Obsidian 4.6 — Pakistan's Agentic Legal Intelligence System
Legal research is not just about finding a case that appears to support your argument. The real work is finding the right authorities, reading what the court actually said, checking the governing statute, identifying cases that go the other way and making sure the citations being relied upon genuinely exist.
Obsidian 4.6 is Pak Legal Database's legal research and analysis system, built on Grok 4.6 and developed specifically around Pakistani law.
It combines a 500,000-token context window with hybrid retrieval, full-judgment analysis, statutory research, citation verification, document intelligence and multi-stage legal research across PLDB's proprietary collection of 166,872 cases and 5,103 statutes.
PostgreSQL • Vector Search • Full Judgments • PDFs • Citation Verification
Not Just a Summary Search
Obsidian does not simply search short AI-generated summaries and then ask the model to fill in the gaps.
Its research system works across several layers of the legal database:
- PostgreSQL: structured case and legislative records, including citations, courts, dates, case information and other legal metadata.
- Vector Search: semantic retrieval that can locate relevant authorities even where the judgment uses different wording from the researcher's query.
- Full Judgment Text: the underlying judgment can be examined before the final answer is formed.
- Original PDFs: the source judgment remains connected to the answer.
- Legislation: relevant statutory provisions can be searched and considered alongside the cases interpreting them.
These sources complement each other. Structured SQL retrieval is useful for precise legal information, vector search helps discover conceptually relevant authorities, and the full judgment provides the context needed to understand what the court actually decided.
It Reads the Judgment Before Forming the Answer
A case summary is useful for discovery, but it is not a substitute for the judgment.
When a case becomes important to the research, Obsidian can work with the full text of the judgment. This allows it to consider the facts, procedural history, arguments, judicial reasoning and conclusion rather than relying only on a short description of the case.
The original PDF remains connected to the authority. Where relevant pages have been identified, clicking the source can take the researcher directly to the page relied upon.
That means a user can move directly from the answer to the actual judgment and check the relevant passage for themselves.
One of the most serious problems with general-purpose AI in legal research is citation hallucination: an authority can look perfectly genuine while being inaccurate or completely fabricated.
Obsidian therefore includes a citation-verification stage as part of its research workflow.
Authorities can be checked against PLDB's underlying legal records before being treated as verified in the final research.
- Does the judgment actually exist?
- Is the citation contained in PLDB's corpus?
- Does the citation correspond with the correct case?
- Can the underlying judgment be opened?
- Can the relevant source material be traced back to the PDF?
Where an authority cannot be independently confirmed from the corpus, Obsidian can expressly mark it as "not verified in corpus" rather than silently presenting it as a verified citation.
That distinction matters. The system does not have to pretend that every authority mentioned during research has passed the same verification test.
Visible Research Trail
Obsidian also shows users what the research system is doing while a request is being processed.
The interface displays a live research trail under Thinking, showing the main research operations being carried out.
This lets the researcher see that the system is not simply waiting for a language model to generate text. It is performing searches, retrieving legislation, checking different formulations of the issue and examining additional authorities before the response is completed.
The research trail shows retrieval and research operations performed by the system; it is not intended to display the model's private internal reasoning.
Obsidian Searches More Than Once
A single legal question can be expressed in several different ways.
Searching only the exact words entered by the user can miss useful judgments. Obsidian can therefore explore different formulations of the same issue during a research request.
For example, a question about vehicle superdari may cause the system to explore concepts such as:
- superdari of vehicle
- interim custody of motor vehicle
- registered owner entitled to custody
- section 516-A Cr.P.C.
- vehicle released pending trial
- absence of rival claimant
- special statutory restrictions on release
The strongest results from those searches can then be compared and analysed together.
It Looks for Cases Against Your Position Too
Good legal research should not stop after finding a few authorities that support the proposition.
Obsidian is designed to identify contrary and distinguishable authorities as part of the same research exercise.
