching for a case right now and suspect the name is the problem, our step-by-step guide on searching cases when a party name is misspelled walks through practical techniques, including how to check whether phonetic search is actually switched on.

04How to combine both for reliable research

The honest answer is that keyword and semantic search solve different problems, and reliable case-law research needs both, not a choice between them.

• Start broad with semantic search when you are researching a legal question and are not sure of the exact vocabulary a judgment might use, or when you are working outside your usual specialisation. This is where semantic search earns its keep: it surfaces judgments you would not have found by guessing keywords.
• Use keyword or Boolean search when you know the exact phrase, section number, or citation you are looking for, and want a precise, auditable match rather than a meaning-based one. Keyword search is also the right tool for confirming a specific fact inside a judgment you have already found.
• Turn on phonetic or fuzzy name matching whenever your search includes a party name, by default rather than as an afterthought, since you usually cannot tell in advance whether the name in the record was spelled the way you typed it.
• Verify before you rely. Whichever method returns a result, read the underlying judgment before citing it. This matters more for semantic results, since a meaning-based match can look right without actually being right, but it is good practice for keyword results too.

A tool that offers only one of these two search modes is asking you to work around its gap manually. A tool that offers both, and lets you tell which one produced a given result, gives you the precision of keyword search and the coverage of semantic search without forcing a trade-off.

05How to test which one a tool actually uses

Vendor marketing pages are not always precise about whether a "smart search" is genuinely semantic, or just keyword search with some autocomplete and synonym expansion layered on top. You can find out yourself in a few minutes.

1. Run a paraphrase test. Take a legal question and phrase it in your own words, deliberately avoiding the exact terminology a judgment would use, for example asking about a "surety let off the hook" instead of "discharge of a surety". If the tool still surfaces the relevant judgment, it is doing real semantic matching, not just keyword lookup.
2. Run a misspelling test. Search for a party name you know exists in the database, but deliberately misspell it the way a clerk plausibly might. If the case still comes up, phonetic or fuzzy name matching is genuinely switched on.
3. Run a precision test. Search for an exact phrase or a section number you know appears in a specific judgment. If a tool that claims to be "AI-powered" cannot reliably find an exact, known phrase, its underlying keyword layer may be weak, which matters when you need a precise match, not just a relevant one.
4. Check whether the source is shown. For any semantic result, look for whether the tool shows you the actual judgment text behind the match. A tool that returns a summary or a citation with no visible source is harder to verify, and that is exactly the gap where an ungrounded or mismatched result can slip through unnoticed.

For a measured, checkable look at how citation accuracy is tested and scored, rather than just claimed, see our citation accuracy benchmark, which sets out the method and discloses where results still go wrong. And if you are weighing AI-based research against how case-law research was traditionally done, our explainer on AI legal research versus traditional legal research covers that comparison in full.

06Where Claw fits

Claw is an all-in-one legaltech platform for Indian advocates, law firms, and corporate legal teams, combining AI-based case search, an AI legal assistant (Legal GPT), case management, and compliance automation across all Indian courts and tribunals.

On this specific question, Claw does not force a choice between keyword and semantic search. Its case search combines semantic and AI-based search, so it can find a relevant judgment even when your query and the judgment use different words, with results returned in under 5 seconds across its coverage of 30 crore-plus judgements spanning 25 High Courts (1980 to 2026) and the Supreme Court (1950 to 2026). Its search is also name-tolerant, applying proximity and phonetic matching so that misspelled and transliterated party names are still found, which is the specific diligence trap this page describes in the names section above.

Claw does not publish a bare "AI accuracy" claim without a way to check it. For a measured, dated citation-accuracy figure with a stated method, not a marketing number, see the benchmark linked above. And if you are new to legal research tools generally and want a fuller sense of the landscape before choosing one, our guide to the best AI tools for law students and junior advocates in India is a good next stop.

07Frequently asked questions