In a nutshell
The platform now consistently treats your documents by the principle of "read once, use everywhere" — and makes visible on every AI answer which sources actually went into it, what was trimmed, and when the AI double-checked the original to be safe.
Why context economy is more than thrift
Anyone working with large case files knows the fundamental problem of every AI application: the amount a model can survey at once is limited. Tools that simply re-read everything for every question don't become more thorough that way — just more expensive — and beyond a certain case size they start silently leaving things out without anyone noticing. That is exactly what users of large law-firm AI tools are currently complaining about in public. Our answer isn't a bigger shovel, but a budget: whatever has been read once is kept as structured knowledge — as facts with evidence, as relationships, as a summary with selectable depth — and reused by every surface of the platform. And where trimming is still unavoidable, the trim is shown instead of concealed.
The same document only costs you once
The platform now recognizes uploaded documents by their content, not just their file name. If you upload a file that already exists — perhaps under a different name — the platform suggests reusing the existing document: the text recognition and analysis you already paid for are preserved, and no duplicate facts or unnecessary clarification cases arise in the fact graph. This applies in the lawyer portal just as in the consumer area, where "Use existing" simply skips the repeated, paid analysis.
One summary, selectable depth
Summaries are now created once per document and reused everywhere — in chat attachments, in pleading drafts, in the letter generator. Previously, the same case document could trigger a complete second summarization chain in every new conversation; now the second access simply costs nothing. You decide how detailed it is: the case file's depth control now also applies to the case summary itself. And perhaps the most important point is a compliance point: if you subsequently exclude a document from the AI, its content also disappears from an already-written summary — the summary is suppressed rather than delivered anyway "with a note".
The AI double-checks the original — and tells you
Agents now prefer working with the stored summaries instead of tons of raw text. To keep that from turning into superficiality, there is a counterweight: before a draft is created, an upstream verification step reviews the assembled context — and if it sees a concrete doubt or a possible contradiction, it re-reads the original document in a targeted way. Each of these lookups appears on the run as an expandable "Double-checked for safety" line, with the reason and a clickable document. This is deliberately built as a trust signal, not as a warning: you see not only WHAT the AI knew, but also that it questioned its own state of knowledge.
The AI knows every fact in your case file — with a clickable citation
The fact graph is now a genuine answer source. The AI assistant can search a case file's recorded facts directly and free of charge, instead of leafing expensively through the full text for every factual question; the especially inexpensive fact-graph chat receives the most important facts together with their supporting quotes — every statement based on them can be verified with a click in the original document. Amounts, currencies and dates are additionally captured as real values, so "Sum up all claims" or "Which deadline is next?" is no longer an estimate but a calculation. Tables stay tables instead of crumbling into individual rows, and a fact that already appears as a deadline in the register is never proposed a second time.
New: the source line on every AI answer
Until now, only an agent run showed what it was built from ("Built from N facts · M documents"). From now on, the AI assistant, Lex & Lexi, the Moot Court and the Canvas editor also carry their own source line on every answer: how many facts, documents and relationship links went in, whether the case summary was included — and, as the honest flip side, when something no longer fit into the context. An answer without a basis now looks different from an answer drawn from a well-digested case file, and that is exactly how it should be. It is the same stance we already described for the autopilot: with us, control isn't a settings menu — it's a product promise.
Staying honest
Even with good budgeting there are limits: very large case files never fit completely into a single AI call, and individual caps (per document or per context block, for instance) deliberately remain in place. The difference is that you now see these limits — as an "n of m" figure, as a trim notice, as a lookup log — instead of guessing them. How to use your quota most effectively is covered in the firm wiki under Costs & Control; what the AI actually sees of a case file is explained by AI case coverage. Not on board yet? For lawyers shows how to get started.