In a superdari question, for example, the system may identify cases granting temporary custody to the registered owner while also finding authorities where release was refused because of a statutory prohibition, disputed ownership, tampered identification numbers or some other distinguishing circumstance.
This allows the final answer to explain not only when the rule applies, but also where its limits lie.
What a Finished Answer Can Contain
Depending on the question and research mode, an Obsidian answer may be organised into:
- Proposition
- Short Answer
- Relevant Statutory Framework
- Authorities
- Detailed Analysis
- Legal Principles Derived from the Cases
- Contrary / Distinguishable Authorities
- Conclusion
- Sources
- Direct links to the underlying PDFs and relevant pages
The purpose is to give the researcher something closer to a structured legal research note than a generic chatbot response.
Fast and High — Two Different Research Modes
Not every legal question needs the same amount of research.
Obsidian therefore has two modes: Fast and High.
Fast is the mode we expect users to use for roughly 90% of normal legal research.
It is quick, focused and capable of producing surprisingly detailed research without making the user wait for a full deep-research cycle.
Fast can still:
- search case law;
- search legislation;
- identify relevant authorities;
- analyse judgments;
- find contrary precedents;
- verify citations;
- link sources and PDFs;
- analyse uploaded documents; and
- produce structured, source-grounded answers.
For most day-to-day questions, Fast is the place to start.
High is deliberately slower.
It is designed for matters where depth is more important than speed and the user wants a broader research exercise rather than the quickest good answer.
High can spend more time exploring the issue, examining additional authorities, retrieving legislation, reading relevant judgment material and building a more detailed synthesis.
The result is generally a longer, richer and more comprehensive answer.
High is particularly useful for:
- complex or unusual questions of law;
- multiple competing precedents;
- cases involving several statutory provisions;
- matters requiring detailed distinction between authorities;
- large or complicated uploaded documents;
- research intended to support substantial legal drafting; and
- questions where the user wants the widest useful analysis available.
Fast vs High
| Feature | Fast | High |
|---|---|---|
| Speed | Fast | Slower by design |
| Best suited for | About 90% of normal research | Complex / research-heavy matters |
| Case-law research | ✅ | ✅ |
| Legislation research | ✅ | ✅ |
| Full-judgment analysis | ✅ | ✅ Deeper |
| Citation verification | ✅ | ✅ |
| Contrary authorities | ✅ | ✅ Broader review |
| Document analysis | ✅ | ✅ |
| Research breadth | Focused | Broader |
| Answer style | Focused and detailed | Longer, richer and more comprehensive |
| Recommended use | Default | When depth matters more than speed |
How a Complex Research Request May Be Handled
The exact workflow varies with the question, but a deeper request can involve:
-
1. Identify the legal issue
Understand the proposition, relevant provisions and likely areas of law. -
2. Search PostgreSQL
Use structured case and legislative data to locate potentially relevant authorities. -
3. Run vector searches
Search semantically for judgments dealing with the same principle using different language. -
4. Search alternative formulations
Explore different legal expressions of the user's question. -
5. Search legislation
Retrieve the relevant statutory provision where applicable. -
6. Read relevant judgment material
Move beyond summaries and examine the full authority. -
7. Compare supporting authorities
Identify the principles emerging from the cases. -
8. Search for contrary authorities
Find cases that refuse, limit or distinguish the proposition. -
9. Verify citations
Check the authorities against PLDB's underlying corpus. -
10. Form the answer
Bring the legislation, verified authorities, distinctions and analysis together into one structured response.
Basic RAG vs Obsidian 4.6
| Feature | Basic RAG | Obsidian 4.6 |
|---|---|---|
| Single vector search | ✅ | Not limited to one search |
| PostgreSQL structured search | Usually No | ✅ |
| Semantic vector search | ✅ | ✅ |
| Multiple query formulations | Usually No | ✅ |
| Full judgment analysis | Limited | ✅ |
| Original PDF sources | Limited | ✅ |
| Direct page-level PDF links | Usually No | ✅ |
| Statutory retrieval | Limited | ✅ |
| Citation verification | Usually No | ✅ |
| Flags unverified citations | Usually No | ✅ |
| Contrary precedent research | Usually No | ✅ |
| Visible research trail | Usually No | ✅ |
| Large-context analysis | Limited | ✅ 500K |
| Fast + High research modes | ❌ | ✅ |
Why 500K Context Matters
Legal research often involves very large amounts of text.
