Insira Kodex
Product, policy and statutory knowledge as a structured hierarchy — queryable in dialogue, every answer backed by the passage it came from.
Trialled in insurance distribution on a curated knowledge base covering product and tariff logic. Reference available on request.
A question to the knowledge base
Question above, answer below. Every statement carries its source — and shows whether it comes from the policy conditions or from statute. Where there is no source, the system says so.
Question
From when is surrender possible under the Vorsorge Musterplan, and with what deductions?
Answer
01
The wrong country, correctly cited
General-purpose models do not fail on Austrian insurance law out of ignorance — they know insurance law well. They just know mostly German insurance law. The German-language legal corpus on the web is overwhelmingly from Germany: commentaries, law-firm articles, forums, model clauses. Austrian material is the exception in it, and a model answers from the mass. The abbreviations make it worse: VersVG and VVG differ by three letters and govern the same subject in two countries, and VAG is the abbreviation of a supervisory statute in Austria and in Germany alike — two different ones. Ask about the surrender of a life policy and the Austrian answer is § 176 VersVG while the German one is § 169 VVG. Both exist, both carry the same heading, both are correctly formatted. That is precisely the danger: the error does not look like an error. It is not an invented citation but a real one from the wrong jurisdiction — and it does not occur reliably, it occurs today and not tomorrow. A hint in the prompt shifts the probability; it does not change what was trained on. In advice you are liable for, a wrong answer you cannot spot is worse than no answer. Insira Kodex does not know Austrian law better — it knows nothing else. There is no German corpus to fall back on, because there is no falling back at all: every statement comes from a node in the knowledge base, or it is reported as unsupported.
02
A hierarchy, not a document dump
The knowledge base is neither a folder of PDFs nor a vector index of equal-length text fragments. The smallest unit is a node: a single normative statement with a stable identifier, held in its place in the tree — statute, section, paragraph, subparagraph on one side; product, tariff generation, policy conditions, section, clause on the other. A node inherits the scope of its parents. That is exactly what is lost when a document is cut into windows: the clause stays readable, the proviso two levels above it does not.
03
Temporal validity is part of the data model
Every node carries the period for which it applies. A contract is judged by the conditions in force when it was concluded, not by today's — to a flat similarity search both versions are the same document, and it will confidently return the wrong one. In Kodex the tariff generation is an axis of the hierarchy, not a note in the filename. Relations between nodes are typed — supersedes, supplements, restricts, refers to. Contradictions are resolved rather than returned side by side.
04
The law is kept current
The statutory layer covers the material relevant to insurance distribution — VAG 2016, VersVG, and the distribution duties arising from the Austrian Trade Act and the IDD. The source is the consolidated version in the Austrian federal legal information system, and changes are picked up automatically. They overwrite nothing silently: affected nodes are flagged as changed and put to the business side for confirmation. Statute and policy conditions remain separate layers, and the statutory layer takes precedence — mandatory law beats the clause.
05
Answers with evidence
Every answer carries the passage it came from, with its layer and its version. Anyone who has to advise can check the statement before passing it on, and open the original when in doubt. The PDF does not disappear — it simply stops being the layer you search, and hangs off the node as evidence.
06
Bounded, not omniscient
The system answers from the stored knowledge base and from nowhere else. What is not in it is marked as unsupported rather than plausibly filled in. Because the hierarchy knows which nodes exist, that answer is a finding rather than a guess. In regulated distribution, that boundary is exactly what makes deployment possible.
07
Local or cloud language model
The language model is interchangeable. Kodex runs against a model hosted on your own hardware just as it runs against a cloud endpoint. The structure is what makes that possible: because retrieval navigates the tree and hands over a small, precisely bounded set of nodes, the model has to phrase rather than to know. A small in-house model is equal to that task — and portfolio data and tariff knowledge never leave the building.
08
Maintained by the business side
The nodes live as structured Markdown under version control. The business side maintains its own knowledge without a detour through engineering, and changes are traceable like code — who changed what, when, and why. The work sits in the modelling, not in the chat interface: without sound domain modelling a language model answers domain questions fluently and wrongly.