What Vitark Does Not Do Well Yet
This page forms part of our Terms and Conditions and is referenced in Section 5 of that document. It is reviewed and re-dated every quarter, and whenever a listed limitation is fixed or a new one is found.
Why this page exists
Most products tell you what they can do. This one also tells you what they cannot.
There are two reasons. The first is that you are using Vitark to make decisions about your own legal position, and a tool that hides its weak spots is more dangerous than one that names them. The second is that we would rather you found these limits here than discovered them in the middle of something that mattered.
If you find a limitation that is not on this list, tell us at ask@vitark.ai. We will investigate, and if it is real we will add it here.
1. What the corpus covers, and what it does not
Covered: 1,04,83,352 judgments from the Supreme Court of India, 25 High Courts, and tribunals. Statutory text from India Code, including central Acts and state collections.
Not covered:
- District and subordinate courts. The overwhelming majority of Indian litigation happens here and almost none of it is in our corpus. If your matter is in a district court, we can tell you what the High Court above it has held. We cannot tell you what that district court has been doing.
- Every tribunal. We hold tribunal material but not all tribunals, and coverage is uneven between them.
- Live case status, cause lists, and orders not yet uploaded. We are a research corpus, not a case-tracking service. We do not know whether your matter is listed tomorrow.
- Unreported judgments. If it was never published, we do not have it.
2. Retrieval is strongest where the case law is dense, and weakest where it is famous
This one surprises people, including us.
Vitark performs best on questions with many similar judgments spread across High Courts: cheque bounce, bail, consumer complaints, tenancy, FIR registration, employment. That is the everyday litigation the product was built for, and the density is what makes the retrieval good.
It performs worse on the small number of very famous constitutional judgments that most people test first. If you open Vitark and search for a landmark you already know by name, there is a real chance it will not come back at the top, or will not come back at all.
This is a known defect. It is on the engineering backlog. We are naming it here because a first-time user testing a famous case and getting a poor result should know that this is the shape of the weakness rather than a sign that the whole system is unreliable.
3. Case names come from the judgment text, not from a name field
Our corpus does not hold party names as structured data. The fields that would ordinarily carry a case title, a petitioner and a respondent are empty across the record. Where a case name is shown to you, it has been read out of the text of the judgment itself.
In the product, this mostly works. Party names are displayed, and where a name genuinely cannot be recovered from the text the product says so rather than showing you a blank.
Through our machine interface it does not. Results returned by our API and by connected AI clients carry the neutral citation, for example 2007 INSC 1266, and not Eastern Book Company v. D.B. Modak. If you reach Vitark through another application rather than through vitark.ai, expect to identify cases by citation, and verify any party name that application shows you against the judgment text.
Two practical consequences, both worth knowing before you rely on a result:
- Searching by party name is unreliable, in the product and through the interface alike. Search by subject, by section, or by citation. A case you are looking for may be on your screen without looking like it.
- A name read out of judgment text can be read wrongly, particularly where the text is a scan. See Section 6.
This is a data and extraction defect, not a retrieval defect. The right judgment is frequently found and then poorly labelled.
4. A result count is a count of passages, not a count of cases
When we tell you a search returned a number of results, that number counts passages of text, not distinct judgments. A single long judgment can contribute several passages to the same result set.
So "247 results" does not mean 247 cases considered your question. It may mean rather fewer cases, some of them appearing more than once.
Do not use a result count as a measure of how settled a point of law is, and do not repeat one as a count of cases in anything you file or publish. If the number of distinct judgments matters to you, count them.
5. Connected matters are not collapsed into one
Where a court decides several connected appeals in one judgment, or where the same dispute generates a series of orders, we do not merge them into a single entry. They appear as separate results.
The effect is that one dispute can look like four, and a line of authority can look deeper than it is. It cuts the other way too: reading one of a set of connected matters may not give you the whole of what the court decided.
Check the case identifiers. Connected matters usually share a date and a bench and differ only in the appeal number.
6. Some older text is scanned, and the scan is uneven
Part of the corpus comes from scanned documents converted to text by optical character recognition. That conversion is not perfect and its quality varies by source and by period. Older material and some regional-language material are worst affected.
What this looks like in practice: a mangled name, a broken number, a section reference that reads as something else, or a paragraph that does not parse. It also means a search may miss a judgment whose text was scanned badly, because the words you searched for are not in the text as we hold it.
Before you rely on any specific words, a figure, a date or a name, open the source order and check it there. This applies with particular force to anything you intend to quote.
7. Statute results return the whole section
Statutory search returns results at section level. For most sections that is exactly right. For long omnibus sections, particularly definitions sections that run to dozens of sub-clauses, you will receive the entire section and will need to find the relevant sub-clause yourself.
We are working toward sub-clause granularity. Until then, use your browser's find function inside long statutory results.
8. Long judgments are truncated
Very long judgments are truncated in the judgment view at approximately 50,000 characters. Where this happens the interface indicates it. If the point you need is deep in a long judgment, open the source order and read it there. Every result gives you enough detail to find the original.
9. Every output cites a source, and that is not the same as being right for you
Our commitment is to process and method: every substantive output carries its citation, and we do not invent citations. That commitment is set out in Section 5 of the Terms and Conditions and it is one we hold ourselves to absolutely.
It is not a guarantee of legal correctness in your matter, for three reasons that no system can engineer away:
- The source can be wrong. Judgments are written by human judges and contain factual errors, internal contradictions, and reasoning later disapproved.
- The source can be out of date. A judgment can be overruled, distinguished, or superseded by statute. We flag treatment where we can. Our treatment flagging is not exhaustive, and the absence of a flag is not a certification that a case is still good law. Check the current status of any judgment you intend to rely on.
- Your facts may differ from the reported facts in ways that change the result entirely. Cases turn on facts.
10. Language
You can search in 11 Indian languages. The judgments themselves are overwhelmingly in English, with some regional-language material.
So a query in Marathi will search the corpus, but most of what it finds will be English text. If you need the reasoning in your own language, ask for it, but treat the translation as a working aid and read the original before relying on it in any proceeding.
11. Vitark is not a lawyer, and this page does not change that
Nothing on this page, and nothing in the product, is legal advice. We are an information service. Section 4 of the Terms and Conditions sets this out in full.
For anything with material legal, financial or personal consequence, take the research to a qualified advocate. The purpose of the product is to make that conversation better informed and shorter, not to replace it.
Reporting something we have missed
If an output does not match its cited source, or you find a limitation not listed here:
Email ask@vitark.ai with the query, the output, and what you found wrong.
We acknowledge within 48 hours and redress within one month.
Where an output materially misrepresents its cited source, you may also be eligible for a refund under Section 3.5 of the Refund and Cancellation Policy.
Change log
| Date | Change |
|---|---|
| 31 July 2026 | First publication. Items 1 to 8 as then listed. |
| 13 August 2026 | Item 3 rewritten: case names are extracted from judgment text and are absent through the machine interface. Three items added: result counts are passages and not cases (item 4), connected matters are not collapsed (item 5), and optical character recognition is uneven (item 6). Publisher stated as TLY SOFT PRIVATE LIMITED. |
TLY SOFT PRIVATE LIMITED | CIN U62013DC2026PTC473011
Registered office: Innov8 Harsha Bhawan, 4th Floor, 13/29 E-Block, Connaught Place, New Delhi 110001
Vitark provides information, not legal advice.