A NotebookLM for one case file. Grounded, cited, zero-hallucination by design.
Case Trail's differentiator. Per-matter retrieval-augmented AI that reads only the documents you uploaded, cites every line back to the page, and refuses to answer when it can't.
What's broken about AI in legal work today.
Lawyers tried generic LLMs, got burned by a fabricated citation, and went back to email. The Notebook fixes the part that matters: trust.
Generic legal LLMs hallucinate citations
Ask ChatGPT for the relevant paragraph and it invents a cause number, a forum, even a judge. By the time you catch it, the draft is in counsel's hands. Free generation has no place in legal work.
Cross-matter bleed is a confidentiality breach
A model fine-tuned on one client's contracts will leak phrases into another's draft. Banks, NBFCs and firms can't put privileged material into a shared model — period.
Counsel still reads the brief from page one
Junior counsel spends two days summarising a 600-page filing before the senior even opens it. The summary is then re-summarised in the pre-hearing meeting. Same work, three times.
Drafts start from a blank page or a stale template
The petition you filed last month had the right structure, the right precedents, the right tone. None of that is reachable when you start the next one — it's locked in the PDF.
What the Notebook delivers, end to end.
Six things every other 'legal AI' product gets at least one of wrong. We get all six right because the architecture starts with grounding, not generation.
Per-matter source rail
Toggle which documents ground each query. Loan agreement only · loan agreement plus the 13(2) notice · the entire matter folder. The Notebook never reaches for anything outside the rail you opened.
Cite-as-you-write
Every claim in an assistant response carries a clickable [cite] chip mapped to a source chunk and page number. Hover to see the paragraph; click to jump to it in the document viewer.
Refuses ungrounded claims by design
If no retrieved chunk supports a sentence, the Notebook says so explicitly — no disclaimers, no creative paraphrase. Refusal is a feature, not a bug.
Tenant-isolated, RBAC-aware
Vector store is keyed by tenant_id and case_id. The Notebook cannot cross matters, cannot cross tenants, cannot ingest a document the requesting user isn't entitled to read.
Pre-built legal prompts
Summarise filing · extract dates · compare petition with counter · draft rejoinder · generate Q&A for the hearing · cross-reference order with prayers. One click, grounded answer, exportable to drafting or vault.
Prompt caching for stable parts
System prompt and retrieved-chunk preamble are cached at the model boundary; only the user question is uncached. Per-tenant cache hit rates above 80% on hearing-prep workloads — enterprise-grade economics.
Why teams pick Case Trail over a generic LLM.
“The first time the Notebook said ‘no grounding for that claim’ — that’s when the recovery team trusted it. Every other tool would have invented an answer; this one refused. We’ve never seen a hallucinated citation come out of it.”
Try the Notebook on a sample matter.
Loaded with Indian filings, orders and notices. Ask it anything — see every cite, see every refusal.