All use cases
Use case · Retrieval-augmented generation (RAG)

Document research

Ask your contracts, policies, filings and manuals a question and get an answer at the level of the clause, each claim cited to the passage behind it. It generalises to questions nobody wrote a rule for, and holds back any answer it cannot cite. It runs in our cloud or in your VPC.

1.6Mpassages indexed
79,268documents with effectivity status
5legal domains

When you need this

Your organisation’s knowledge sits in thousands of PDFs behind a search box nobody trusts. Generic chatbots answer fluently and wrongly, and in legal, compliance or operations a confident wrong answer is worse than none. The documents keep changing, so yesterday’s index serves outdated rules. Some of it is too sensitive to send to anyone’s cloud.

What we build

  • The right passage first: each question is split into sub-queries that run on semantic and keyword search, and the results are merged and reranked.
  • A source for every claim: each claim carries a citation number that opens the exact passage used, with its document number and effectivity status.
  • Documents in force first: each document’s status, where it is known, is refreshed on a schedule and checked at query time, so search prefers documents in force.
  • Runs where your data lives: our cloud or your VPC, with an option to write answers in the user’s browser on a small local language model.
  • Every failure becomes a test: a golden set of questions scores retrieval and faithfulness against the live system, each failure we find is added to it, and a gate withholds any answer that has no citation or cites a repealed document as current.

How it’s built

Your corpus → continuous crawl · embed · index → hybrid retrieval + rank fusion → LLM rerank + answer with citations → citation and currency gate

Methods: hybrid retrieval · LLM pipelines & agents

In our products

vnadvise answers questions on Vietnamese law in five domains (investment, tax, real estate, securities and foreign exchange) at article and clause level, over 1,606,838 indexed passages. Each claim cites the passage behind it, with the document number and its effectivity status, known for 79,268 documents; retrieval prefers documents in force, and documents no longer in force can still appear. A Local mode runs answer generation in the browser on Qwen 3.5 0.8B; questions still go to the server for document lookup.