A single judgment may contain dozens or hundreds of pages. A difficult matter may involve several judgments, statutory provisions, pleadings, contracts and other documents at the same time.
Obsidian 4.6 has a 500,000-token context window, giving it room to work across unusually large quantities of legal material within the same research context.
That becomes particularly useful when the answer depends on comparing several judgments rather than extracting one paragraph from one case.
- Built on Grok 4.6
- 500,000-token context window
- 166,872 Pakistani cases
- 5,103 statutes
- Hybrid retrieval using PostgreSQL + vector search + full judgments
- Does not depend only on summaries
- Can work with the underlying full judgment before answering
- Connects authorities to the original PDF
- Can open sources at the relevant PDF page
- Performs citation verification
- Can flag authorities as not verified in corpus where verification is unavailable
- Searches for contrary and distinguishable precedent
- Displays a visible research trail while working
- Searches case law and legislation as part of the same research exercise
- Fast handles roughly 90% of normal research quickly
- High performs deeper research and produces longer, richer answers
- Supports uploaded document analysis
- Supports English, Urdu, Roman Urdu and Punjabi
A query asking when Pakistani courts have granted superdari of a vehicle does not have to end with a list of cases.
Obsidian can identify the governing provision (section 516-A Cr.P.C.), retrieve grant cases, examine the relevant pages of those judgments and explain the principles applied by the courts.
It can then separately identify contrary or distinguishable authorities, such as cases involving statutory restrictions under the CNSA or disputes over title and possession.
The final answer can therefore explain both: when superdari is granted and when the same principle does not apply.
Sources remain connected to the answer and can be opened directly in their original PDFs.
Source-Grounded Does Not Mean Blind Trust
No AI system is completely immune from error.
Obsidian's retrieval, full-document access and citation-verification layers are designed to reduce common legal-AI failures, particularly fabricated authorities and unsupported citations.
But confirming that a judgment exists is different from guaranteeing that every interpretation of that judgment is correct.
The important difference is that the researcher does not have to take the answer on faith. The underlying judgment can be opened and checked.
For professional legal work, the original case or statutory provision should always be reviewed before being relied upon in court, formal advice or legal drafting.
TL;DR
- Obsidian 4.6, built on Grok 4.6
- 166,872 Pakistani cases
- 5,103 statutes
- 500K context window
- PostgreSQL + vector retrieval + full judgment analysis
- Searches multiple formulations of a legal issue
- Searches case law and legislation
- Reads full judgments rather than relying only on summaries
- Connects results to original PDFs
- Can open the relevant PDF page directly
- Verifies citations before treating them as confirmed
- Flags citations that are not verified in the corpus
- Searches for contrary and distinguishable judgments
- Shows a visible research trail while working
- Fast: excellent for roughly 90% of everyday legal research
- High: slower, deeper and designed for richer, longer and more comprehensive research
- Document analysis
- English, Urdu, Roman Urdu and Punjabi support
Best Practice
- Start with Fast. It is designed to handle the vast majority of normal research and is usually all you need.
- Use High when the question deserves deeper treatment. It takes longer because it is intended to return a broader and more detailed research answer.
- Be specific. Give Obsidian the facts, provision, jurisdiction or procedural issue where possible.
- Upload relevant documents. The more relevant material available to the research system, the more useful the analysis can be.
- Watch the Research Trail. It shows the searches, legislation retrieval and research operations being performed while the answer is prepared.
- Check contrary authorities. They often matter as much as the cases supporting the proposition.
- Open the source. Use the linked judgment and page references to verify the authority behind the answer.
Ready to try it? Visit www.paklegaldatabase.com and start researching with Obsidian 4.6.